École Nationale Supérieure de
Techniques Avancées

Abstract

Deep species distribution models have recently shown superior predictive capabilities to classical methods used in ecology, especially when trained on very large volumes of occurrence data. One problem with these models, however, is that they can only predict the probability of a species conditional on an observation but they cannot predict species assemblages. In this study, we explore adaptive thresholding methods whose objective will be to predict more relevant assemblages than naive methods such as absolute thresholding or the k most probable species. We rely on statistical tests rejecting a null hypothesis, and a posteriori probability calibration method.