The Performance and Potential of Deep Learning for Predicting Species Distributions

ABSTRACT Aim Species distribution models (SDMs) address the whereabouts of species and are central to ecology. Deep learning (DL) is poised to further elevate the already significant role of SDMs in ecology and conservation, but the potential and limitations of this transformation are still largely unassessed. Location North America. Time Period 2009–2021. Major Taxa Studied Mammals, amphibians, reptiles, ants, and butterflies. Methods We evaluate DL SDMs for 2299 terrestrial vertebrate and invertebrate species at continental scale and 0.00833° resolution in a like‐for‐like comparison with latest implementation of conventional SDMs. We compare two DL methods (a multi‐layer perceptron (MLP) on point covariates and a convolutional neural network (CNN) on geospatial patches) against non‐DL SDMs (Maxent and Random Forest), emphasising fair comparison and induced pitfalls from information leakage, species imbalances, and location biases. Results On average, DL models match, but do not surpass, the performance of non‐DL methods. DL performance is moderately to substantially weaker for species with narrow geographic ranges, fewer data points, and those assessed as threatened and hence often of greatest conservation concern. Furthermore, information leakage across dataset splits substantially inflates performance metrics, especially of CNNs. Main Conclusions Our results indicate that biases known from conventional SDM settings are strongly amplified for DL models. Although recent advancements in DL draw promising new avenues for ecological process modelling, their benefits beyond improved numerical performance can only be met when pitfalls are accounted for. Realising the potential of DL in its entirety will thus require a closer collaboration between ecology and machine learning disciplines.