Improving the quality of distribution models for conservation by addressing shortcomings in the field collection of training data

Conservation biology can benefit greatly from models that relate species' distributions to their environments. The foundation of successful modeling is a high-quality set of field data, and distribution models have specialized data requirements. The role of a distribution model may be primarily predictive or, alternatively, may emphasize relationships between an organism and its habitat. For the latter application, the environmental variables recorded should have direct, biological relationships with the organism. Interacting species may be valuable predictors and can improve understanding of distribution patterns. Sampling should cover the full range of environmental conditions within the study region, with samples stratified across major environmental gradients to ensure thorough coverage. Failure to sample correctly can lead to erroneous organism-environment relationships, affecting predictive ability and interpretation. Sampling ideally should examine a series of spatial scales, increasing the understanding of organism-environment relationships, identifying the most effective scales for predictive modeling and complementing the spatial hierarchies often used in conservation planning. Consideration of statistical issues could benefit most studies. The ratio of sample sites to environmental variables considered should ideally exceed a ratio of 10: 1 to improve the analytical power and reliability of subsequent modeling. Presence and/or absence models may suffer bias if training data detect the study organism at an atypical proportion of sites. We considered different strategies for spatial autocorrelation and recommend it be included wherever possible for the benefits in biological realism, predictive accuracy, and model versatility. Finally, we stress the importance of collecting independent evaluation data and suggest that, as with the training data, a systematic approach be used to ensure broad environmental coverage, rather than relying on a random selection of test sites.