2022 ESA Annual Meeting (August 14 - 19)

COS 138-4 Habitat phenology has a lot to offer to deep species distribution models (Deep-SDMs).

8:45 AM-9:00 AM
518A
Joaquim Estopinan, INRIA, LIRMM;Maximilien Servajean,LIRMM, Paul Valéry Montpellier University;Pierre Bonnet,CIRAD, UMR AMAP;Alexis Joly,INRIA, LIRMM;François Munoz,LIPHY, Grenoble Alpes University;
Background/Question/Methods

Exploiting the right ecological variables driving species presence is key to capture true distributions and enhance conservation science. Recent satellite missions provide global multi-spectral data of high spatial and temporal resolution. Coupled with precise biodiversity data including citizen observations, there is an unprecedented opportunity to train global SDMs fed with rich high-dimensional inputs. Moreover, deep neural networks architectures can learn and generalize from complex spatio-temporal patterns. Our goal is to prove the efficiency of such global Deep-SDMs as a means to seize species ecological niche. We will illustrate our point with an ambitious Sentinel-2 twelve-month-long image time-series dataset built around orchid occurrences spread worldwide (almost 1M observations after taxonomic and geographic filtering, fourteen thousand species). What is the contribution of landscape phenology to SDM’s predictive power? Is the information crystallized at species level helpful to estimate IUCN statuses?We tailored a method comparing SDM performances when its input’s temporal dimension was randomly permuted, sampled or averaged. An adaptation of the Inception v3 convolutional neural network is being used. Regarding orchids risk extinction, we began to add complementary SDM entries: lon./lat. coordinates, human pressure, elevation, bioclimatic variables and ecoregions. We compared risk categories classification from species model features to existing methods.

Results/Conclusions

Compared to a model trained with one single-view of species habitat only, having access to habitat phenology increases macro-average top-30 accuracy by 92 %. Model capacity to discriminate habitats worldwide is sharper. Interestingly, performances in occurrence-poor and species-rich regions are especially improved. Species features are condensing the high dimensional input and the SDM behaves then as an expressive information extractor. IUCN status classification is promising. Efforts are invested to develop a scalable and interpretable framework where this work could be easily reproduced on another taxon associated with a set of geolocated occurrences. In the face of climate change and other anthropogenic threats, we believe that our method has great potential to capture the fine-grained ecological preferences and vulnerabilities of species and thus enable informed decision making.