Explainable artificial intelligence enhances the ecological interpretability of black-box species distribution models
Abstract
Species distribution models (SDMs) are widely used in ecology, biogeography and conservation biology to estimate relationships between environmental variables and species occurrence data and make predictions of how their distributions vary in space and time. During the past two decades, the field has increasingly made use of machine learning approaches for constructing and validating SDMs. Model accuracy has steadily increased as a result, but the interpretability of the fitted models, for example the relative importance of predictor variables or their causal effects on focal species, has not always kept pace. Here we draw attention to an emerging subdiscipline of artificial intelligence, explainable AI (xAI), as a toolbox for better interpreting SDMs. xAI aims at deciphering the behavior of complex statistical or machine learning models (e.g. neural networks, random forests, boosted regression trees), and can produce more transparent and understandable SDM predictions. We describe the rationale behind xAI and provide a list of tools that can be used to help ecological modelers better understand complex model behavior at different scales. As an example, we perform a reproducible SDM analysis in R on the African elephant and showcase some xAI tools such as local interpretable model‐agnostic explanation (LIME) to help interpret local‐scale behavior of the model. We conclude with what we see as the benefits and caveats of these techniques and advocate for their use to improve the interpretability of machine learning SDMs.
What the paper shows and why it matters
(AI-generated)
Species distribution models (SDMs), which relate where a species occurs to environmental conditions, have become more accurate as ecologists adopted machine learning, but how these models arrive at their predictions has not necessarily become easier to understand. Modellers therefore appear to face a trade-off between accuracy and interpretability (Breiman 2001). Asking whether ecologists should give up interpretability for slight gains in accuracy is reasonable, and some argue that simpler models, which can be read directly, are the better choice (Rudin 2019). Ecology already used some tools for looking inside fitted models, but explainable AI (xAI), which brings such tools together, had not yet been applied to SDM research, according to the paper. The paper challenges the assumption that a complex model has to remain an uninterpretable ‘black box’.
The paper argues that explainable AI (xAI) lets ecologists keep complex species distribution models and still interpret them, both across the whole modelled area (global) and at single sites (local). It sorts interpretation methods into those that use simple, directly readable models and those applied afterwards to a fitted model, and sorts the latter into global and local ones. Among the tools that work with any algorithm, it lists permutation importance (how much accuracy drops when a variable is scrambled), partial dependence (the average effect of a variable on the predictions), Shapley values (each variable’s contribution to a single prediction) and LIME. LIME explains one prediction by fitting a simple model that mimics the complex one in the neighbourhood of that site (Ribeiro et al. 2016). A random-forest model for the African elephant, built purely as a demonstration, shows what this can reveal: the variable that matters most over the whole modelled area is not among the top variables listed at three randomly chosen sites.
The paper concludes that such explanations can help ecologists judge whether a fitted model is ecologically plausible, notice where a variable’s importance changes across space, and spot predictors that may not act causally before the model is used for extrapolation or management. It stresses that explaining a model is not the same as explaining real causal relationships, that a local surrogate only approximates the model it explains, and that the case for simpler models remains under debate.
References
- Breiman (2001). Statistical Modeling: The Two Cultures (with comments and a rejoinder by the author). Statistical Science 16, 199–231. doi:10.1214/ss/1009213726
- Rudin (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1, 206–215. doi:10.1038/s42256-019-0048-x
- Ribeiro et al. (2016). "Why Should I Trust You?": Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1135–1144. doi:10.1145/2939672.2939778