1
Information Processing and Extraction Group
Group Leader: Dr. Miguel Vélez-Reyes
Members: Dr. Sandra Cruz-Pol, Dr. Shawn Hunt, Dr. Hamed Parsiani, Dr. Luis Jménez,
Dr. Bienvenido Vélez, Dr. Manuel Rodríguez, and Dr. Pedro Rivera
1 Introduction
NASA’s Office of Earth Science (OES) studies Earth as an interconnected system of atmosphere, oceans, continents and life to understand the total Earth system and the effects of natural and human-induced changes on the global environment. Using higher spectral and spatial resolution imagery, and the availability of multi-mode sensing modalities that are (or will be) available from space and sub-orbital platforms, NASA acquires, processes, and make available very large (gigabyte to terabyte) volumes of remote sensing data, and related observations and information to private, public, and governmental entities. This information is, or has the potential to be, used by scientists to understand and solve major scientific mysteries, and by the practitioners and policy makers to solve practical, societal problems, and/or establish sound policy decisions. To reach the full potential of using the available information, we need to improve our ability to extract and manage information about our environment provided by remote sensing data, and related observations.
Signal processing is a broad field that encompasses the acquisition of data from the world around us, manipulating or processing that data into a useful form, the extraction of information from that data, and the interpretation of that information. The breadth and power of signal processing is what makes it one of the key enabling technologies of the information age. Our group research focuses on the development of signal processing algorithms to extract information from remotely sensed and related data.
2 Summary of Accomplishments
This is a summary of the main project accomplishments during the reporting period. A detailed description of the individual projects is given in the following section.
• A new version of the Hyperspectral Image Analysis Toolbox has been released and is available on the WEB from censsis.neu.edu. Spatial features have been combined with spectral features to perform classification of hyperspectral images. Considerable improvement in classification accuracies has been obtained. The spatial features considered are statistical and multiresolution texture features. These feature sets are being added to the HSI analysis toolbox in LARSIP.
• The main accomplishments of this reporting period for the Terrascope project include the development of various image clustering algorithms and their integration into a re-designed graphical user interface capable of assisting the user in situations when a query retrieves many cluttered images. The new graphical user interface automatically groups such images into folders at the client side thus yielding much improved response time and less server load.
• Using rainfall data from a 2Dimensional video disdrometer (2DVD) located at NWS site in San Juan, PR, the microwave remote sensing group found that at least two different types of drop size distribution (DSD) are involved in PR's rainfall systems. The study was conducted during Tropical Storm Jeanne pass over the island of PR in Sept 2004, with Ancillary data from NCDC rain gauges and Doppler radar data from NEXrad, in addition to using the NASA TRMM data to pinpoint the times and site of maximum rainfall. More events will be required to validate rainfall types differences found in this work. In August, the 2DVD was relocated at the near the TCESS facilities at the UPRM R&D Center to study and characterize the rainfall statistics of the West part of the island.
• The material type determination neural network algorithm was extended from pure soil types of sand, loam, and clay to their 50% mixtures, using GPR at 1.5 GHz. A limited database of MCFDs of the three soil types and their 50% mixtures was created by MCFD-NN in a Graphical User Interface (GUI) format, to be used as a desirable product for the end user. The material type determination has gone beyond the laboratory environment into open field measurements. A method of elimination of rocks and other clutter present in the soil has been devised.
A detailed description of the research work is given in Section 7.2. Sections 7.4-7.9 give detailed information about outcomes such as publications, students, and collaborative efforts. Table 7.1.1 gives a summary of important IPEG outcome statistics.
Table 7.1.1 IPEG Outcomes Statistical Data.
|# Investigators (Total # faculty and research associates) |8 |
|# Undergrad students supported (total) |8 |
|# Undergrad students supported (US Citizen, UMD only) |8 |
|# Undergraduate degrees awarded (US Citizen, UMD only) |1 |
|# Masters students supported |9 |
|# Masters students supported (U.S. Citizen, UMD only) |8 |
|# Doctoral Students Supported |3 |
|# Doctoral Students Supported (U.S. Citizen, UMD only) |1 |
|# Grad students supported (total) | 12 |
|# Grad students supported (U.S. Citizen, UMD only) |9 |
|# Masters degrees awarded (US Citizen, UMD only) |6 |
|# Doctoral degrees awarded (U.S. Citizen, UMD only) |0 |
|# Post-docs supported (total) |2 |
|# Post-docs supported (U.S. Citizen, UMD only) |1 |
|# Refereed Papers and Book Chapters Accepted/Published |17 |
|# of student authors or co-authors |18 |
|# Refereed Papers and Book Chapters submitted but not yet accepted or published |6 |
|# of student authors or co-authors | |
| |2 |
|# Presentations at NASA Installations |6 |
|# Presentations national/international research conferences |22 |
|# given by students |4 |
|# Presentations at Faculty seminars | |
|# Presentations to K-12 Schools | |
|# K-12 Students Contacted | |
|# K-12 Educators Contacted |30 |
|# Patents applied/awarded | |
|# NASA MUREP Panels served on (peer review, advisory, etc.) | |
|# Other NASA Panels served on (peer review, advisory, etc.) | |
|Total research $ awarded by NASA other than URC |$90,000 |
|Total research $ awarded by other Agencies | |
|Infrastructure $ leveraged from NASA sources other than URC | |
|Infrastructure $ leveraged from other Agencies during 1998 as a direct result from URC | |
| | |
|UMD = Underrepresented Minority and Disabled | |
3 Research Projects
1 Hyperspectral Image Processing: Miguel Vélez-Reyes, Luis Jiménez, and Vidya Manian
1 Problem Statement
The main objective of this project is the development of information extraction techniques based on hyperspectral imagery (HSI) for environmental applications. In Hyperspectral Imaging or Imaging Spectroscopy, high-spectral resolution and spatial information of the scene under study is collected. As the object of interest is embedded in a translucent media (i.e. atmosphere, or coastal waters), the measured spectral signature is a distorted version of the original object signature mixed with clutter. The high spectral resolution could allow discrimination between media and object contributions enabling the retrieval of information about the object of interest. Our approach is based on integrating physical models and statistical methodologies where appropriate. Algorithm validation is done using simulated data, real data collected using laboratory setups, as well as data available from existing AVIRIS and HYPERION imagery. Validated algorithms are incorporated into the Hyperspectral Image Analysis Toolbox being developed at LARSIP.
The work on the this project can be divided into two components
• Object Recognition
• Estimation Problems
The object recognition has been the prime focus of our work for many years. Mature classification, feature extraction, and training algorithms are being incorporated into a hyperspectral image processing toolbox being developed in the MATLAB environment. Estimation problems have been focused in developing algorithms for unmixing of hyperspectral images.
2 Relevance to NASA
Hyperspectral sensors are being used by NASA scientists for different applications such as land cover classification and mineralogy in earth and planetary[1] surfaces. The proposed research in information extraction algorithms from hyperspectral imagery can provide NASA scientist the tools needed for information extraction to study important earth and extraterrestrial planetary processes.
3 Goals of the Project
Object recognition: The main challenge of this component is how to develop robust algorithms for hyperspectral subsurface recognition problems that can deal with the high dimensionality of the data, and with an embedding media that can have adverse optical properties, and high uncertainty and time variability in the optical parameters.
Parameter estimation: The main challenges in this component arise from the uncertainty associated with the media optical parameters and the need for computationally efficient forward and inverse models for light propagation in a highly dispersive media.
In addition, examples in the literature [41] show the need to look at non-Gaussian models to model variability in hyperspectral data. This will have an impact in the underlying statistical models used for object recognition and parameter estimation using hyperspectral data. Thus we need to take a deeper look at the statistical modeling of hyperspectral data and the incorporation of these models into recognition and estimation algorithms.
As we discussed previously, accurate measurements of land/surface properties from space rely on precise removal of the influence of the atmosphere on the measured signal. Atmospheric correction algorithms estimate atmospheric effects ignoring surface effects. Once the atmospheric effects are eliminated, the remaining signal is used to estimate the surface contribution to the measured spectra. This decoupling can lead to high uncertainty in the retrieved surface parameters [42]. Therefore it is of importance to develop (or incorporate from the literature) models and corresponding inversion procedures that take into consideration this coupling. This will further improve our ability to tackle challenging problems in many environmental applications where the media has challenging optical properties and a challenging interaction with the objects of interest such as benthic habitat monitoring and management in the presence of turbid waters.
4 Component Accomplishments
• Developed of a new iterative algorithm for unmixing of hyperspectral data that solves the fully constrained (positivity and sum to one) unmixing problem.
• Finished implementation of a covariance estimation algorithm using regularization that improves classification accuracy enabling better handling of the problem of insufficient training samples for quadratic classifiers.
• Applied of classifiers developed during Yrs 1-3 to Benthic Habitat classification in Enrique Reef.
• Finished first version for public release of the Hyperspectral Image Analysis MATLAB Toolbox. This toolbox has all the validated algorithms from previous years for supervised and unsupervised classification, spatial spectral integration, band subset selection, and feature extraction using projection pursuit.
5 Technical Summary
The main objective of this project is the development of techniques for information extraction from hyperspectral remote sensing. Research sponsored under the NASA grant focused on three areas this year:
• Integration of spatial and spectral information for hyperspectral image classification
• Spectral unmising
• Hyperspectral image analysis toolbox development
A. Integration of Spectral and Spatial Information in Hyperspectral Classification
Our main objective is the development of tools for detection and classification in hyperspectral imagery that can deal with the high dimensionality of the image and take full advantage of the information in the spectral and spatial domains. Work this year focused primarily in the integration of spectral and spatial information in hyperspectral image classification.
A..1. Introduction
Hyperspectral images are characterized by large amounts of data spread over several spectral bands. Traditionally these images are classified by spectral information alone [1] employing a classifier. There are some works reporting the performance of spatial methods in classifying these images. Gaussian Markov random fields have been used to model texture information in these images [2]. Filter banks have been used in [3] and [4] for recognition of regions. In this work we present a set of statistical and multiresolution features for improving hyperspectral image classification.
A..2. Spatial Texture Features
Let[pic], represent a hyperspectral image where N1 and N2 and B are the number of rows, number of columns and number of bands, respectively. The spatial texture features are a set of 6 statistical and multiresolution features that capture the gradient, directional variations and the residual energies of texture regions in a spectral band. The texture features include the average (f1), the standard deviation (f2), the average deviation of gradient magnitude (f3), the average residual energy (f4), and the average deviation of the horizontal (f5) and vertical directional residuals (f6) of pixel intensities given by
[pic], [pic],
[pic],[pic],[pic],
and
[pic],
respectively ([pic]is the mean value of the spectral texture). These features capture the edge information of the texture and are, therefore, useful in characterizing the coarseness and randomness of the texture.
The second set includes multiresolution features computed from wavelet decomposition of texture regions. A 2-level wavelet transform of the region is constructed by convolving it with a wavelet filter of length n. This produces 4 sub-images: the approximate, vertical, diagonal and horizontal detail images. The approximate sub-images are further decomposed in a similar manner; specifically, the sub-images at level r are half the size of the sub-images at level r-1. In this work, wavelet features are the wavelet coefficients of the sub-images.
A..3. Hyperspectral Image Classification Algorithm
Given a set of training samples for the classes present in the image, the algorithm partitions it into two sets – training set and testing set. Spectral features are the pixel values in each band. Texture features as described above are extracted for each of the samples. A texture class is represesented by the mean feature vector of the training samples of that class. A testing samples is classified to a particular class using the minimum distance obtained from the distance metric [pic]where fp is the feature vector, [pic]is the feature vector for texture class i,[pic]is the feature vector for texture class j, and P is the number of features. (( fp ) is the standard deviation of the features over the texture classes.
[pic] [pic]
(a) (b)
Fig. 7.2.1.1. (a) Subset of Moffett field image with ground truth classes (RGB bands 10,17 and 25), and (c) Classification using spatial method.
TABLE 7.2.1
CLASSIFICATION RESULTS OF AVIRIS MOFFETT FIELD DATASET USING SPATIAL METHOD
|Classes |Accuracy |Number of testing pixels |
|Clear water (1) |100 |973 |
|Sparse vegetation (2) |100 |1349 |
|Water (3) (not as clear as 1) |88.75 |5303 |
|Water (4) (with suspended solid material) |90.18 |2272 |
|Evaporation pond (5) |91.67 |1239 |
|Total |94.12 |11136 |
A..4. Classification Results and Discussion
Fig. 7.2.1(a) shows a 256x256 subset section of the original image of the Moffett field, California at the southern end of San Francisco Bay acquired by JPL’s AVIRIS sensor on August 20, 1992. This image has 224 bands in the 0.4 (m to 2.4 (m spectral range. The 5 classes are clear water, sparse vegetation, not very clear water, water with suspended material and evaporation pond. After eliminating the water absorption and noise bands, 126 bands are used in this experiment. The spatial features are computed from a total of 226 8x8 training samples using Daubehies wavelet filters. The feature selection process is applied to the original feature set of length 630, resulting in a set of 9 optimal features. The first is the spectral feature from band 29 and the rest are wavelet horizontal and vertical energy features computed from bands 27, 29, 97, 112 113, 115, 116 and 117. A pixel by pixel classification map is generated using 8x8 scanning windows with 7 pixel overlap. Fig. 7.2.1(b) shows the resultant map. The classification accuracies for the testing sites (different from training samples), with the number of pixels for each class are given in Table 7.2.1. The total accuracy for this experiment is about 94.12%.
Spatial texture features have been effectively computed and combined with spectral features to improve classification of hyperspectral images. It is shown that spatial extents of spectral channels contain higher discriminatory textural information. The Daubechies wavelet features and statistical features have performed best in all the experiments. In general, use of spatial features for classification requires lesser number of training samples. A simple distance metric has been used in this work, parametric classifiers such as maximum likelihood and neural network architectures can also be utilized with these features.
A..5. References
1] D. A. Landgrebe, Signal theory methods in multispectral remote sensing, Wiley, NJ, 2003.
2] G. Rellier, X. Descombes, F. Falzon and J. Zerubia, “Texture feature analysis using a Gauss-Markov model in hyperspectral image classification, IEEE. Trans. GRSS, Vol. 42, No. 7, pp. 1543-1551, July 2004.
3] P. S. Hong, L. M. Kaplan and M. J. T. Smith, “Hyperspectral image segmentation using filter banks for texture augmentation,” Proceedings. IEEE Workshop on Advances in Techniques for Analysis of Remotely Sensed Data, pp. 254-258, 2004.
4] M. Shi and G. Healey, “Hyperspectral texture recognition using a multiscale opponent representation,” IEEE Trans. GRSS, Vol. 41, No. 5, pp. 1090-1095, May 2003.
B. Spectral Unmixing
The spatial resolution of most HSI flown nowadays is larger than the size of the objects being observed; in this case the high spectral sensor resolution can be used to detect and classify subpixel objects by their contribution to the measured spectral signal. The problem of interest is to decompose the measured reflectance (or radiance) into its basic elements. This is the so-called unmixing problem [29,30] in HSI. Spectral unmixing is the procedure by which the measured spectrum of a pixel is decomposed into a collection of constituent spectra, or endmembers, and a set of corresponding fractions or abundances.
B.1 Mixing Model
Analytical models for the mixing of materials provide the basis for the development of techniques to recover estimates of the constituents and their proportions from mixed pixels. In the typical mixing process, the surface is portrayed as a checkerboard mixture. The incident light interacts with the surface and the received light conveys the characteristics of the media in a proportion equal to the area covered by the endmembers and the reflectivity of the media. The typical model for such interaction is given by [1]
|[pic] | |
| |(1) |
where [pic] is the pixel of interest, [pic] is the spectral signature of the i-th endmember, xi is the corresponding fractional abundance, w is the measurement noise, m is the number of spectral channel, and n is the number of endmembers. The matrix [pic] is the matrix of endmembers and [pic]is the vector of spectral abundances. Notice that all elements of b, A, and x are constrained to be positive and [pic]. Typically m>n so we are dealing with an overconstrained linear system of equations. This is called in the literature [1] the linear mixing model (LMM).
B.2 Abundance Estimation
For most applications, the measurement noise vector w in (1) is assumed to be independent identically distributed white Gaussian noise. That is
w~N(0,(2I)
where I is the m(m identity matrix and (2 is the noise variance. The maximum likelihood estimate of x based on b is then given by
|[pic] | |
| |(2) |
where [pic]is the estimate of x, and || ||2 is the Euclidean norm. We will refer to (2) as the fully constrained unmixing problem. We could have arrived at LS minimization by plain curve fitting with no statistical considerations.
In the literature, solutions presented to the unmixing problems partially enforce the constraints on the abundances. If no constraints are enforced in (2), the unconstrained least squares problem results for which direct solutions are available (see [2]). Direct solutions are also available for the sum-to-one constrained problem (see also [2]). In the non-negative least squares (NNLS) problem, only the positive constraint is enforced. For the NNLS problem, only iterative methods can be used. The most used algorithm for the NNLS is an active set strategy of Lawson and Hanson described in [2]. Multiplicative iterative algorithms are an alternative for the solution of the positively constrained abundance estimation problem are discussed in [3]. A penalty method to solve the fully constrained problem is proposed in [4].
This year we developed an approach that solves the fully constrained problem exactly by solving a least distance problem [5,6].
4 B.3 Abundance Estimation as a Least Distance Minimization Problem
To derive the proposed algorithm, we will transform (2) to a least distance minimization problem. The first transformation is based on the QR decomposition of the endmember matrix A given by
[pic]
where Q is an orthonormal matrix and R is an nxn upper triangular matrix. Define
[pic]
where c1 is an nx1 vector. Using this transformation, (2) can be transformed to [2]
|[pic] | |
| |(4) |
We relaxed the sum to one to sum less than or equal to one since it allows the inclusion of a dark endmember easily. It is easily to shown (see [2]), that (2) can be transformed to an inequality constrained similar to (4). Define
|[pic] |(5) |
Substituting (5) into (4), we get
|[pic] | |
| |(6) |
where
[pic], [pic]
and 1 is an nx1 vector of ones. This is a least distance (LD) problem [2]. The solution of the least distance problem using a dual method is based on the following equivalence between a LD and a particular NNLS problem [2].
Theorem: Consider the LD problem (6). Let u be the solution of the NNLS problem
|[pic] | |
| |(7) |
where
[pic]
Let the residual corresponding to the solution be
|[pic] |(8) |
and put (=||r||2. If (=0, then the constraints g(Gz are inconsistent and (6) has no solution. If ((0, then the vector defined by
[pic]
is the unique solution to (6). Proof: See Lawson and Hanson [2].
Once z is obtained, the abundance estimates are obtained from (5) as follows
|[pic] |(9) |
To solve (7), we can use any of the algorithms for NNSL described in [3] and [2].
B.3.1 Experimental Results
To compare the effects on the abundance estimates of the constraints, we use a Hyperion Image taken over Enrique Reef in La Parguera, southwest Puerto Rico. The Enrique Reef data used for this experiment consists of first 92 bands from 435-890nm with a spatial resolution of 30 meters. The data was pre processed with Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) to remove atmospheric effects in the data. Figure 7.2.2 shows an aerial photo of La Parguera and highlights the location of the reef.
[pic]
Figure 7.2.1.2: Aerial Photographic of La Parguera, Lajas, Puerto Rico
[pic]
Figure 7.2.1.3: Enrique Reef Image from Hyperion Sensor
[pic]
Figure 7.2.1.4: Enrique Reef Image taken by IKONOS Sensor (1-meter resolution)
Figure 7.2.1.3 shows a color composite image using the Enrique Reef HSI data and Figure 7.2.1.4 shows an image taken with the IKONOS sensor at 1 meter resolution. Clearly, we can see the degradation caused by the low spatial resolution of the Hyperion sensor (Figure 7.2.1.3) when compared with the IKONOS image (Figure 7.2.1.4). In addition, Figure 7.2.1.5 shows the spectral response of the selected endmembers: thalassia (sea grass), reef flat, rhizophora (mangrove), sand lagoon and sea water. The Pixel Purity Index (PPI) method of ENVI was used to extract the endmembers from the Enrique Reef HYPERION image.
[pic]
Figure 7.2.1.5: Enrique Reef Endmember.
[pic]
Figure 7.2.1.6: La Parguera Classification Map.
B.3.2 Abundance Estimation
The unconstrained (ULS), sum-to-one (STOLS), nonnegative (NNLS), and nonnegative sum-less-than or equal to one (NNSLO) were implemented in ENVI and tested with the Enrique Reef data. All comparisons will be visual at this point. Figure 7.2.1.6 shows a classification map of La Parguera area done by the Biogeography Program of the National Oceanic and Atmospheric Administration (NOAA) National Centers for Coastal Ocean Science, and is part of the Center for Coastal Monitoring and Assessment (CCMA) [7]. The method for producing the classifications maps is using aerial photographs to determine what kind of habitats, objects or structures are in the images, we refer to [7] for more information about the classification process.
Figure 7.2.1.7 shows the abundance estimation results obtained for sea grass with ULS, STOLS, NNLS and NNSLO. The region estimated for sea grass is similar to that visually identified in the IKONOS image. We can see the negative values were estimates obtained by the ULS and STOLS algorithms.
Abundance estimation results for sand with the ULS, STOLS, NNLS and NNSLO are shown in Figure 7.2.1.8. We also obtain negative abundances estimates with ULS and STOLS. Also, we can observe that the results obtained with NNSLO are closer to the IKONOS image. The results of the abundance estimates for sea water, mangrove and reef head endmembers were similar behavior with all algorithms. The NNSLO estimates are closer to the IKONOS image.
Figure 7.2.1.9 shows a comparison of the abundance estimates of CFC and NNSLO for the different endmembers. We can observe that both algorithms satisfy the positivity and the sum to one constraint. The estimates of sea grass, reef crest, mangrove and sea water results for both algorithms are very similar.
|[pic] |[pic] |
|(a) |(b) |
|[pic] |[pic] |
|(c) |(d) |
|[pic]>100%[pic]100-81% [pic]80-61%[pic]60-41% |
|[pic]40-21% [pic]20-0% [pic] negative |
Figure 7.2.1.7: Sea Grass Abundance Estimates with (a) ULS (b) STOLS (c) NNLS (d) NNSLO.
|[pic] |[pic] |
|(a) |(b) |
|[pic] |[pic] |
|(c) |(d) |
|[pic]>100%[pic]100-81% [pic]80-61%[pic]60-41% |
|[pic]40-21% [pic]20-0% [pic] negative |
Figure 7.2.1.8: Sand Abundance Estimates with (a) ULS (b) STOLS (c) NNLS (d) NNSLO.
B.3.3 Conclusions
An algorithm to solve the fully constrained abundance estimation problem has been developed. The algorithm is based on transforming the fully constrained abundance estimation problem to an equivalent least distance problem and using a Theorem from Lawson and Hanson that solves this problem using dual theory as a nonnegative least squares problem. We also present result of comparison of different algorithms: ULS, STOLS, NNLS and CFC in unmixing a Hyperion image of the La Parguera, Puerto Rico and compared them with results obtained with NNSLO. As the results shows, the NNSLO algorithm abundance estimates that gives the best visual agreement with a high spatial resolution IKONOS image.
[pic]
Figure 7.2.1.9: Unsupervised unmixing.
[pic]
(a)
[pic]
(b)
[pic]
(c)
[pic]
(d)
Figure 7.2.1.10: Estimated endmembers for different values of the parameter p (number of endmembers): (a) 2, (b) 3, (c) 4, (d) 5.
5 B.4 Unsupervised Unmixing Problem as Positive Matrix Factorization.
For abundance estimation, we need to have previously determined the endmembers (matrix A). Most published methods deal with full unmixing in two stages as shown previously. In the first stage, the endmembers are determined by one of several methods [1]. In the second stage, the abundances are estimated. The endmember identification stage requires significant interaction by the user. It is our interest to develop methods where the endmemembers and their abundances are determined simultaneously with minimal user intervention (unsupervised) as depicted in Figure 7.2.1.9.
In our work, we are focusing on using the positive matrix factorization (PMF) to solve the unsupervised unmixing problem. If we group all pixels of an image in a matrix B, the unsupervised unmixing problem can be stated as finding an integer p, a positive mxp matriz A, and a positive pxn matrix X such that
|[pic] |(15) |
subject to
1- [pic]
2- [pic]
where A is the matrix of endmembers and X is the matrix of abundances. This is a constrained positive matrix factorization (PMF) problem [9]. Initial efforts to solve the fully constrained PMF have been focused on a penalty type method to enforce the constraints. The new cost function takes the form
|[pic] |(16) |
where ((0 is a positive constant.
B.4.1 Experimental Results
Initial results using the Enrique reef Hyperion data are presented here. The estimated endmembers for different number of endmembers are shown in Figure 7.2.1.10.
Estimated endmembers that correspond to pure pixels (water and mangrove) as identified with ENVI PPI previously shown in Fig. 7.2.1.5 are extracted by the approach. Furthermore once a good endmember is extracted it is not discarded even if we increase the number of endmember. Further work is needed particularly in terms of determining the number of endmembers unsupervised.
C. Hyperspectral Image Analysis Toolbox
The Hyperspectral Image Analysis Toolbox is an image analysis software using MATLAB [9, 10]. It is intended for the analysis of multispectral and hyperspectral images. A new version of the toolbox is now available. It provides Supervised and Unsupervised classification algorithms plus additional features like dimensionality reduction, post processing techniques, as well as a class statistics and a pixel’s spectral response utility. Tab panels were implemented for both supervised and unsupervised classification windows. The Class Statistics utility was redesigned visually and functionally. Pixel´s Spectral Response utility was modified to provide proportionality in the plots. Also, in order to provide better data visualization, a 3-D Visualizing utility was created, where the user can rotate any hyperspectral or multispectral image cube and obtain spectral characteristics by looking at contiguous band colormapping. Warning dialogs are now present for certain action confirmations. These dialog windows were implemented taking into account human-computer interaction (HCI) guidelines and software usability techniques. All the GUI features of the toolbox were designed using MATLAB Graphical objects. The development of a user’s manual will also provide new users with samples and guides on using the toolbox. These improvements are of great importance for the release of the first version of the toolbox
Figure 7.2.1.11 shows a functional block diagram of the toolbox and Figure 7.2.1.12 shows an example of the toolbox GUI. The toolbox can be downloaded from
D. References
1] N. Keshava and J.F. Mustard, “Spectral Unmixing.” In IEEE Signal Processing Magazine, pp. 44-57, January 2002.
2] C.L. Lawson and R.J. Hanson, Solving Least Squares Problems, Prentice-Hall, 1974.
3] M. Velez-Reyes, S. Rosario, A. Puetz, R. B. Lockwood, “Iterative algorithms for unmixing of hyperspectral imagery.” In Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery IX, Proceedings of SPIE Vol. 5093, April 2003.
4] Chang C. I.; Ren H.; et. al. ‘‘Subpixel Target Size Estimation for Remotely Sensed Imagery". In Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery IX, Proceedings of SPIE Vol. 5093, April 2003.
5] M Vélez-Reyes and S. Rosario, “Solving Adundance Estimation in Hyperspectral Unmixing as
a Least Distance Problem” In Proceedings of International Geoscience and Remote Sensing Symposium, September 2004.
6] S. Rosario-Torres and M. Vélez-Reyes, “An algorithm for fully constrained abundance estimation in hyperspectral unmixing.” In Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XI, Proceedings of SPIE Vol. 5806, April 2005.
7] Kendall, M.S., M.E. Monaco, K.R. Buja, J.D. Christensen, C.R. Kruer, and M. Finkbeiner, R.A. Warner. 2001. “Methods Used to Map the Benthic Habitats of Puerto Rico and the U.S. Virgin Islands”,
8] Lee, D. D. and H. S. Seung, ”Learning the parts of objects by non-negative matrix factorization.” In Nature, 401 (1999) 788-791.
9] E. Arzuaga-Cruz, L.O. Jimenez-Rodriguez, M. Velez-Reyes, D. Kaeli, E. Rodriguez-Diaz, H.T. Velazquez-Santana, A. Castrodad-Carrau, L.E. Santos-Campis, and C. Santiago. “A MATLAB Toolbox for Hyperspectral Image Analysis.” In Proceedings IEEE International Geosciences and Remote Sensing Symposium, Alaska, September 2004.
10] S. Rosario-Torres, E. Arzuaga-Cruz, M. Velez-Reyes and L.O. Jiménez-Rodríguez, “An Update on the MATLAB Hyperspectral Image Analysis Toolbox.” In Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XI, Proceedings of SPIE Vol. 5806, April 2005.
[pic]
Figure 7.2.1.11: Toolbox functional block diagram.
[pic]
Figure 7.2.1.12: Hyperspectral Image Analysis Toolbox.
1 Future Work
• Continue work on algorithms for integration of spectral and spatial information in hyperspectral image classification.
• Continue work in unsupervised unmixing algorithms. Focus on validation and fast implementation.
• Integrate developed algorithms in the Hyperspectral Image Analysis Toolbox.
2 Technology Transfers
• A new version of the Hyperspectral Image Analysis MATLABTM Toolbox has released to community and described in several publications [9, 10]. Over 100 downloads already.
2 Digital Signal Quantization by Dr. Shawn Hunt
Most signals today are converted into digital form because of the ease of transmission, manipulation and storage. The process of conversion from analog to digital, or from a high resolution digital to a lower resolution digital signal adds quantization noise. This research is investigating new and novel methods of quantization. This includes studies of optimal quantization and also adaptive quantization.
1 Relevance to NASA
All sensor signals are digitized before processing or transmission. In order to reduce the transmission or storage requirements, the minimum number of bits to accurately represent the signal must be used. Having algorithms that reduce the number of bits used by improving the digital representation means that the signals will have smaller transmission or storage requirements.
2 Goals of the Component
• Determining the noise level of a signal before quantization.
• Developing an algorithm to optimally dither a signal off line.
• Developing an algorithm to adaptively dither a signal.
3 Component Accomplishments
• C code has been developed to analyze digital signals. This is being used in the ongoing effort to finish the first component goal.
4 Technical Summary
This research is investigating new and novel methods of quantization. The process of conversion from analog to digital, or from a high resolution digital to a lower resolution digital signal adds quantization noise. Quantization is the process of reducing the amplitude resolution of a signal. This happens when an analog signal is converted into a digital signal, or when a high resolution digital signal is converted into a lower resolution signal.
Quantization is unavoidable, when converting from analog to digital, and so the noise associated with the process, quantization noise is also unavoidable. Digital signals are typically represented in binary digit form, and so the number of bits used to represent each sample is directly related to the amplitude resolution. The greater the number of bits used the higher the resolution. Unfortunately, the higher the number of bits, the greater the memory requirements for storing and transmitting. In general, we want the highest number of bits possible to retain amplitude resolution, but the lowest number of bits to reduce the signal size. These mutually exclusive requirements lead to better methods of quantization.
The total quantization noise added to a signal during quantization is determined by the number of bits used. The distribution of this noise is not determined however, and can be changed according to the signal being quantized in order to minimize its effects.
A typical quantization signal model is
sq[n] = s[n] + q[n],
Where sq[n] is the quantized signal, s[n] the original signal and q[n] the quantization noise. When quantizing a signal, it is generally desired that the quantization noise be independent of the signal being quantized. If they are not independent, this leads to signal related harmonics in one dimensional signals and blocking effects in images. If the signal being quantized is large relative to the quantization noise, then the independence of the signal and noise is generally true. If the signal is small however, then they are generally not independent. In order to improve the independence, or at least the uncorrelatedness, of the signal and quantization noise, a small noise signal is added to the signal before the quantization process. This noise, called dither, can guarantee the uncorrelatedness of the signal and quantization noise. The drawback of course, is that adding the quantization noise increases the noise level. It is generally accepted that the increase in noise level is a small price to pay to eliminate the undesired effects of correlated quantization noise. The noise level is also related to the uncorrelatedness. By adding a white noise signal with a uniform distribution over one half a quantization level will decorrelate the first moment of the signal and quantization noise. This added dither noise increases the total noise level by 3dB. Adding another white noise signal, independent of the first will decorrelate the second moment. This can be done to decorrelate the signal and quantization noise to the desired degree. The disadvantage is that each noise signal adds 3dB to the total noise level, or decreases the signal to noise level by 3dB.
This research has as objective investigate two novel methods for performing quantization of one dimensional signals. First, determine a dither signal to make the signal and quantization noise independent, not only uncorrelated. There have been attempts at finding optimal dither for images, but not for one dimensional signals. Second, investigate adaptive dither in order to change the type of dither depending on the signal being quantized.
The first objective is to find a dither signal to make the signal and quantization noise independent. When quantizing a signal in real time, the best that can be done is to uncorrelate each moment with additional noise. Many signals today, however, are quantized processed and requantized in a computer or computational environment done off line, with the whole signal present. This allows the optimization of the dither noise with respect to the whole signal, and the possibility of reducing. A similar process [1] was used to make the principal components of multispectral images independent, not only uncorrelated as is the general case.
The second objective is to find adaptive dithers that depend on the signal being quantized. As mentioned above, the total amount of quantization noise is determined by the number of bits representing the sample, but the spectral distribution of this noise is arbitrary. This means that the spectrum of the noise can be selected as desired. In audio signal processing, it is typical to select the spectrum of the quantization noise to make it less audible [2,3,4]. The spectrum is designed so that there is less noise where the ear is most sensitive, and more noise where it is less sensitive. This works well, but the noise spectrum remains the same at all times. In an adaptive case, the noise spectrum can be selected so that the signal masks the quantization noise. In other words, the signal being quantized can be exploited to work as a dither signal. This self-dithering has to be adaptive, as it must change as the amplitude and frequency of the signal changes.
The first step in working with quantization is to be able to analyze the signal to be quantized. Fundamental is being able to determine the noise level of a signal before performing the quantization. Since we are dealing with real signals, the signal of interest will always have a noise component. This noise component must be determined in order to determine the minimum number of bits needed to represent the signal without adding any additional noise. In other words, if a signal has a large noise level, then using many bits for representation will only be used to accurately represent the noise. A signal with a 12 bit per sample representation may have a noise level high enough so that 8 bits could accurately represent the signal without adding noise.
It turns out that determining this noise level is not trivial, and there are no references in literature on how to do this. The difficulty comes from the lack of knowledge about the signal. In general the signal is arbitrary, so some assumptions must be made about the noise. Assuming white noise, then any correlation must come from the signal itself. What is to be determined is the level of the noise, so the correlations at different bit levels is what is needed. Looking at each bit level independently will not work because lower level bits can be independent even if they are part of a deterministic signal. The difficulty comes in part from the nature of a digital signal. When a signal with a large amplitude is quantized, there is a large amplitude change from one sample to the next, and so the lower level bit or bits between samples are generally independent. This means that a correlation study of the bit levels will not work. The typical technique of analyzing the signal in the frequency domain can be used, but is inconclusive. This method calculates the lowest levels attained by the signal across the spectrum, and uses this as the estimate of the noise level. This method only works if the signal does not cover the whole spectrum.
5 References
1. J. Kern and J. S. Tyo, "Multivariate Histogram Shaping and Statistically Independent Principle Components," Proc. SPIE vol. 5093: Algorithms for Multispectral, hyperspectral, and Ultraspectral imaging IX, S. S. Shen and P. E. Lewis, Eds., pp. 263 – 274.
2. R. Wannamaker, “Psychoacoustically Optimal Noise Shaping,” JAES Volume 40 Number 7/8 pp. 611-620; July/August 1992.
3. S. Lipshitz, J. Vanderkooy, R. “Wannamaker, Minimally Audible Noise Shaping,” JAES Volume 39 Number 11 pp. 836-852; November 1991.
4. J. Hawksford, “Time-Quantized Frequency Modulation, Time-Domain Dither, Dispersive Codes, and Parametrically Controlled Noise Shaping in SDM, ” JAES Volume 52 Number 6 pp. 587-617; June 2004.
3 Microwave Remote Sensing of the Atmosphere by Sandra Cruz Pol
1 General Overview
Microwave remote sensing work is being performed at the UPRM/ECE Laboratory for Cloud Microwave Measurements of Atmospheric Events (CLiMMATE). Our main goals are the development and calibration of atmospheric and cloud models using data from radar and radiometer measurements. We are currently working with models for stratus clouds, cirrus ice crystals, atmospheric attenuation and precipitation.
Our work includes the development of algorithms for the calibration of the atmospheric models to better retrieve the physical and radiative characteristics of the atmosphere. This includes water vapor content, vertical air motion, liquid water content, and raindrop distribution for clouds and/or precipitation. We also look at the atmospheric attenuation suffered by the radar signal as it travels through the clear atmosphere due to water vapor and oxygen gases. This attenuation varies, among other factors, with frequency, radar scanning angle, air temperature and pressure. We employ data collected with the UMass Cloud Profiling Radar System (CPRS) operating at Ka and W-bands, NASA TRMM PR at Ku-band, NASA EDOP at X-band, NWS WSR-88D at S-band, NOAA wind profiler at S-band, microwave radiometers and NWS radiosondes, among others sources. In addition, we are in the initial stages of developing an atmospheric X-band Doppler radar for weather applications.
Current work includes the uses of NCAR cirrus crystal measurements with radar reflectivities measured with the UMass CPRS at the Blackwell Airport, CO in spring 2000. In this project, we modeled the ice crystals found in cirrus clouds with data measured by an airborne instrument, known as the Video Ice Particle Sampler (VIPS), from the National Center for Atmospheric Research (NCAR). We computed the backscattering of the ice crystals from the ice particle’s size distribution N(D), and compare them with in-situ radar reflectivity measurements. In addition, two Doppler radars working at W-band and at S-band together with radiosonde observations were used to retrieve physical parameters such as rainfall rate and vertical air motion in rain. These instruments were deployed at the U.S. Department of Energy Atmospheric Radiation Measurement (DOE ARM) Cloud and Radiation Testbed (CART) site in Lamont, Oklahoma where rain data was collected on November, 2001.
2 Relevance to NASA
Understanding the dynamics of the lower atmosphere is vital for meteorological studies and weather forecasts. For this purpose a current network of 165 NEXRAD Doppler weather radars operates across the United States. Unfortunately this network cannot detect the lower layers of the atmosphere due primarily to blockage from topography and due to the Earth’s curvature. This work focuses on the calibration of current rain retrieval algorithms for the WSR-88D radar using ancillary data from local rain gauges and from overpassing TRMM precipitation radar.
This research involves fundamental concepts in the areas of remote sensing of the atmosphere, atmospheric attenuation and radar design. Our students have the opportunity to participate in the development stages of new radar algorithms that could complement and improve current technologies.
3 Benefits to Society
The output of our algorithms for precipitation estimation can ultimately feed hydrological models for determining potential mud slide and flooding areas. We are evaluating current algorithms and looking at the regions were current technology cannot adequately sample the lower atmosphere.
4 Goals of the Component
• To calibrate the quantitative precipitation algorithms used by the National Weather Service to retrieve the amount of precipitation in a tropical environment.
• To use NASA Tropical Rain Measurement Mission (TRMM) Ku-band radar data to validate our algorithms together with data from other sources such as the NWS NEXRAD S-band radar in Cayey, Puerto Rico.
• Develop IDL code to read the TRMM data and align it with the other data sets in time and space.
5 Component Accomplishments
• Obtained data from NASA TRMM and rain gauges and in-situ disdrometer for tropical storm Jeanne pass over the island on September of 2004.
• Two different types of DSD were found in PR’s rainfall systems. This validates findings in [1] that DSDs are highly variable between coastal and mountainous zones, even for a small region as the island of Puerto Rico, which posses a very complex orography. The importance of considering variations of Z-R relationships between ocean and land surfaces, as proposed by Ulbrich et al (1999), is therefore preliminary confirmed. This work contributes with supplementary data analysis for tropical environments that has been understood as necessary.
• Conducted study of rainfall rate characteristics during Tropical Storm Jeanne pass over the island of Puerto Rico during Sept 2004. Preliminary results show that overall event varied between Z = 239.8 R1.25 and Z = 443.07 R1.29 compared to currently used Rosenfeld algorithm Z = 250 R1.2 which underestimates or overestimates actual rainfall received.
6 Student Accomplishments
• The student José Maeso performed study of rainfall statistics using rain gauges at the local NWS office in San Juan and a 2D video disdrometer and will be presenting his work at the International Geoscience and Remote Sensing Symposium in Seoul, Korea this summer.
• The alumni Shannon Rodríguez now at NASA Goddard under Gerald Heymsfield group, working in collaboration with UPRM in the development of an X-band polarimetric radar for meteorological applications.
7 Technical Summary
This microwave remote sensing work was performed at the UPRM/ECE Laboratory for Cloud Microwave Measurements of Atmospheric Events (CLiMMATE). Our main objectives are to develop and calibrate atmospheric models using data from radar and radiometer measurements.
• In order to develop a full characterization of rainfall in the lower atmosphere of the island of Puerto Rico we gathered NASA TRMM data from the Ku precipitation radar to validate and calibrate current rainfall rain algorithms used by the local NWS.
• We are in the process of aligning data in space and time to evaluate the algorithm currently being used by the NWS WSR-88D radar (TJUA); Z= 250R1.2, by comparing radar reflectivity and rainfall rate data from the NASA TRMM Precipitation Radar (PR) Ku-band, a 2D Video disdrometer, the National Weather Service S-band NEXRAD radar and local rain gauges.
8 Quantitative Precipitation Estimation using Microwave Sensors
Understanding the role of clouds in the Earth’s heat budget and the radiation transfer processes is vital for global climate models and meteorological studies. This research comprises the areas of remote sensing of the atmosphere, including rain and clouds, using microwave sensors such as radars and radiometers at various frequencies.
Our main goals in this work are to:
• Develop a full characterization of rainfall in the lower atmosphere of the island of Puerto Rico within the rain areas designated by the NWS to develop statistics for pr zones.
• Evaluate the algorithm currently being used by the local National Weather Service (NWS) WSR-88D radar (TJUA); Z= 250R1.2, by comparing radar reflectivity and rainfall rate data from the NASA TRMM Precipitation Radar (PR) Ku-band, the National Weather Service S-band NEXRAD radar and local rain gauges.
• To overcome the problem of the curvature of the earth to study the lower atmosphere where weather occurs, we are planning to set up a small network of X-band radars at key sites to complement the NEXRAD’s measurements.
9 The state of the art
Current algorithms used by the Weather Surveillance Radar- 1988 Doppler (WSR-88D) tend to either underestimate or overestimate rainfall amounts [Vasquez and Roche, 2001]. Several of the reflectivity- rainfall (Z-R) relationships approved by NOAA Radar Operation Center (ROC) to retrieve precipitation rate for non-polarimetric Doppler radars are given below.
[pic]
where R is rain rate and Zh is the horizontal co-polar radar reflectivity factor.
Puerto Rico is a rectangular shape, 230 x 65 km island with complex tropical rainfall events especially difficult to forecast due to the influence of easterly trade winds and to the complex topography of the island. The local NWS office uses the Rosenfeld Tropical algorithm to estimate rainfall rate overall the island. However several studies and personal communications with the local office indicate that this algorithm either overestimates or underestimates the actual precipitation. This is in part due to the complex topography of the island and to the inability of the radar to look at the lower layers of the atmosphere. Another source of uncertainty is that rain gauge data used to calibrate these algorithms provides cumulative point measurements which are hard to compare with the spatial and temporal resolution provided by the WSR-88D [Vasquez and Roche, 2000]. Elsner and Carter, 1997 have divided the island into six regions according to the rainfall statistics using a factor analysis. We plan to use NASA TRMM to further validate and calibrate this algorithm with focus to the western part of the island where the NWS radar cannot sample below 1km of the troposphere due to earth curvature at the given distance from the radar. We will also used local rain gauge network for comparison and in-situ disdrometer measurements. This latter is a 2D Video disdrometer that provides instant in-situ measurements of rain rate, raindrop size distribution, oblate drop shape, among others, and it’s provided by the Colorado State University.
The general expression to relate rainfall rate with radars backscattered power which is proportional to the volume reflectivity, Zh,is given by
[pic]
where N(D) is the drop size distribution also known as DSD, D is the diameter of the raindrops, K is the dielectric factor, λ is the wavelength in air of the frequency of operation, and σhh is the backscattering coefficient for a single drop of size D. One of the main problems with X-band is the high attenuation compared to Ku and S-band in rain [Akeyama et al., 1980]. Rain attenuation makes rainfall estimation difficult. Attenuation induced effects will affect Zh.
To reduce errors in estimation it is necessary to reduce as much as possible the a priori variability interval of the drop size distribution (DSD) parameters. This can be done through microphysical simulation to connect the DSD to different situations and altitudes instead of using a random a priori DSD. A statistical study of rainfall in Puerto Rico should aid us in this respect. In addition, the number of independent measurements at any given region should be large enough and should be compared to other sensor measurements such as the NASA TRMM.
10 Rain rate and Drop Size Distribution Characterization in Tropical Environment
A two-dimensional video disdrometer (2DVD) has been deployed in Puerto Rico northern coastal zone since mid-August 2004. An initial drop-size distribution (DSD) characterization has been performed to compare with previous results of other studies made in the island of Puerto Rico. The event studied was Tropical Storm Jeanne, affecting the region on 15th-16th of September 2004. Preliminary results confirmed that DSDs are highly variable between coastal and mountainous regions, even in a small island (~9000 km2), as suggested by Ulbrich in [1]. In addition, this work intends to improve the reliability of rain algorithms currently used in tropical regions, contributing to better estimation of rainfall rates (R). The expected radar reflectivity Z is calculated from 2DVD-measured DSDs and compared with measured Z from the National Weather Service WSR-88D radar (S-band Doppler radar a.k.a. NEXRAD). Different Z-R relationships – used to determine the rainfall amount from measured Z – for each of the two days of the storm are obtained from these calculations. Discrepancies between storm days are adduced to differences in stratiform and convective rain components.
• An initial drop-size distribution N(D) characterization has been performed to compare with previous results of other studies made in the island of Puerto Rico.
• 2-D Video Disdrometer(2DVD)-- N(D) data was used
• Rainfall rate data from NASA TRMM satellite, 2DVD and rain gauges, was compared.
1 Introduction
Drop size distribution (DSD) is the most fundamental component in rainfall estimation techniques, since it governs all rainfall integral relations. Therefore, its accurate estimation for all rain-rates is necessary in order to develop and validate rainfall retrieval algorithms. The Two-Dimensional Video Disdrometer (2DVD) offers a new approach to measuring DSDs [2]. Under low-wind conditions it provides accurate and detailed information about drop size, terminal velocity, and drop shape. It allows for a detailed verification of rain measurements, in some cases more detailed than traditional rain gauges, using the DSD information.
Rainfall rates obtained from the 2DVD are required to be compared with the Rosenfeld Tropical Z-R relationship – defined for use in tropical convective systems – in order to confirm its validity for tropical areas or to calibrate accordingly. These results will contribute to solve discrepancies between different findings regarding radar and ground instruments rainfall estimation. In the past, only 4 out of 21 2DVDs deploy-locations have been in tropical environments.
DSD estimations derived from radar reflectivities have shown huge overestimations of rain-rate, especially because the overestimation of the number of big drops; in tropical climates, the number of drops tends to be more –in number– and smaller –in size– than in moderate climates. On the contrary, previous work on the island of Puerto Rico using a Joss-Waldvogel disdrometer claimed that there is not significant difference between the computed reflectivity from the measured DSD and the one observed by NEXRAD. Therefore the use of the 2DVD in a tropical location will increase the confidence of the information about drops distribution in tropical climates. Puerto Rico is the smaller of the Greater Antilles in the Caribbean, east of the Dominican Republic and west of St. Thomas, one of the U.S. Virgin Islands.
The Collaborative Adaptive Sensing of the Atmosphere (CASA) Engineering Research Center (ERC), who is partially funding this work, is aimed towards creating a new engineering paradigm in observing, detecting and predicting weather and other atmospheric phenomena [3]. One significant part of this new effort is Quantitative Precipitation Estimation (QPE), which pursues to improve the precipitation estimates and enhance the reliability of flood prediction. The results of this work will present important information for the QPE algorithms that will be used in the near future within the CASA ERC. This will yield enhanced rainfall estimations, much needed for the tropical zones communities. Data analysis was performed for rainfall during Tropical Storm Jeanne accumulation affecting the island of Puerto Rico on the 15th-16th of September 2004.
2 Dsd Characterization
Procedure
DSD information is obtained using a program provided by Colorado State University, dubbed FIRM_DSD. This program reads 2DVD proprietary-formatted files and converts them to text files with the following information: time, diameter range, DSD values, and rain rate. For each minute a DSD value is reported for each diameter range, varying from 0 to 0.25 mm to 10 to 10.25 mm. In addition, the computed rain rate is provided for each minute.
To characterize DSDs, two parameters were computed: the mass-weighted mean diameter, Dm, and the normalized intercept, Nw, using
[pic] (1)
[pic] (2)
Both parameters are used to normalize DSDs, reducing the scattering of data points. This is useful in comparing the shapes of distributions with widely different rain rates [4].
In order to compare to results from similar studies, stratiform rain was defined as that with R greater than 0.5 mm/hr and standard deviation from R less than 1.5 mm/hr. Conversely convective rain was defined as that with R greater than 5 mm/hr and standard deviation from R greater than 1.5 mm/hr.
For the 15th of September 712 minutes of data were recorded by the 2DVD, whereas for the 16th of September 970 minutes are available for analysis. To reduce computing times, data was averaged every 2 minutes, for DSD characterization analysis, thus 356 data points were used for the 15th and 485 for the 16th.
B. Dm and Nw values
Table 7.2.3.1 summarizes values found for both and log10; these were as expected.
Figure 7.2.3.1(a) shows results from previous studies on stratiform rain parameters as well as findings from this work (See #11 in the figure). Regarding stratiform rain, about 50% of data points were classified as this type on September 15, while about 33% were selected on September 16, 2004.
Table 7.2.3.1: Parameters Summary
|Day/Rain Type | [mm] |log10 |
|15 Sep 2004 stratiform |1.14 |4.19 |
|15 Sep 2004 convective |1.38 |3.96 |
|16 Sep 2004 tratiform |0.98 |4.61 |
|16 Sep 2004 convective |1.04 |4.80 |
On the other hand, as opposed as found by Ulbrich et al in 1999 in a mountainous region of the island, when continental properties were found in the DSDs, log10 versus plot shows characteristics similar to the Maritime Cluster (see Figure 1(b)).
Since less than 20% of data points were classified as convective rain in any of the two days, these results require further analysis to make it statistically significant.
3 Radar Reflectivity
Determination of proper DSDs – or N(D) – is crucial to calculate reflectivity Z. The integral form
[pic] (3)
is used for this purpose, and defines Z as the sixth power of the hydrometeor diameter summed over all hydrometeors in a unit volume [6]. Radars, such as National Weather Service (NWS) WSR-88D, measure Z and use rainfall retrieval algorithms to determine the amount of precipitation expected, hence the importance of an accurate determination of DSDs. Even if smaller drops are more numerous, calculating the sixth power of D causes that the fewer larger diameter drops contributes more to Z. In many cases DSD models overestimates the number of big drops for tropical climates (see Figure 7.2.3.2), then is clearly understood the reason for rain overestimation in some cases. This makes more important to obtain detailed information about DSDs, especially for smaller diameter drops, as less weight is given to them in this calculation. An accurate number of drops will account for their appropriate contribution to Z.
[pic]
(a)
[pic]
(b)
Figure 7.2.3.1. (a) The value of log10 (with 1s std dev bars) versus from disdrometer data (numbered open circles) and dual-polarization radar retrievals (open squares as marked) for stratiform rain. Dotted line is the least squares fit. (b) As in (a) except data for convective rain. Note that Nw is the 'normalized' intercept parameter and Dm is the mass-weighted mean diameter of a 'normalized' gamma DSD.
[pic]
Figure 7.2.3.2. DSD out of tropical rain, recorded on 30 August 1995, 22:10-22:30, in Lae, Papua New Guinea. MP-DSD (yellow), JT-DSD (red), and JD-DSD (green) indicated. Mean R = 25.24 mm/hr. As shown in the figure all three methods overestimate the amount of large raindrops.
Rainfall rate R and Z have been related through
[pic] (4)
an equation known as Z-R relationship. The WSR-88D precipitation processing system (PPS) converts Z to R using a Z-R relationship [5]. A list of NWS Radar Operational Center (ROC) accepted Z-R relationships s is presented in Table 7.2.3.2. Other relationship are being suggested by former studies made in Puerto Rico, such as [5], but variations in rainfall characteristics made them suitable for some but not all cases.
In Puerto Rico, the Rosenfeld Tropical relationship, Z=250R1.2, is widely used, though lately it has been changed for certain events, following local NWS findings from their research.
With the intention of comparing and determining Z-R relationships, new Rs are calculated from 2DVD DSD information. To calculate R the following equation was used [4].
[pic] (5)
This equation yields R in mmhr-1 when D is in mm, v(D) is in ms-1, and N(D)dD is in drops m-3. Drop terminal velocity v(D) was found from [4].
[pic] (6)
After calculating Z and R using (3) and (5), respectively, it is necessary to determine Z-R relationships between them, fitting data to find coefficients a and b that optimally fulfills equation (4). Base 10 logarithms were found on each side of (4) to make lineal instead of power or exponential fitting.
[pic]
Table 7.2.3. presents the summary of results from the coefficients found on each day for different Z-R relationships. It also shows the differences caused by differences in R between using FIRM_DSD data and calculating R using (5). As FIRM_DSD uses measured fall velocities instead of (6), results account for that difference.
4 Conclusions
Fundamentally, at least two different types of DSD are involved in PR’s rainfall systems. This validates findings in [1] that DSDs are highly variable between coastal and mountainous zones, even for a small region as the island of Puerto Rico, which posses a very complex orography. The importance of considering variations of Z-R relationships between ocean and land surfaces, as proposed by Ulbrich et al (1999), is therefore preliminary confirmed. This work contributes with supplementary data analysis for tropical environments that has been understood as necessary.
More events will be required to validate rainfall types differences found in this work. It was expected to find a bigger convective component in a tropical storm event, so further analysis of events with considerable amount of precipitation are suggested.
Differences in measured and computed fall velocities caused differences in Z-R relationships when using FIRM_DSD rain rate output and calculated R from measured DSDs. Therefore, is necessary to compare both results with rain gauge or other validated data as well as with NEXRAD data, to decide what would be the best path to follow when computing expected Z. It is known that measured velocity will have a strong horizontal component that will affect the measurements, especially in strong tropical storm winds as is this case. For that reason 2DVD literature specifies that it provides accurate data under low-wind conditions.
5 Future Work
Additional significant rainfall data from other events is available; these will be used to compare and validate these results. Data from other tropical regions will be considered and used for comparison purposes. Next July 2005, the 2D video disdrometer will be transported from San Juan to Mayagüez, to be deployed at the UPRM campus, so that rainfall statistics of this region of the island can be studied and characterized.
In terms of Z, some low-resolution data (3-minutes, 5 dBZ intervals) has been obtained from NWS local weather forecast office, and even when it will not serve for exact comparisons, it will impart confidence in our results.
Finally, existing techniques to filter errors that can be introduced by drops coming from different regions of the clouds are being analyzed. As bigger drops have different fall velocities than smaller drops, DSD estimation on the ground could not be in accordance with what radars are measuring above, causing errors in DSDs and therefore in Z computations. These errors are known as observational noise. One of these techniques is the Sequential Intensity Filtering Technique (SIFT) which concentrates on the stability of the R-Z relationship during a physically uniform situation [7]. SIFT, aimed to filter out observational noise, will be investigated applying it to data sets considered here and to future sets as they become available from this ongoing measurement experiment.
TRMM/PR data: The NASA Tropical Rain Measurement Mission (TRMM) was launched in November 1997 to study the 3D structure of rain events from space. TRMM carries among its instruments a microwave precipitation radar (PR) that operates at 13.8GHz (Ku-band). TRMM passes over the island sixteen times every day during 1.14 seconds each time with a horizontal resolution of 4.3 km. We obtained data from rain event in April 2005 to compare with the local NWS radar retrieved rain rate and in the calibration of current rain algorithm for several regions in the island. We are currently working in the development of IDL code to read the data and align it with ancillary data sets.
11 References
[1] Ulbrich, C. W., M. Petitdidier, E.F. Campos, “Radar Properties of Tropical Rain found from Disdrometer Data at Arecibo, P.R.”, 29th International Conference on Radar Meteorology, Montreal, Canada, July 1999.
[2] Kruger, A., W.F. Krajewski, “Two-Dimensional Video Disdrometer: A description”, Journal of Atmospheric and Oceanic Technology, Vol. 19, pp. 602-617, 2001
[3] Collaborative Adaptive Sensing of the Atmosphere (CASA) Engineering Research Center, “Overview” [Online]. Available from .
[4] Bringi, V.N. and V. Chandrasekar, Polarimetric Doppler Weather Radar: Principles and Applications, 1st ed., United Kingdom: Cambridge University Press, 2001.
[5] Vázquez, M. and A. Roche, “An evaluation of WSR-88D rainfall estimates across Puerto Rico during Hurricane Debbie”, NEXRAD Weather Forecast Office San Juan, Carolina, Puerto Rico.
[6] Doviak, R.J. and D.S. Zrnić, Doppler Radar and Weather Observations, 2nd. edition, Academic Press, 1993.
[7] Lee, G. and I. Zawadzki, Sequential intensity filtering technique (SIFT): filtering out noise to highlight the physical variability of drop size distribution. 31st International Conference on Radar Meteorology, Aug 2003.
4 Material type determination of Subsurface Large Objects or Layers Using Ground Penetrating Radar, PI: Dr. Hamed Parsiani
1 Technical Summary
In the previous research work, it was shown that the novel method of the Material Characteristics in Frequency Domain (MCFD) manifested variations as the type of the material was changed. This property was exploited for the proper determination of soil types. The work was accomplished in an ideal laboratory environment.
In the present research, the material type determination was extended from pure soil types of sand, loam, and clay to their 50% mixtures, using GPR at 1.5 GHz. A limited database of MCFDs of the three soil types and their 50% mixtures was created by MCFD-NN in a Graphical User Interface (GUI) format, to be used as a desirable product for the end user.
In this work the material type determination has gone beyond the laboratory environment into open field measurements, using mainly the university baseball field.
It has been noted that the GPR images of soil types in a homogeneous are of land will be affected to a great extent by the rocks and other items present in the soil. A method of elimination of these external materials from soil has been devised.
The direction of the research has been guided by the need to cover rapidly more land using GPR installed at the back of a vehicle. However, the anticipated vehicular motion and the irregularities of the land will clutter the GPR image and will require methods to combat both problems. The effect of the vertical motion of the GPR has been analyzed, and a method is recommended to distinguish between the air reflection and soil reflection, as long as the two are not convolved.
2 Development of a graphical user interface for the determination of soil type, referred to as MCFD/NN-GUI
In this work we are concentrating on the interactions of the GPR electromagnetic wave with the first layer of the ground under analysis. A representation of this interaction is shown in Figure 2.1. The transmitted and received signals are, St , and Sr, respectively. The scattered signal, Ss, is ignored, given that it is difficult to record. It is assumed that the soil behaves as a linear system, as illustrated in Figure 2.2. Here we can see that the problem is simply solved by identifying the unknown soil characteristics. As established in [1],[2], system identification is done in the frequency domain, where Sr(u) = St(u) ∙ M(u).
[pic]
[pic]
Therefore, the Material Characteristics in Frequency Domain (MCFD) is
[pic]
The general methodology is illustrated in Figure 3. A two-layer backpropagation Neural Network was used for this work. The Neural Network’s output is a number which represents the material or mixture of materials that compose the soil in question, given the training set of MCFD as input. Figure 2.4 depicts an example of a training set along with a representation of the Neural Network.
[pic]
[pic]
NN training was accomplished with several samples of sand, loam, clay, and their mixtures. With more samples, the NN is more robust. Good testing results were obtained for each of these material types. The testing process consisted of presenting a sample at the input of the trained NN, with an already known material type which was not used in the training process.
3 Graphical User Interface Development
A Graphical User Interface (GUI) has been developed to make the MCFD/NN algorithm accessible to a potential user. The GUI links the MCFD/NN software along with all its capabilities, implemented in C language, into a single window with several options. The GUI supports input of an unknown soil GPR scan to determine its type and also its moisture content. The language used in the GUI implementation was the Tool Command Language and its graphical toolkit (Tcl/Tk). Figure 2.5 shows the GUI Main Window.
[pic]
The Interface options include as a basis, Material Type or Moisture Content Determination calculation selectors. The user may choose one of these two options by clicking on the checkbutton that is followed by the desired calculation (Material Type or Moisture Content). When performing Material Type Determination, the currently available material types for analysis are sand, loam, clay and their mixtures. The user should know preliminarily if the constitution of the terrain under consideration is mainly a pure material or a mixture. Selection of the approximate soil composition is performed through the list box that appears by clicking the leftmost arrow button, as seen in Figure 2.6. The Moisture Content Determination extension is made possible through additional funding from NOAA.
[pic]
Once the calculation type is selected, the user can proceed to train a NN or to actually determine the Material Type or Moisture Content of an unknown GPR scan. To train a NN, the user must have a training set of already analyzed GPR scans of the terrain studied. NN training is executed by entering the training filename in the upper Filename field, or by browsing method using the Browse button. Next, clicking on the Start Training button begins the training process. The progress bar on the right fills and a message is displayed when training is completed, as shown in Figure 2.7.
[pic]
Actual soil type determination takes place after a GPR scan is entered at the lower Filename field. Again, the user may enter the filename or browse for the file. After all desired options have been selected and a Neural Network has been trained, clicking in the Find Match button runs the procedure to find a NN match. A test performed for a mixture of sand and clay is seen in Figure 2.8. The NN result along with its Field Result (if any, previously stored in a .txt file) is then displayed in a message box as shown in Figure 2.8b.
[pic]
4 Selection of the wavelets in the presence of external materials present in the soil
For the automatic processing of GPR images, an algorithm in C-Language was developed to extract the best possible wavelet from the various scans in a GPR image, based on a few criteria which will be addressed subsequently. These were designed with the purpose of preserving the wavelet main characteristic, which is used in later processing.
The GPR grayscale image as shown in Fig. 3.1 is a collection of all the wavelet signals acquired during the time the radar is active. The signal appears as an array of successive scans (columns) with each scan digitized in 512 (rows) samples. The algorithm functions by detecting the reflection from the boundaries present in the images provided by the GPR. Individual scans as shown in Fig. 3.2 of the same material (soil) can vary based on interference or reflections from other objects such as rocks within the soil.
Evaluation of the wavelet main characteristic allows for a rating system to which each scan is subjected. This system checks that the scan contain three identifiable maximums and minimums (Fig. 3.3), verifies for attenuation greater than the air reflection (Fig. 3.4), and confirms the presence of a separation between the boundary of the desired wavelet and the beginning of noisy signals due to reduction of the signal to noise ratio at the deeper levels (Fig. 3.5). The algorithm evaluates these three characteristics for each scan and extracts the wavelet which best fits these parameters.
[pic]
This wavelet may have been subjected to perturbations caused by the effect of reflections in the soil due to rocks and other non-soil items, denoted by the addition of sharp curves to the signal in the time domain, as shown in Fig. 3.6. To alleviate this, the wavelet is passed through the Discrete Fourier Transform at which point the algorithm analyses the frequency content of the signal. In Fourier Domain, the perturbations are easily visible in the frequencies approximately above the three gigahertz (Fig. 3.7). The algorithm then utilizes an Ideal Filter to remove the higher frequency content, thus reducing the effect of the unwanted reflections. Both an Ideal Filter and a Butterworth Filter were tested for the adequate removal of this content, but the former proved more suitable as it did not change the signal due to the soil type.
[pic] [pic]
After the removal of the high frequency content, the signal is passed through the Inverse Discrete Fourier Transform and is processed by the previously mentioned three criteria to evaluate the wavelet main characteristic. Based on the results of the aforementioned evaluation, the wavelet is either kept or discarded by the algorithm. The complete process is shown in the flow chart of Fig. 3.8.
[pic]
5 Preliminary analysis of the effects of GPR vertical motion on soil images
One of the difficulties in a rapid open field soil type determination is the cluttering of the GPR image due to the vehicular motion, as depicted in Fig. 4.1
[pic]
The Vertical motion of the GPR creates signals where the air reflection signal and the soil reflection signals are well separated, Fig 4. 2, or they are very well convolved, or in various degrees of mixture.
[pic]
6 GPR at its high level
If the antenna is held high enough over the soil, the received signal will be composed of two main reflections, the air reflection and the reflection from the soil, and other reflections. The goal is to separate the reflections from the air and the soil. To accomplish this task, an experiment is performed by correlating a standard air signal with the actual received signal (reflection from the soil). The correlation function is defined by
The maximum value of the correlation function will tell us when maximum similarity between the standard air signal, y(n), and the actual received signal, x(n),occurs.
This allows us to extract out the air signal and look for the next reflection, which should be the reflection from the soil. This operation is depicted in figures 4.3, 4.4 and 4.5. Although not yet thoroughly tested, this method shows promise as a tool to separate the air and soil signals.
7 Conclusions
Material Characteristics in Frequency Domain is a vital tool for Material Type Determination, since it is possible to extract and analyze these characteristics to discriminate between different soil types in a terrain using their respective GPR scans. This technique is non-invasive and less time-consuming than current methods of Material Type Determination. A MCFD/NN/GUI approach has been accomplished for sand, loam, clay, and their mixtures with a satisfactory outcome. A method is devised which requires further improvements in the recognition of appropriate wavelets reflected from relatively homogenous soil cluttered with external objects. An analysis of the images with GPR motion was commenced.
8 Future Work:
A larger database is necessary for the improved usage of the MCFD/NN-GUI algorithm. The automatic wavelet extraction requires further testing with more rocks and other objects mixed in soil. The rapid imaging of the land with GPR installed is an important task which requires proper extraction of wavelets from the clutter due to both vehicular motion and land irregularities.
The usage of 1.5 GHz and the recently acquired 2.2 GHz antenna in the clutter remover due to motion and irregularities based on sensor fusion.
9 Summary of proposed work for next year:
• Rapid open field determination of soil characteristics, by attaching the GPR antenna to a moving vehicle.
• Elimination of the clutter caused by motion and land irregularities using mathematical approximations, and two antenna approach (at 1.5 and 2.2 GHz)
• Collaboration with the GSSI in the open field measurements
• Improvements of the database for the MCFD/NN-GUI algorithm
• Publish the improved results in a peer-review journal.
[pic]
[pic]
[pic]
Fig. 4.5. Extracted soil reflection
10 Reference:
[1] Hamed Parsiani, Maritza Torres, Pedro Rodriguez,” Material Characteristics in Fourier Domain (MCFD) formulation, a signature to determine soil type, moisture, and vegetation health, based on multilayer ground penetrating radar reflection”. Proceedings of IASTED SIP 2004, Aug. 23-25, 2004, Hawaii.
[2] Hamed Parsiani, Pedro Rodriguez, “Subsurface Material Type Determination from Ground Penetrating Radar Signatures”, Proceedings of SPIE 2004, Gran Canaria, Spain, Sept. 13-16, 2004.
5 TerraScope: Dynamic Image Retrieval and Composition Services in Distributed Information Systems by Bienvenido Vélez, Manuel Rodríguez and Pedro I. Rivera
1 Problem Statement
The emergence of large-scale distributed environments, such as the Internet, has provided access to vast collections of satellite images, hyper spectral images, GIS data, measurements and other kinds of data products essential to carry out high impact research in many scientific and engineering disciplines [11,17,19]. Many universities, research laboratories and government agencies provide access to these data sets using a Web-based interface, which allows the user to select data products from a pre-defined set of available data holdings [17]. However, these solutions provide limited support for novel services that:
• Efficiently integrate data from multiple data sources into a single data stream that can be used by institutions to create new data products and services.
• Enable end-users (e.g. DSP engineers) to supply custom tailored processing algorithms that can be used during the process of generating new image products.
• Respect the administrative autonomy of information providers while at the same time supporting collaboration to provide integrated access to all available data.
The TerraScope system embodies our vision of a novel Distributed Information System where multiple data sources (e.g. hyper spectral image databases) can be integrated into a coherent system to harvest the data products that each data center stores. These data products may include images, metadata, GIS features, and any other relevant type of information. In this system, new data products can be generated on demand from base products, and users can supply their own application-specific code to assist in the dynamic generation of the new data products. The system should support operations such as:
• Retrieval and visualization of images such as hyper spectral images, MODIS, AVHRR and ETM+ from multiple data sources. This could be extended to images obtained from different sensor modalities.
• Dynamic sub-setting (slicing) of images to create mosaics from two or more images extracted from different data sources, or generated on-the-fly by a previous query.
• Overlay of images to form composite images with different levels of details on each layer and higher information content.
• Dynamic generation of MPEG, QuickTime or Flash movies that illustrate changes in the time domain. Examples of such changes are coral reef destruction, growth of internal tumors inside live organisms, changes in ocean bottom topology, and other kinds of natural phenomena captured via different sensor modalities and subsurface sensing techniques.
2 Relevance to NASA
According to the NASA Administrator, Honorable Sean O’Keefe, NASA’s mandate is: to improve life here, to extend life to there and to find life beyond. To aid in the achievement of this mission, NASA continuously maintains one of the largest collections of multi-media data under both its Space Sciences and Earth Sciences enterprises. Information is collected both internally by NASA as well as by NASA-supported research centers across the world such as the Tropical Center for Earth and Space Studies hosted at the University of Puerto Rico Mayagüez. This information is used to reach into the depths of hidden world not reachable by more standard means including deep space and deep earth.
Such information can only be of value if it becomes accessible to the scientists and engineers who in the end can use it to carry out their research and generate new knowledge. A system like TerraScope could serve as an important vehicle for the effective sharing and dissemination of this valuable information. TerraScope’s design goals have been chosen with this objective in mind. For instance, the system follows a peer-to-peer architecture in which any information provider serves information to both clients as well as other information providers. Each TerraScope server is capable of handling queries and contacts other peer servers to gather all information relevant to these queries. This type of architecture facilitates load balancing among all available peer servers and avoids single points of failure that may render the system inaccessible and are the norm in more centralized approaches including integration servers. Furthermore, the TerraScope user interface will support dynamic data product composition potentially using user-provided data processing algorithms.
3 Benefits to Society
The TerraScope System has been designed using only standard and open protocols and off-the-shelf components available in multiple platforms. This choice provides the greatest opportunities for the wide proliferation of tools and information among data providers and users. For instance, the TerraScope Image Navigator (TIN) was completely developed using Macromedia Flash MX, a tool originally designed for interactive movie authoring. Users wishing to access TIN can do so from any web browser like Internet Explorer or Netscape Navigator and only need to install a freely available Flash player to run the application. Thus the required software tools are available on virtually all major platforms including MS Windows and Linux.
Open standards and protocols facilitate the dissemination of NASA information beyond NASA’s scientific partners. Society at large can greatly benefit from having access to user friendly tools and information collected by NASA. In particular, K-12 students can use this information to learn about the earth and the universe in a way amenable to individual creativity and exploration that extends beyond simple navigation of a static repository of information. Kids can exploit TerraScope’s support for dynamic image and video subsetting and composition in order to, not only find, but create the answers to their inquiries.
4 Goals of the Component (Short Term)
• Design a middleware framework to support a peer-to-peer distributed image database system
• Design a web-enabled client capable of supporting image browsing and retrieval via the Internet and using standard protocols and off-the-shelf software components.
• Disseminate research results in peer-reviewed conferences and thru the web
5 Component Accomplishments
• Designed one or more algorithms to classify satellite images in order to effectively support browsing.
• Redesign the TIN GUI to support effective image clustering-based browsing and searching.
• Design various alternative image clustering algorithms
• Measure the performance of the various clustering algorithms.
• Measure the effectiveness of the new Graphical User Interface (GUI).
• Published articles in peer-reviewed scientific conferences.
• Conducted several presentations at both local and national levels.
• Developed a website holding all information and software relevant to the TerraScope project.
• Developed various proposals requesting supplemental funding for the TerraScope project.
6 Student Accomplishments
The Terrascope subcomponent of the IPEG component supported student Lizvette Malavé. Lizvette completed his MS dissertation and obtained the MS degree in Computer Engineering in May of 2005. She published her work in a paper entitled “Terrascope Image Clustering: Applying Clustering Techniques to Image Agglomeration in Image Retrieval Systems” at the Third IASTED International Conference on Communications, Internet, and Information Technology (CIIT 2004). Lizvette herself presented her work at the conference.
7 Technical Summary
The emergence of multimedia technologies and the possibility of sharing and distributing geospatial satellite data through broadband computer networks have exacerbated the need for geographical information technologies. In one typical scenario, geospatial data is periodically collected at geographically distributed ground stations which often span several technical, administrative and even political domains. Solutions attempting to offer integrated access to distributed data by relying on centralization of data repositories are often not feasible.
Frequently, scientists around the world need to effectively access information necessary for research. Data Many satellite data collector centers are distributed around the world. They provide web access to their databases of satellite images and other geographic useful for the earth. Terrascope is a new system developed at the University of Puerto Rico-Mayagüez Campus to provide effective and transparent access to these inherently distributed images.
The inherently distributed nature of geospatial data should be of no concern to scientists, students or other data consumers. The information system should enable and facilitate collaboration among a set of distributed data providers who wish to collaborate to provide an image of a single data repository with minimal loss of their individual autonomy over their data. Based on this principle we are designing and developing the TerraScope Earth Science distributed peer-to-peer database middleware system conformed by a Search and Retrieval Engine (SRE) [33] and TIN, the graphical user interface (GUI) module.
Terrascope follows the client server architecture depicted in Figure 1. The server module is called the Search and Retrieval Engine (SRE) [1] and the client module is called the Terrascope Image Navigator (TIN) [2]. The server (SRE) consists of a set of JAVA servlets modules running inside a web server. These modules implement an abstraction of a single data repository by communicating with multiple TerraScope SRE’s peers. The SRE computes the set of results by potentially forwarding queries to other SRE’s believed to hold data pertaining to the query specified by the user. Hence, SREs act as servers to TIN clients, and also as clients of other SREs. This type of architecture is often called peer-to-peer. For more details of the SRE and TIN, the reader is referred to [1] and to [2].
[pic]
Figure 1. Terrascope Architecture
The TerraScope Image Navigator (TIN) provides the end user graphical interface to access the collection of SREs. It is an image browser that provides ubiquitous and efficient access to distributed information. The TIN prototype delivers satellite images with their corresponding metadata, GIS characteristics, and other information to any web browser with a Flash MX player installed. Initially, TIN displays a map showing the overall geographical region covered by the satellite ground stations contributing data to the distributed repository (see Figure 2). The current prototype includes data collected by the Tropical Center for Earth and Space Studies (TCESS), the Center for Subsurface Sensing and Imaging Systems (CenSSIS) both at the University of Puerto Rico Mayaguez Campus and data collected by the Aster sensor from the Southern Urals Region of Rusia. Using familiar GUI controls TIN users can restrict the scope of their search to a specific data repository, geographical region, type of satellite sensor (e.g. MODIS, RADARSAT and Landsat 7) and data collection date.
TIN displays polygons of the images found in the database corresponding to the user query and automatically geo-references these images within the query image (see Figure 3). The user can click on the geo representation he/she is interested in and the image will be displayed in the main window. Once an image of interest is selected and displayed, users may easily search the database for images contained within this image (sub-images) or for images overlapping it (see Figure 4). In other words, each browsed image provides a geospatial context from which future exploration may proceed; a feature that we call recursive navigation.
[pic]
Figure 2. TIN’s Initial Window
[pic]
Figure 3. Geo Representation of images retrieved by TIN in response to a query
Previous user studies have evidenced that geo-representation of images helps users visualize the location of images and provide an effective way to present this kind of data. However, when the database consists of multiple highly overlapping images, geo representation of images may worsen the ability of users to find images.
[pic]
Figure 4. Recursive Navigation Example in TIN
During the reporting period we focused our research on image clustering techniques that could be applied to ameliorate the image clobbering problem. The databases that Terrascope is intended to support contain many images that overlap one another because the satellites that recollect the images are continuously rounding the same geospatial region. The system must support temporal queries across all such overlapping images. As shown in Figure 5, the problem is that the user is unable to access all the data requested when the images overlap and can not request to download the images hidden by other images. In order to access the data the user may be forced to make different queries to retrieve less information per query. Coming up with suitable queries to get rid of unwanted information can be often a tedious and time consuming process.
[pic]
Figure 5. Example of Image Overlapping in TIN
Another problem confronted by the user is that properties of data cannot be effectively visualized in the result set. For instance, the results in Figure 5 come from different sources and sensors, but this is not easily identified by just looking at the geo-referenced polygons.
[pic]
Figure 6. Initial Context Geographical Area
TIC exploits image clustering in order to automatically classify the retrieved images into smaller groups and present them to the user in a way that facilitates browsing and finding images of interest. The central hypothesis of our work is that image clustering can be used to improve the effectiveness of image retrieval/browsing systems with similar characteristics to TIN. We test our hypothesis by conducting two types of experiments. The first one is a comparison of several image clustering algorithms and the second is a user study. Our experimental results provide evidence in favor of the effectiveness of TIC to improve the presentation of results and the ability of the users to visualize the properties of data.
TIC: A Re-designed Graphical User Interface to Support Clustered Image Browsing
The clustering provided by TIC maximizes the benefits of TIN’s image navigation. Initially, TIC displays a map from where the user may select the search area (see Figure 6). The user may use the GUI controls to explore the areas with zooming and panning, or he can use the controls to make queries.
The user can make restrict the scope of the query by selecting the area with a rubber band tool. The query can be modified selecting the sensors or sources from where the user needs the data. Also, the user can do temporal queries selecting a specific date or range of dates (see Figure 7).
[pic]
Figure 7. Selecting a search area with the rubber band tool
After submitting the query to the server, TIC parses the XML result message and automatically geo-references the retrieved images into the previous displayed image using their corresponding geographical metadata (see Figure 8). In addition to the visual representation of the result set, a text representation is displayed on a separate panel. Each image in the result set is identified by a unique id number that is extracted from the image metadata. Note that the result set on Figure 8 only contains images which overlap with the selected search area.
[pic]
Figure 8. Geo-referenced and Textual Result Set
The geo represented polygons are rendered as buttons. The buttons and the different panels are synchronized. If the user rolls over a polygon it changes colors and metadata of the corresponding image is displayed on the lower panel. If the user cannot rollover an image because it has other images that overlap, the user may select an image from the list, the system responds by highlighting the geo-representation and displaying its metadata.
From the display in Figure 8 the user may proceed in one of three ways. He/She may submit a different query after realizing that the results were inadequate, he/she may use the GUI controls to explore the areas with zooming and panning, or he/she may click on one of the embedded polygons in order to download the image and navigate into it (recursive navigation) (see Figure 10).
Recursive navigation facilitates the exploration of smaller regions using higher resolution images. Once an image of interest is downloaded, the user can explore the area by using zooming and panning tools or can save the image into their hard drive. The user can also navigate into the area using GUI controls to make another query. The selected image becomes the new geographical context and the query is automatically restricted to this area. The query can be further refined using the controls, searching by source, sensor, and/or dates.
[pic]
Figure 9. Image Selection
[pic]
Figure 10. Recursive Navigation
The geo representation is a graphical way to work with some of the characteristics (latitude and longitude points) included in the image metadata but the images have other characteristics such as source, or satellite type that are not easily target by just looking at the result set. Using TIC the user can have a graphical and structured view of all the data by making use of the clustering options that take advantage of the images metadata. This process requires no interaction with the server and no human pre-processing.
TIC’s user interface provides several options to categorize data. When the user categorizes images, the list of the result set becomes a tree. Each node gathers images with similar characteristics according to the user request. The following are the options currently supported by TIN to cluster the data:
• By Source: TIC clusters the result set according to the source database that holds the image. Each node corresponds to a source or database that contributes data to the repository. In the example presented in Figure 11 the images are classified by source. The sources that currently contribute data are CENSSIS, TCESS, and LARSIP. The user can select the node labeled LARSIP to display only the images held by this source.
• By Sensor: TIC clusters the result set according to the kind of sensor that was used to collect the image. In the example presented in Figure 12 the images are classified by sensor. The images where taken using the sensors: RSAT and LSAT. For instance, the user can select the node labeled LSAT to display only the images within the result set taken by this type of satellite sensor.
• By Date: TIC clusters the result set by date for the first one to the latest one. Each node label identifies the date of the images contained in the node. In the example presented in Figure 13 the images are classified by date. The user can select the node labeled Sat April 15, 2000 to display only the four images taken that day.
• By Minimum Overlapping: TIC clusters the result set into approximately ten (10) nodes in a way that minimizes overlapping inside each node.
[pic]
Figure 11. Image Clustering by Source
[pic]
Figure 12. Image Clustering by Sensor
.
[pic]
Figure 13. Image Clustering by Collection Date
Each branch’s label contains the quantity of leafs that it contains. All clustering options allow users to conveniently expand or collapse nodes to facilitate focusing on the information of interest. Also, at any time the user can click the button labeled UNCLUSTER to start over with the result set list as seen in Figure 8.
The cluster tree can be changed dynamically and recursively in order to capture the clustering that better suits the user’s information need. As shown in Figure 14, the user may select to group images by source (peer) and group again by sensor types within each source group. When the user selects one group (or subgroup), only the images contained in it will be visible in the map; the others will be hidden
Recursively clustering the result the user can easily answer complex queries that are commonly needed by scientists. For example, consider the following query “Search all images from any source that took images of America last week”. Using the query parameters of TIN the user can select the area corresponding to America and the dates corresponding to “last week”, submit the query and the result set will be geo represented in the map. This query was suggested to us by a visitor scientist from NASA during a TCESS review board.
After the result set is displayed the user might be interested in getting detailed information of the images. As shown in Figure 14, using TIC the user might easily find out which sources took images of the area and which sensors where used by each source. Also, the user might be interested in knowing which images were taken which day of “last week”. To achieve this, the user may use the clustering options available in TIC. The image taken by CENSSIS using a RSAT sensor on Sun Apr 8 2001 can be easily targeted by traversing the tree. If the user clicks on the polygon the image will be displayed and will become the new geographical context as shown in Figure 15.
[pic]
Figure 14. Recursive Clustering Example
[pic]
Figure 15. Selected image becomes new search context
At any point the user can start over the clustering by clicking the button labeled UNCLUSTER to start over with a flat result set list. Also, the user can make a new query at any point and TIC will automatically update the result set.
Image Clustering Algorithms
In this chapter we describe various clustering algorithms that we have experimented with in order to support TIC graphical user interface. Each algorithm supports clustering of the result set tree to help the user target the desired data. All algorithms alleviate image agglomeration because images grouped in each subgroup have less probability of overlapping. As explained in Chapter 4, images contained within a group can be displayed by clicking on the corresponding node label in the tree.
Data Structures
Terrascope uses an array structure to hold the images in the result set together with their metadata. Each element in the array represents an image. The attributes of each element are the image’s characteristics retrieved from the metadata in the XML message returned by the server. This array structure is used to geo-reference the clickable polygons in the context image.
Terrascope Image Clustering organizes result sets using the structure of a tree. The tree is implemented as a XML object. Each leaf of the tree represents an image and it is represented as a pair of a label and a value. The current version of TIC uses a unique identifier for each image provided by the databases to label the corresponding leaf node. The value of the leaf node is a pointer to the in the result set array. This mapping is shown in
Figure 16. Branch nodes are labeled with a meaningful string that describes the data it contains.
We use a tree structure because it can easily support recursive clustering and it makes easier the indexing used by the clustering algorithms when traversing the tree. The data attribute keeps a mapping from each leaf node to the corresponding element in the results array and to the corresponding geo representation object. This mapping is needed to retrieve the images metadata saved in the array and map it to the geo representations used in the clustering algorithms.
Figure 16. Array-Tree Mapping
Clustering by Source, Sensor and Dates
TIC currently clusters the result sets according to metadata attributes associated with images such as: source, sensor and date to cluster the images. The algorithm removes leafs of the current tree and according to the required clustering attribute it add them to a descriptive branch in a temporary tree. To access the attribute value we map the data value of the leaf in the tree with the corresponding element in the array (see Figure 16). After all leafs are added to the corresponding branch the labels of the branches are updated to display the quantity of leafs they contain and then GUI’s tree is updated. Figure 17 shows the clustering by source algorithm. The clustering by sensor algorithm follows the same idea. The clustering by date differs from the others because it also sorts. Figure 18 demonstrates the clustering by date algorithm.
[pic]
Figure 17. Source Clustering Algorithm
[pic]
Figure 18. Date Clustering Algorithm
Clustering by Minimum Overlapping
When the database consists of multiple highly overlapping images, geo representation of images may worsen the ability of users to find images. We designed several algorithms to address this problem. The first algorithm that we designed segregates the images into clusters with no overlapping (as shown in Figure 19). The algorithm begins with one cluster and iterates over each image in the results array. The algorithm goes trough the available clusters and add the image in the first available cluster where that image does not overlap with the images previously placed in the cluster. If an image overlaps with at least one image in the cluster, a new cluster is created to add the image in.
Figure 19. No Overlapping Algorithm
The result of running this algorithm was that we often obtained many tiny clusters resulting in a very wide tree. The reason for this is that almost every image overlaps with some other image in the result set. We decided to allow some kind of overlapping in order to narrow the tree. We achieved this by restricting the number of clusters while simultaneously maintaining a balance in the amount of overlap across the different clusters.
The number of clusters should not be a fixed number, because if we cluster a large result set into a small number of clusters it may result in too much overlapping. The number of clusters that we use is dependent of the size of the result set. On The number of clusters allowed is calculated using the following formula:
MAX_CLUSTERS = 10 ( log(result set size) (
We developed two algorithms that cluster the result set into a fixed number of clusters with minimum overlapping. The two algorithms are presented in the next sections.
Algorithm #1: Minimum overlapping by collapsing neighbors
This algorithm (see Figure 20) clusters the result set into clusters with no overlapping. Then the algorithm narrows the tree by searching for clusters that are too small and collapse them with neighboring clusters. In our current prototype “too small” is defined as the number of images that would be held in each cluster if every cluster ended up with the same size.
[pic]
Figure 20. Minimum Overlapping: Collapsing Neighbors
Algorithm #2: Minimum overlapping by allocating clusters
This algorithm (see Figure 21) allocates the maximum number of clusters. For each image, it counts the overlapping hits in each cluster. In our current prototype, an overlapping hit is when an image overlaps with other image. The image is then assigned to the cluster which where it has the minimum overlapping.
Figure 21. Minimum Overlapping: Allocating Clusters
We define:
■ N = number of images in the result set
■ C = number of clusters with no overlapping
■ K = number of clusters with minimum overlapping
■ Typically C is greater than K.
The complexity of the minimum overlapping algorithms will be:
■ Alg #1 Collapsing Neighbors : N3 + C2
■ Alg #2 Allocating Clusters: N3 + K
Algorithm #2 Allocating Clusters is more efficient than Algorithm #1 Collapsing Neighbors.
Experimental Results
In this chapter we present the results of two types of experiments that we conducted to evaluate the effectiveness of our algorithms. The first type of study is a comparison of the algorithm effectiveness to eliminate overlapping. The second is a user study to measure the effectiveness of TIC’s GUI.
Comparison of Overlapping Algorithms
We compared the effectiveness of the two algorithms: “minimum overlapping by collapsing neighbors” and “minimum overlapping by allocating clusters” in solving the agglomeration of images problem.
Algorithm #1, “minimum overlapping by collapsing neighbors “, clusters the result set into clusters with no overlapping. Then the algorithm narrows the tree, searching for clusters that are too small according to the Maximum number of clusters and tries to fill them with images of the next cluster.
Algorithm #2, “minimum overlapping by allocating clusters”, allocates the maximum number of clusters. For each image, it counts the overlapping hits in each cluster. The image is then assigned to the cluster with where it has the minimum overlapping.
Both algorithms create a number of clusters that is dependent of the size of the result set. The number of clusters allowed is calculated using the following formula:
MAX_CLUSTERS = 10 ( log(result set size) (
In order to measure the effectiveness of the algorithms we calculate the maximum of overlapping hits occurred over an image for each cluster. An overlapping hit occurs each time an image is overlapped by another image. For example: If an image in the first cluster was overlapped two times and another image was overlapped three times, the maximum overlapping hits per image in the first cluster is three. We ran the experiments using two different result sets with high overlapping. Figure 22 and Figure 23 demonstrate the results of the experiments. The horizontal axis of the graphs represents the clusters created by the algorithms and the vertical axis represents the maximum overlapping hits over an image in the corresponding cluster.
[pic]
Figure 22. Experiment Results for Result Set 1
[pic]
Figure 23. Experiment Results for Result Set 2
Algorithm #2 demonstrated to cluster the results in a better manner minimizing the overlapping occurrences in the result set. Also, as explained in Chapter 5, Algorithm #2 is faster than Algorithm #1.
User Study
This chapter presents the results of a user study conducted to evaluate the effectiveness of TIC improving the presentation of results and the ability of the users to visualize the properties of data. TIC provides the option to distribute the images in different clusters so the geo representation does not loose usability. Also, TIC provides a way to dynamically and recursively organize the results in a hierarchical way.
The user study was design to measure the benefits and effectiveness that TIC provides to image browsers.
Training
Before beginning the experiment, each participant was educated in the use of image browsers in general. We wanted to be sure each participant had a clear understanding of the assignment they were about to perform. Participants completed some pre-tasks. The participants had no time limit to do the pre-tasks. All the participants were informed that they did not have to proceed with any further tasks until they felt comfortable with the browsing system. The goal of the training was to verify that our subjects obtainded minimal familiarity with TIC.
Subjects Profile
There were 20 participants involved in this experiment; most of them were students at the University of Puerto Rico at Mayagüez, with various backgrounds including Computer Engineering and Electrical Engineering. The participants’ ages range between 25-29 years. The majority of users reported to be experts users of personal computers (PC), using them an average of 37 hours per week. We believe that this is a representative sample of the type of user that would use Terrascope.
Methodology
The users were asked to perform various image retrieval tasks designed to evaluate three features of TIC. The three tasks attempted to measure are: the effectiveness of the “minimum overlapping” clustering, effectiveness of TIC helping to answer complex queries that require clustering, and the effectiveness of the recursive clustering of the images. In each task the users were asked to find some data making use of the geo representation alone and making use of the image clustering.
Results:
Task #1
The users were asked to find an image with much overlapping. They were given the id of the image and were asked to find the metadata and to select the geo representation in order to view the actual image.
80% (16 out of 20) of the users found that it was easier and faster to target the image using TIC’s clustering for minimum overlapping than using only the geo-represented polygons.
Task #2
The users were asked to find “Which sources took images of South America”. To find the answer the user had to use the geo representation alone and had to use the clustering options of TIC. To use TIC the users where asked to cluster the images by Source and then select the clusters and view the according geo representations drawn in the map.
100% (20 out of 20) of the users found easier to find answers to queries using TIC’s clustering and structure view of the results.
Task #3
The users were asked to make a common query done by scientist: “Which sensor used which source to take the images of South America”. To find the answer the user had to use the geo representation alone and had to use the clustering options of TIC. To use TIC the users where asked to cluster the images “by Source” and recursively cluster “by Sensor”.
100% (20 out of 20) of the users found easier to find answers to complex queries using TIC’s recursive clustering and structure view of the results.
Conclusions
Our experiments provide evidence that TIC improves the effectiveness of manipulation and image navigation by providing an organized and structured display of the retrieved images. Most users preferred to use the clustering methods to target the answer to their queries. All subjects reported that they consider the tools for dynamic clustering and manipulation of the results useful.
References
[1] M. Rodríguez and E. Coronado, SRE Search and Retrieval Engine of TerraScope Earth Science Information System, Proceedings of IASTED International Conference, Computer Science and Technology, Cancun, Mexico, 2003
[2] B. Vélez, A. Cabarcas, L. Malavé: TIN: An Interactive Image Navigator Providing Ubiquitous Access To Distributed Geo-Spatial Data, Proceedings of International Conference on Information Technology: Coding and Computing Las Vegas, Nevada, USA, 2004
[3] XML, From the Inside Out.
[4] C. Moock. ActionScript for Flash MX: The Definitive Guide, Second Edition. O’Reilly & Associates, Inc. December 2002.
[5] Quicklook Swath Browser: A web-based tool host by The Canada Centre for Remote Sensing.
[6] P. Mouginis-Mark, L. Glaze, P. Flament. Virtually Hawaii by University of Hawaii.
[7] U.S. Geological Survey (USGS) Landsat 7 Image Viewer:
[8] T. Barclay, J. Gray, and D. Slutz, Microsoft TerraServer: A Spatial Data Warehouse, Proceedings of International Conference on Management of Data and Symposium on Principles of Database Systems, ACM SIGMOD, 2000.
[9] E. Lim, Goh, D., et al., G-Portal: A Map-based Digital Library for Distributed Geospatial and Georeferenced Resources, Proceeding of the second ACM/IEEE-CS joint conference, International Conference on Digital Libraries. Portland, Oregon, United States, 2002.
[10] B. Vélez, J. Valiente, Interactive Query Hierarchy Generation Algorithms for Search Result Visualization, Proceedings of the IASTED International Conference Internet and Multimedia Systems and Applications, August 2001.
6 Publications
1 Publications in journals (students underlined):
|Authors: |Title |Name of Publication |
|L.O. Jiménez, J. Rivera-Medina,|Integration of Spatial and Spectral Information by means of |IEEE Transactions on Geoscience and Remote |
|E. Rodríguez-Díaz, E. Arzuaga, |Unsupervised Extraction and Classification for Homogenous |Sensing, Volume 43, Issue 4, April 2005, |
|M. Ramírez |Objects Applied to Multispectral and Hyperspectral Data |Page(s):844 – 851 |
|V. Manian and A. Ross |Face detection using statistical and multi-resolution texture |Multimedia Cyberspace Journal, Special Issue |
| |features |on Pattern Recognition and Bioinformatics, |
| | |Vol. 3, No. 3, pp. 1-9, 2005 |
2 Publications submitted to journals
|Authors: |Title |Name of Publication |Publication |
| | | |status: |
|Julio Duarte, Miguel Vélez-Reyes, |Sensitivity Analysis of the Water-leaving Remote Sensing |Remote Sensing of the |In revision |
|Stefano Tarantola, and Roy Armstrong |Reflectance in Coastal Shallow Waters |Environment | |
|Julio M. Duarte C., Paul Castillo and |Comparative Study of Semi-implicit Schemes for Nonlinear |IEEE Transactions on |Submitted |
|Miguel Velez-Reyes |Diffusion in Hyperspectral Imagery |Image Processing | |
|Vidya Manian, Luis O. Jimenez and |A comparison of statistical and multiresolution texture |IEEE Transactions on |Submitted |
|Miguel Velez-Reyes |features for improving hyperspectral image classification |Remote Sensing | |
|Luis O. Jiménez-Rodríguez, Emmanuel |Unsupervised Feature Extraction Techniques for |IEEE Transactions on |Submitted |
|Arzuaga-Cruz, and Miguel Vélez-Reyes |Hyperspectral Data and its Effects on Supervised and |Remote Sensing | |
| |Unsupervised Classification | | |
|V. Manian, R. Vasquez and A. Ross |Nonlinear dynamics in two dimensional texture pattern |Optical Engineering |Submitted |
| |discrimination | | |
3 Publications in Conference Proceedings:
|Authors: |Title |Name of Publication |
|M.Vélez-Reyes and S. Rosario |Solving Adundance Estimation in Hyperspectral |Proceedings IEEE International Geosciences |
| |Unmixing as a Least Distance Problem. |and Remote Sensing Symposium, Alaska, 2004. |
|E. Arzuaga-Cruz, L.O. Jimenez-Rodriguez, M. |A MATLAB Toolbox for Hyperspectral Image |Proceedings IEEE International Geosciences |
|Velez-Reyes, D. Kaeli, E. Rodriguez-Diaz, H.T. |Analysis. |and Remote Sensing Symposium, Alaska, 2004. |
|Velazquez-Santana, A. Castrodad-Carrau, L.E. | | |
|Santos-Campis, and C. Santiago. | | |
|Y.M. Masalmah, M. Vélez-Reyes, and S. |“An Algorithm for Unsupervised Unmixing of |In Algorithms and Technologies for |
|Rosario-Torres, |Hyperspectral Imagery using Positive Matrix |Multispectral, Hyperspectral, and |
| |Factorization.” |Ultraspectral Imagery XI, Proceedings of |
| | |SPIE Vol. 5806, April 2005. |
|S. Rosario-Torres, E. Arzuaga-Cruz, M. |“An Update on the MATLAB Hyperspectral Image |In Algorithms and Technologies for |
|Velez-Reyes and L.O. Jiménez-Rodríguez, |Analysis Toolbox.” |Multispectral, Hyperspectral, and |
| | |Ultraspectral Imagery XI, Proceedings of |
| | |SPIE Vol. 5806, April 2005. |
|S. Rosario-Torres and M. Vélez-Reyes | “An algorithm for fully constrained abundance |In Algorithms and Technologies for |
| |estimation in hyperspectral unmixing.” |Multispectral, Hyperspectral, and |
| | |Ultraspectral Imagery XI, Proceedings of |
| | |SPIE Vol. 5806, April 2005. |
|S. Morillo-Contreras, M. Vélez-Reyes, and S.D. |, “A comparison of noise reduction methods for |In Algorithms and Technologies for |
|Hunt |image enhancement in classification of |Multispectral, Hyperspectral, and |
| |Hyperspectral imagery.” |Ultraspectral Imagery XI, Proceedings of |
| | |SPIE Vol. 5806, April 2005. |
|J.A. Goodman, F. Gilbes, M. Vélez-Reyes, S.D. |SeaBED: A Controlled Laboratory and Field Test |Proceedings of 8th International Conference |
|Hunt, R.A. Armstrong |Environment for the Validation of Coastal |on Remote Sensing for Marine and Coastal |
| |Hyperspectral Image Analysis Algorithms |Environments : Halifax, Nova Scotia, May |
| | |2005 |
|J.A. Goodman |Hyperspectral Remote Sensing of Shallow Coral |Proceedings of 8th International Conference |
| |Reef Ecosystems Using AVIRIS and HYPERION |on Remote Sensing for Marine and Coastal |
| | |Environments : Halifax, Nova Scotia, May |
| | |2005 |
|V. Manian and M. Velez-Reyes, |A boosted algorithm for texture classification |In Visual Information Processing XIV, |
| |and object detection, |Proceedings of SPIE, Vol. 5817, 2005. pp. |
| | |25-33 |
|Shawn Hunt, and Leila S. Rodríguez |Fast Piecewise Linear Predictors for Lossless |Proceedings IEEE International Geosciences |
| |Compression of Hyperspectral Imagery |and Remote Sensing Symposium, Alaska, 2004. |
|J. Morales, J.; Trabal, J., Cruz-Pol, S. and |"Cirrus Clouds Millimeter-Wave Reflectivity |Microwave Remote Sensing of the Atmosphere |
|Sekelsky, S. |Comparison with In-Situ Ice Crystal Airborne |and Environment IV, September, 2004, |
| |Data", |Proceedings of SPIE Vol. 5654 |
|Cruz Pol, S. L., José Maeso, V. Bringi and V. |“DSD characterization and computations of |Proceedings IEEE International Geosciences |
|Chandrasekar |expected reflectivity using data from a |and Remote Sensing Symposium,, Seoul, Korea,|
| |Two-Dimensional Video Disdrometer deployed in a|2005. |
| |Tropical Environment”, | |
|Lizvette Malavé and Bienvenido Vélez |Terrascope Image Clustering: Applying |Third IASTED International Conference on |
| |Clustering Techniques to Image Agglomeration in|Communications, Internet, and Information |
| |Image Retrieval Systems |Technology (CIIT 2004). |
|Hamed Parsiani, Maritza Torres, Pedro Rodriguez |Material Characteristics in Fourier Domain |Prodeedings of IASTED-SIP-2004 |
| |Formulation, a signature to determine soil | |
| |type, moisture, and vegetation health, based on| |
| |multilayer GPR reflection | |
|Hamed Parsiani, Pedro Rodriguez |Subsurface Material Type Determination from GPR|Prodeedings of SPIE-2004 |
| |Signatures | |
|J.M. Ortiz and M. Vélez-Reyes, In |“The Use of Edge-Enhancing Smoothing |Proceeding 4th WSEAS International |
| |Pre-Filters to Aid in the Detection of Oceanic |Conference in Electronics, Signal |
| |Features.” |Processing, and Control, Rio de Janeiro, |
| | |Brazil, April 2005. |
4 Publications submitted to Conferences:
|Authors: |Title |Name of Publication |Publication status: |
|V. Manian, L. O. |A comparison of statistical and multiresolution |Image and Signal Processing|In print. |
|Jiménez-Rodríguez, M. |texture features for improving hyperspectral image|for Remote Sensing XI, | |
|Vélez-Reyes |classification |Proceedings of SPIE Vol. | |
| | |5982 | |
5 Presentations (not in conference proceedings)
|Presenters: Underline student|Title |Year and month of |Name of the conference, |City, State, or country |
|names put main author first | |the presentation |symposium, etc. |where the presentation |
| | | | |was given |
|M. Vélez-Reyes and Y. |An Algorithm for Unsupervised Unmixing |May 24 -27, 2005 |NASA Earth Sciences and |Pasadena, CA |
|Masalmah |of Hyperspectral Imagery Using Positive| |Applications Workshop, | |
| |Matrix Factorization | | | |
|J. Goodman, L. Guild, R. |A Summary of the 2004 AVIRIS Mission to|May 24 -27, 2005 |NASA Earth Sciences and |Pasadena, CA |
|Armstrong, F. Gilbes, R. |Florida and Puerto Rico | |Applications Workshop, | |
|Berthold, A. Gleason | | | | |
|M. Vélez-Reyes |“Tropical Center for Earth and Space |November 16, 2004. |TCESS Open House at NASA|NASA Goddard Space Flight|
| |Studies: An Overview.” | |GSFC |Center |
|Shawn D. Hunt |Spectral Sensing and Image Processing |November 16, 2004. |TCESS Open House at NASA|NASA Goddard Space Flight|
| |Research @ LARSIP | |GSFC |Center |
|Jose Maeso and Sandrra |Microwave Remote Sensing of the |November 16, 2004. |TCESS Open House at NASA|NASA Goddard Space Flight|
|Cruz-Pol |Atmosphere | |GSFC |Center |
|Bienvenido Vélez |Ubiquitous Access to Autonomous |November 16, 2004. |TCESS Open House at NASA|NASA Goddard Space Flight|
| |Satellite Image Repositories in | |GSFC |Center |
| |TerraScope | | | |
7 New Grants
|Principal Investigator |Title |Agency |Amount & |
| | | |Duration |
|Sandra Cruz Pol |Student Developed Radar Network for NASA |Puerto Rico NASA Space Grant Announcement of |$90,000, 1 year |
| |Tropical Satellites Validation |Opportunity (AO) 2005-2006 IDEAS-ER program | |
|James Goodman and Wilson|Utilizing High-Performance Computing to |Puerto Rico Space Grant Consortium (PRSGC), Innovative |$29,940, 1 year |
|Rivera |Investigate Performance and |Developments | |
| |Sensitivity of an Inversion Model for |to Enhance Aerospace Education and Research | |
| |Hyperspectral Remote Sensing of | | |
| |Shallow Coral Ecosystems | | |
8 Student Supported
1 Graduate Students (fully or partially) Supported by TCESS
|Student Name |Project Title |Advisor |Graduation Date |Degree |If graduated, |
| | | | | |Plans |
|Enrico Mattei | |Hamed Parsiani |May 2006 |MSEE | |
|Lizbeth Malavé | |Bienvenido Vélez |December 2004 |MSCpE |Industry |
| | | | | |Raytheon |
|Idalis Rodríguez| |Manuel Rodríguez |December 2004 |MSCpE |Industry |
|Yahya Masalmah |Unsupervised unmixing of hyperspectral imagery |Miguel Velez-Reyes |May 2006 |Ph.D. | |
|Christian Nieves|Development of a Website for the SeaBED TestBED |Miguel Velez-Reyes |December 2005 |MSCpE | |
| |Facility | | | | |
|Vanssa Ortiz |Hyperspectral Change Detection using Temporal |Miguel Vélez-Reyes |May 2005 |MSEE |Graduate School, |
| |Principal Components Analysis | | | |RPI |
2 Graduate Students in Associated Research (use TCESS Facilities)
|Student Name |Project Title |Advisor |FundingSource |Graduation Date|Degree |If graduated, |
| | | | | | |Plans |
|Julio M. Duarte |Sensitivity Analysis of Forward Models for |Miguel |NSF CenSSIS |May 2006 |Ph.D. CISE | |
| | |Velez-Reyes | | | | |
|Alejandra Umaña |Determining the dimensionality of hyperspectral|Miguel |NSF CenSSIS |May 2008 |Ph.D. CISE | |
| |imagery for unsupervised band selection |Velez-Reyes | | | | |
|Alexey |Retrieval of Bottom Properties in Shallow |Miguel |NSF CenSSIS |December 2005 |MSEE | |
|Castrodad-Carrau |Waters from Hyperspectral Imagery |Velez-Reyes | | | | |
|Shirley |A Comparison of Resolutiuon Enhancement Methods|Miguel |NSF CenSSIS |May 2005 |MSCpE |Industry |
|Morillo-Contreras|as Pre-Processing for Classification of |Velez-Reyes | | | | |
| |Hyperspectral Images | | | | | |
|Maritza Torres | |Hamed Parsiani|NOAA CREST |December 2004 |MSEE | |
|Mairim Ramos | |Hamed Parsiani|NOAA |May 2006 |MSEE |Unknown |
| | | |CREST | | | |
3 Undergraduate students (partially or fully) supported by TCESS
|Student Name |Project Title |Advisor |Graduation Date |Degree |If graduated, plans |
|Carl A. Lizzattaga | |Hamed Parsiani |May 2007 |BSEE | |
|Joshua Benitez | |Hamed Parsiani |May 2006 |BSEE | |
|Angel Villalaín | |Bienvenido Vélez |December 2004 |BSCpE |Industry |
4 Undergraduate Students in Associate Research (use TCESS Facilities)
|Student Name |Project Title |Advisor |Funding Source |Graduation Date |Degree |If graduated, plans |
|Soralis Pimentel | |Sandra Cruz Pol |NASA FAR |May 2006 |BSEE | |
|Luis Pérez Clifford | |Sandra Cruz Pol | |May 2006 |BSEE | |
|Ricardo Ríos | |Sandra Cruz Pol | |May 2007 |BSEE | |
|Carlos Rodríguez | |Sandra Cruz Pol | |May 2006 |BSEE | |
|José Padovani | |Sandra Cruz Pol | |May 2007 |BSEE | |
9 Post-doctoral Fellows
|Fellow Name |Project Title |Advisor |
|Vidya Manian |Integration of Spectral and Spatial Information in |Miguel Vélez-Reyes, |
| |Hyperspectral Image Classification |Luis O. Jiménez |
| | |(CenSSIS co-funding) |
|James Goodman |Remote Sensing of Benthic habitats using Hyperspectral |Miguel Vélez-Reyes |
| |Imaging |(CenSSIS co-funding) |
10 Service Activities
• The Terrascope system was applied to management of data over Rusia as part of a project for the National Geospatial Agency.
• Dr. Sandra Cruz-Pol was a reviewer for IEEE Transactions for Geosciences and Remote Sensing.
• Dr. Sandra Cruz is Associate Editor of the Newsletter of the IEEE Geoscience and Remote Sensing Society.
• Dr. Miguel Vélez-Reyes was a reviewer for IEEE Transactions for Geosciences and Remote Sensing.
• Dr. Miguel Vélez-Reyes was a member of the program committee for Multispectral, Hyperspectral, and Ultraspectral Imagery XI, part of the SPIE Defense and Sequrity Symposium held in Orlando, FL in April 2004.
• Dr. Luis Jiménez was a session Co-Chair at Remote Sensing of the Ocean, Sea Ice, and Large Water Regions 2005, Conference part of 11th SPIE International Symposium on Remote Sensing 19–22 September, 2005 Bruges, Belgium
-----------------------
[1] 2005 Mars Reconnaissance Orbiter mission
-----------------------
Figure 2.7. Neural Network Training
Figure 2.6. Calculation Selection for Material Determination
Figure 2.5. MCFD/NN Graphical User Interface – Main Window
Neural Network
Output Layer
Hidden Layer
Input Layer
MCFD
Calculation
Figure 2.4. Training samples and representation of a Neural Network
GPR antenna over clay sample
GPR scan
Training Set
Validation
of results
Field Observations
Trained Neural Network
Get unknown soil sample
Calculate
MCFD
Extract an appropriate wavelet
GPR Image
file.bmp
Figure 2.3. General Methodology
[pic]
[pic]
[pic]
Figure 2.2 System representation of the layer characteristics, with the transmitted and reflected waves
Figure 2.1. Cross section of the ground
[pic]
[pic]
[pic]
Air
Layer I
La Parguera
Allocate clusters
For all images
For all branches in new tree
Find minimum overlapping
Add image to branch were minimum overlapping
For all images
For all branches in new tree
If image does not overlap other
Add image to branch
If image overlapped in every branch
Add branch
Add image to branch
For all clusters
While cluster is too small
Get image from next cluster
For all images
For all branches in new tree
If image does not overlap other
Add image to branch
If image was not added
Add branch
Add image to branch
For all branches in current tree
For all images in branch
Get image source from Array
For all source branches in new tree
If image’s source is branch’s source
Add image to branch
For all images
Get image’s date from array
For all date branches in new tree
,-.T—§¨ÔØèéö÷)
*
*+—›ÊîprIf image’s date is less than branch’s
Add branch before current branch
Add image to branch
If image’s date is branch’s
Add image to branch
Fig. 4.5 Extracted soil reflection
Fig. 4.4 Extracted air reflection
Fig. 4.3 GPR signal with air reflection and soil reflection
[pic]
Fig. 4.2: Air and soil reflections are well separated
Fig 4.1 . 300 GPR scans showing the effect of vertical antenna motion
Appropriate wavelet
Fig 3.8 Appropriate Wavelet Extraction Method
Yes
No
*
*
Non-soil Elements Renoval
(Fig 3.6, 3.7)
Best Scan in Wavelet Succession?
Wavelet Noisy Tail Removal (Fig 3.5)
Max/Min and Attenuation Test (Fig 3.3, .4)
Full Wavelet
Extraction
(Fig 3.2)
GPR
Image
Fig 3.1
Fig 3.7 Fourier Transform of manipulated wavelet
Fig 3.6 Wavelet with added artificial perturbation
Fig 3.5 Verification of separation of boundaries and elimination of noise and extraction of the desired wavelet
Fig 3.3 GPR maxima and minima verification
Fig 3.4 Check attenuation greater than air reflection
Fig 3.2 Single full-length GPR scan.
Fig 3.1 GPR grayscale image displaying
successive scans with 512 rows.
a))
b))
Figure 2.8. Material Type Determination test: a) Main Window, b) Result Message Box
................
................
In order to avoid copyright disputes, this page is only a partial summary.
To fulfill the demand for quickly locating and searching documents.
It is intelligent file search solution for home and business.
Related searches
- 1 or 2 374 374 1 0 0 0 1 168 1 1 default username and password
- 1 or 3 374 374 1 0 0 0 1 168 1 1 default username and password
- 1 or 2 711 711 1 0 0 0 1 168 1 1 default username and password
- 1 or 3 711 711 1 0 0 0 1 168 1 1 default username and password
- 1 or 2 693 693 1 0 0 0 1 168 1 1 default username and password
- 1 or 3 693 693 1 0 0 0 1 168 1 1 default username and password
- 1 or 2 593 593 1 0 0 0 1 or 2dvchrbu 168 1 1 default username and password
- 1 or 3 593 593 1 0 0 0 1 or 2dvchrbu 168 1 1 default username and password
- 1 or 2 910 910 1 0 0 0 1 168 1 1 default username and password
- 1 or 3 910 910 1 0 0 0 1 168 1 1 default username and password
- 192 1 or 2 33 33 1 0 0 0 1 1 1 default username and password
- 1 or 2 364 364 1 0 0 0 1 168 1 1 admin username and password