Glenn Archetto Use of predictive modeling in landscape ecology

[Pages:7]Glenn Archetto

Use of predictive modeling in landscape ecology

Landscape pattern change has a wide-ranging effect on how ecosystems and human communities function. These effects can have large spatial and temporal scales which have made examining these processes difficult. Early in the conception of landscape ecology, much focus was put into creating effective data sources, metrics and statistical approaches that could capture spatial and temporal patterns (Gausan et al). With the refinement of remote sensing and GIS techniques, they have become an important tool in understanding how these patterns affect process. Unfortunately, these technologies can only bring us so far in our understanding as changes in land use result from complex interactions of many factors, some of which are of human origin and some which arise from the environment (Pijanowski et al). Predictive models help us understand how these processes work by creating statistical models which include these complicated factors and relate them to patterns that are observed on the ground. By combining strengths traditional in situ data with the ability of remote sensing and GIS to capture landscape patterns at vast spatial and temporal scales, researchers have developed a new methodology in understanding the landscape. In this paper, I will discuss how researchers create and predictive models to help understand these complex interactions and give the strengths and weaknesses of this method.

Human development has long been known to be a cause of habitat degradation and loss of biodiversity. As this development further creeps across the landscape, interactions between natural and man-made landscapes become all the more relevant to researchers attempting to understand this relationship. Due to the importance of understanding how land use affects and environmental change affect species distribution, predictive modeling has taken a major role in conservation and management. (Guisan et al) In order to begin understanding these complex issues, models must be developed. The first step in creating a model is forming a conceptual network which includes defining a study area and developing a hypothesis which will be studied. When dealing with habitat mapping or predicting the presence or absence of a particular species in those habitats, the next step is often determining variables which will be applicable to the research subject.(Guisan) The most important of these are the abiotic and biotic factors which may limit the distribution of organisms or ecosystems. According to Guisan et al, the ecological gradients formed by these factors form a boundary in which one direction is physically stressful while the other is biologically stressful. For the purpose of a model, these create the upper and lower limits in which an organism can be found within the landscape. For example, in Anderson et als study, researchers attempted to determine the range of two species of spiny pocket mice in South America abiotic factors such as elevation, aspect, soil conditions and climate data and biotic factors such as competition and coarse vegetation cover were chosen as variables which may help predict their range. In studies which attempt to determine future land-cover, not only do abioitc and biotic factors have to be considered, but human related elements must also be included. From here the actual process of model building can begin. Due to the heterogeneous nature of the landscape, it is too

complex to be predicted accurately in every temporal or spatial instance (Drew et al). Model development is therefore separated into two specific groups; calculating tools and theoretical models. (Guisan et al). Calculating tools are static models which aim to only provide information about the configuration of the landscape. Theoretical models are developed to predict responses in the landscape from the abiotic and biotic variables included in the model. (Guisan et al) While these two types of models differ in scope and goals, they both allow for statistical determination of which variables actually describe a landscape. This is typically done using a step-wise regression in which variables which show close correlation to the feature or organism being studied are kept in the model while those which show low correlation are eliminated (Garza-Perez et al). The use of statistical methods to quantify the relationship between pattern and process has allowed for a step forward in the practical application of landscape ecology.

The ability to I found that the greatest strength of these models is in predicting species distribution over large spatial ranges. For example, Osborne et al successfully used an empirical model to map potential great bustard habitat in Spain by accounting for human development and habitat preferences. Accurate prediction of habitat location is of prime importance to conservation of endangered species and this is a fundamental strength of this type of analysis. The reasons for these importances are two-fold. The first is that it provides a relatively cost effect way of gaining preliminary information on where organisms may be located and where ground collection of data should be focused. The second importance is understanding the spatial pattern between useful habitat and developed land (Osborne et al). By outlining the location of probable habitat for a species, managers can focus on conserving lands which have the least likelihood of being affected by human incursion. Another major asset of predictive modeling is in prediction of future human development on the landscape. Work done in predictive modeling of agricultural development in Belize has shown positive results in the ability to predict road development as a function of increased agricultural development Osborne et al) (Rutherford et al). While models which include human choice are hard to quantify, there has been enough success to use these models in the planning of future agricultural planning and road development. This is of importance to landscape ecology as a whole because road development is a major cause of fragmentation and degradation associated with increased ability to penetrate natural habitats (Osborne et al). Finally, predictive modeling is useful in baseline monitoring and studying as expert knowledge can be used to create preliminary observations about habitat pattern and species allocation (Drew et al). These models allow for the testing of a hypothesis based on multiple future models which taken into account multiple scenarios of land use. In a study done by White et al, predictive models were created using expert knowledge to determine populations which were at risk to development by considering expert knowledge on each species range as well as six maps of possible future development. Studies like this are extremely important to understand how species should be managed for even though it is not based on ground truth data (White et al).

While there are many benefits associated with the use of predictive modeling, there are also weaknesses which must be addressed. A primary frustration is the reliance on ground truthing to provide base data to calibrate models. Remote sensing data acquisition outpaces

the ability to collect ground data, especially when doing large regional studies. While remote sensing can provide a large quantity of data on environmental pattern and general composition, information such as species distributions soil conditions must be collected directly. This problem is compounded in undeveloped areas such as southern Belize where little ground data has been collected and they have a weak GIS infrastructure.(Chomitz) Lack ground truthing leads to type I error in that areas will be misidentified as possible habitat or incorrectly identified. Another weakness in predictive models is accounting for direct human interaction or other un-accounted for variables in landscape function. In a study done by Rutherford et al on transition between forest and agricultural land uses in Switzerland, a major problem in their study was the planting of trees and hedgerows by farmers. This problem arose in a high rate of transition from meadow to closed tree canopy in their model, which would be impossible due the twelve year time period. Due to the complexity of a landscape, not all variables can be accounted for in a model no matter how complicated it is; therefore factors outside of the study may have an influence which will not be accounted for by the model. (Guisan et al) Finally, the nature of landscapes leads to problems in study design. Landscapes are comprised of a mosaic of interrelated patches, which makes collecting independent data points difficult. Spatial autocorrelation was a common problem in many of the papers reviewed. The lack of independent sample points leads to an overconfidence in the fit of the model and incorrect predictions. In Pijanowski et als study, autocorrelation was found to be the cause of most errors in their model.

Overall, predictive models have been shown to be able to predict and quantify how abiotic, biotic and human related variables can produce changes in the landscape or identify habitat that is may be suitable for conservation. The ability to take ground data and apply remote sensing and GIS technology to extrapolate shows great promise in the future as predictive modeling continues to evolve. While the extreme complexity of the processes which lead to landscape pattern can cause error in the estimation, the quantification of land characteristics makes this a viable and powerful tool for landscape ecologists to apply to understand and predict ecological outcomes. As human development continues to disrupt the way landscapes function, there will be an ever increasing demand for these type of analysis's in order make informed management decisions based on statistical inquiry into previously un-quantifiable characteristics.

Annotated Bibliography

Gillian N. Rutherford, Peter Bebi, Peter J. Edwards, Niklaus E. Zimmermann.2008. Assessing land-use statistics to model land cover change in a mountainous landscape in the European Alps. Ecological Modelling vol. 212, pg 460-471

In this paper, Rutherford et al use previously collected land-use land-cover information to determine its usefulness in creating a model to predict land change between agricultural lands and forest cover. To create the model, they separated data into response and predictor variables. The response variable was Land-cover data was taken from the Swiss-federal department of statistics. The land-cover data was The data was originally done in 100 meter pixel resolution, so all other data was brought into this resolution.

J. R. Garza-Perez, A. Lehmann, J. E. Arias-Gonzalez. Spatial prediction of coral reef habitats: integrating ecology withi spatial modeling and remote sensing. Marine Ecology Progress Series Vol. 269, Pg -141-151

In this paper, Garza-Perez et al use a combination of remote sensing data an in situ data to predict benthic communities in Akumal Reef, Mexico. To do this, 54 sampling stations were set up in various reef-zones. At these stations, a modified version of the Aronson & Swanson video transect method was used in order to have control points for benthic structure. A digital topographic model was then created using 6752 echo-sounding and satellite image data points which were extrapolated using a TIN function. Using the bathymetric data in combination with high resolution satellite data, two consecutive modeling runs were performed using GRASP methodology. They found that the use of satellite imagery was useful in predicting benthic communities, with 68.3 to 86.2% of correlation in the models and greater than >50% accuracy for predicted maps.

Robert P. Anderson, Marcela Gomez-Laverde, A Townsend Peterson. Geographical distributions of psiny pocket mice in South America: insights from predictive models. 2002. Global Ecology & Biogeography. Vol 11, pg 131-141

Anderson et al attempt to create a predictive model of two species of spiny pocket mice in north-western South America using a combination of occurrence records and environmental data from GIS maps. Occurrence records included 56 collection localities for H. Australis and 40 locations for H. anomalus. The GIS environmental data included elevation, slope, aspect, soil conditions, coarse vegetation zones and three coverages for solar radiation, temperature and precipitation. A model was created using Genetic Algorithm for Rule-Set Prediction (GARP). This model searches for associations between environmental characteristics of known occurrence locations and those in the rest of the study. This method uses iterative processes to determine the predictive accuracy of a particular rule and whether it should be included in the study. The final model consists of if-then statements to determine whether a particular pixel will predicted absent or present for the particular spiny pocket mouse species. They found that while their best models predicted presence in habitat which is known to be not suitable to either species, they found that this process was useful in showing the power of models in presence absence prediction.

Juan E. Malo, Francisco Suarez, Alberto Diez. Can we mitigate animal-vehicle accidents using predictive models? 2004. Journal of Applied Ecology vol. 41, pg 701-710.

In this paper, Malo et al create a model which can be used to predict areas with high rates of vehicle collisions with wild animals. To do this a database of 2067 animal collisions were used to create two data sets at different spatial scales. The first dataset was comprised of

1 km road sections with both high and low numbers of animal collisions. The second dataset comprised of collision and no collision incidences at points on the road system at a sub-1-km scale. Logistical regression was used to determine relationships between collisions and land- cover in adjacent areas. They found that this technique was highly significant in identifying areas of high collision occurrence, and resulted in more than 70% correct classification of cases.

Antoine Guisan, Niklaus E. Zimmermann. Predictive habitat distribution models in ecology. 2000. Ecological Modelling. Vol 135, pg 147-186

This paper gives a review of the process used in creating predictive models for habitat. It goes over step by step the process of choosing variables, how to implement study design, the appropriateness of using an empirical or theoretical model and what statistical analyses should be used in order to draw significance from the data. I found this paper to be extremely informative on the types of models being used and the statistical methods behind them and gave a great background to the implementation and development of predictive landscape models.

P.E. Osborne, J.C. Alonso, R.G. Bryant. Modelling landscape-scale habitat use using GIS and remote sensing: a case study with great bustards. 2001. Journal of Applied Ecology. Vol 38, pg 458 - 471

Attempt to create a model which can predict breeding habitat of the great bustard in central Spain at the landscape scale using remote sensing and GIS data. This paper attempts to provide distribution data for managers as they can't keep up with ever changing land-use which affects breeding habitat for this endangered species. To create the model, GIS layers were created for roads, buildings, railways and river systems. This data was taken from digitized infrastructure maps and were rasterized to 80-m pixel resolution. A new variable was created with each infrastructure layer in which each central pixel in a 13x13 window was assigned with the proportion of pixels which featured that infrastructure type. This was combined with a digital terrain model to give each pixel an altitude as bustard breeding habitat choices are related to slope. Satellite imagery was then used to perform an NDVI analysis to determine vegetative type. Finally, bird census data was introduced as control data to determine habitat types used by bustards. They found that this model was successful at mapping predicted breeding habitat for Greater Bustards with it correctly predicting 93% of occupied sites and 78.9% of unoccupied sites.

Kennth M. Chomitz and David A. Gray. Roads, Land Use, and Deforestation: A Spatial Model Applied to Belize. The World Bank Economic Review. Vol 10, No 3: pg 487 ? 512.

In this paper, Chomitz and Gray attempt to create a model which can predict road building and land use change in Belizean forests. They used the assumption that market access, land quality

and tenure status affected the probability of agricultural land use and whether semisubsistance or commercial will occur there. Chomitz and Gray used GIS data on slope, aspect, soil quality and distance to road as independent variables and Land-use land-cover data derived from SPOT satellite imagery as the dependent variable. They found that market distance, land quality and tenure have a strong effect on the likelihood and type of cultivation. Areas which are closer, are on relatively flat, fertile soil are more likely to be converted to commericial lands while poorer soils were mostly relegated to semisubsistance farming. This means that land managers may want to focus on control of semisubstance farming which often uses land on forest frontiers, which leads to road building and substantial fragmentation of forest patches.

Bryan C. Pijanowski, Daniel G. Brown, Bradley A. Shellito, Guarav A. Manik. Using neural networks and GIS to forecast land use changes: a Land Transformation Model. 2002. Computers, Environment and Urban Systems. Vol 7, pg 23.

Pijanowski et al use this paper as a case study in the use of The Land Transformation Model (LTM) and artificial neural networks (ANN's) in creating predicative land use models. They use these techniques to determine how made-made features such as roads and agricultural density and natural features such as lakes and rivers affect urban development patterns in Michigan's Grand Traverse Bay watershed. This was an innovative paper which used by empirical (ANN) and theoretical (LTM) to determine a model which accounted for predicted land use at multiple scales. This paper also gave an insight into why autocorrelation negatively affects predictive model studies and how it can be accounted for.

C. Ashton Drew, Ajith H. Perera. Expert Knowledge as a Basis for Landscape Ecological Predictive Models. Predictive Species and Habitat Modeling in Landscape Ecology: Concepts and Applications. Ch 12.

This chapter provides an overview on the use of expert knowledge in the creating of predictive modeling. Expert knowledge has both benefits and drawbacks to predictive modeling. Because expert knowledge is by definition biased, it can introduce error into modeling predictions. It outlines how predictive models are often subject to making decisions between cost effectiveness and accuracy. Expert knowledge is often cheaper to obtain than ground data, so it is often used in its place for determining appropriate variables and species or habitat boundaries. Special consideration must be used when including this type of data, most often in the form of using a Bayesian model and must be supplemented with monitoring programs to validate the model.

Denis White, Priscilla G. Minotti, Mary J. Barczak, Jean C. Sineos, Kathyrn E Freemark, Mary V Santelmann, Carl F. Steinitz, A. Ross Kiester, Eric M. Preston. Assessing Risks to Biodiveristy from Future Landscape Change. 1997. Conservation Biology, Vol. 11; pg 349 ? 360.

In this paper White et al use expert knowledge, ground data and remote sensing information to create a model which can be used to predict whether species will become extirpated from their natural ranges. By using a combination of expert knowledge and ground data, White et al was able to determine minimum amounts of habitat required for the continuous of the study species. He then used a predictive model to create 6 future land-cover scenarios used the percentage of habitat left to calculate extinction rates. This paper was insightful into how expert knowledge can be incorporated into a statistical model.

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download