Presenter/Author Information

Andrei Kirilenko
R. S. Hanley

Keywords

artificial neural networks, neuroniche, model ensemble, ecological niche, model uncertainty

Start Date

1-7-2008 12:00 AM

Description

Transformation of spatial distributions of species is a key element for understanding global and regional impact of climate change on environment in the future. Yet little is known about current distributions of species, especially in relation with physical parameters of environment. This is primarily due to limited availability of data on species occurrences, which prevents reconstruction of their geographical distribution with standard methods of spatial statistics. To address this problem, specialized applications called distributional models, based on the hypothesis of fundamental niche, are being developed. We have built a new tool for predicting species’ geographical distributions based on presence-only data. The tool is developed around an Artificial Neural Network (ANN) simulation engine, which employs the Stuttgart Neural Network Simulator (SNNS). Our tool is capable of using multiple environmental layers as predictors to generate the patterns at the species’ presence and pseudo-absence localities, selecting and training an ANN using these patterns, selecting the optimal ANN, testing the selected ANNs on independent sets of data, applying the selected model to project species distribution at current or modified climate conditions, and porting the resultant presence probability maps to ESRI ArcGis. One of the frequent criticisms with ANN-based computer applications is connected with their generalization ability, demonstrated through “model overtraining”. When the pre-selected number of training steps for the model is too high, the modelgenerated predictive surface looses its smoothness, becoming too tightly locked at the training dataset. To reduce the generalization abilities of ANN predictions of insect species’ occurrences, we employed model ensembles. The redundant ensemble of ANN models was used to simulate spatially distributed probabilities of species presence from the train presence data set; the results of modeling with each individual network were then pooled together using their linear combination. Such modeling based on employing redundant model ensembles yielded a significant improvement in overall model performance and reduced model-related uncertainty.

Share

COinS
 
Jul 1st, 12:00 AM

Reducing Uncertainty in Ecological Niche Models with ANN Ensembles

Transformation of spatial distributions of species is a key element for understanding global and regional impact of climate change on environment in the future. Yet little is known about current distributions of species, especially in relation with physical parameters of environment. This is primarily due to limited availability of data on species occurrences, which prevents reconstruction of their geographical distribution with standard methods of spatial statistics. To address this problem, specialized applications called distributional models, based on the hypothesis of fundamental niche, are being developed. We have built a new tool for predicting species’ geographical distributions based on presence-only data. The tool is developed around an Artificial Neural Network (ANN) simulation engine, which employs the Stuttgart Neural Network Simulator (SNNS). Our tool is capable of using multiple environmental layers as predictors to generate the patterns at the species’ presence and pseudo-absence localities, selecting and training an ANN using these patterns, selecting the optimal ANN, testing the selected ANNs on independent sets of data, applying the selected model to project species distribution at current or modified climate conditions, and porting the resultant presence probability maps to ESRI ArcGis. One of the frequent criticisms with ANN-based computer applications is connected with their generalization ability, demonstrated through “model overtraining”. When the pre-selected number of training steps for the model is too high, the modelgenerated predictive surface looses its smoothness, becoming too tightly locked at the training dataset. To reduce the generalization abilities of ANN predictions of insect species’ occurrences, we employed model ensembles. The redundant ensemble of ANN models was used to simulate spatially distributed probabilities of species presence from the train presence data set; the results of modeling with each individual network were then pooled together using their linear combination. Such modeling based on employing redundant model ensembles yielded a significant improvement in overall model performance and reduced model-related uncertainty.