Interpretability of distribution models of plant species communities learned through deep learning

Very recently, plant species distribution models based on convolutional neural networks have appeared. Although they have proven to be more efficient than state-of-the-art models (such as MaxEnt, Random forest, boosted trees, etc.), especially on poorly represented species, they are often criticized for their lack of interpretability. We present here the first results of a study that aims to clarify the predictive power of these new methods by linking the information contained in the input data to the quality of the response using ablation experiments.