Abstract
Species distribution models link the geographic distribution of a species to its environment. The objectives of using these models are multiple. They can be used to extract knowledge on species and their environmental preferences, to help with conservation plans and policies, to monitor and anticipate the spread of invasive species, or to simulate environmental changes and their impacts on species. To best meet these objectives, it is necessary to design efficient, accurate and interpretable models. Most of the models used today are relatively simple models. These models have the advantage of being easy to interpret by producing simple relationships between a species and its environment. However, they often share some shortcomings such as sensitivity to overfitting, which requires a careful choice of descriptive data of the environment to avoid interpretation errors. Models based on machine learning approaches have shown performances that are often as good or even better, with a stronger robustness against overfitting. However, these methods are more often criticized for their lack of interpretability. This is the case with convolutional neural networks whose first experiments have shown promising results for their use in species distribution modeling. Convolutional neural networks are known for their particularly high performance in all image processing tasks (classification, object detection, counting, etc.). They have the particularity of being able to use very large data with little risk of overfitting. Even more than other machine learning approaches, these models are often described as black boxes that are difficult to interpret. We propose to study the use of these models, called Deep-SDMs, in the context of species distribution prediction with a particular attention to interpretation in order to highlight the potential interests of this new approach while trying to clarify the mechanisms involved.We present the use and analysis of Deep-SDMs with several interpretability experiments in different contexts. We conduct comparisons on some aspects with more state-of-the-art models. We propose qualitative and quantitative analyses on the interpretation of Deep-SDMs learning. In particular, we propose to study what the model captures, either by analyzing the differences in performance according to the data used and the information they contain, or directly by studying the learned representation space of the model (the last layer of the model).Overall we show that it is possible to analyze and interpret model learning in several ways, leading to interesting ecological conclusions. We show an interesting potential of Deep-SDMs that allow: (1) to learn a single model for many species simultaneously and using observation data without absence data, (2) to use more complex and richer representations of the environment thanks to their ability to use very high dimensional data, (3) often better performances than other models, especially on rare species, (4) learning on a very large scale (thousands of species and regions the size of countries) and at a very fine resolution (around ten meters) thanks to remote sensing data, and (5) possible reuse of models in similar contexts, taking advantage of the learning already done.