Raphaël Benerradi is a second-year PhD student in the IROKO team at Inria, based at the LIRMM laboratory in Montpellier, supervised by Alexis Joly, Christophe Botella and Maximilien Servajean. Trained in computer science and mathematics (CentraleSupélec) and agronomy (AgroParisTech), he works at the interface of statistics, machine learning and ecology. Pl@ntNet interviewed him about his research; here is what matters most for anyone reusing the data on this portal.
The question
Where are plant species present, and how does that change over time? Raphaël builds models that estimate the probability of presence of a species across space and time, using the opportunistic observations collected through Pl@ntNet — the very records shared on GBIF.
The catch: sampling bias
Pl@ntNet observations are extremely numerous, but they follow no sampling protocol. People photograph plants wherever and whenever they like, which over-represents cities, tourist spots and springtime, and under-represents other places and seasons. Any naive map drawn from raw counts mostly shows where the observers were.
The approach
His models separate two processes: an ecological one, describing why a species occurs somewhere (climate, land use…), and an observation one, describing the probability that it is actually detected by users depending on location, season and other factors. These occupancy models already existed for data collected under strict protocols; the work adapts them to Pl@ntNet’s context and scales them up with neural networks to handle millions of records.
The goal is not to replace botanical surveys, which remain higher-quality but costly and local, but to offer a complementary approach able to produce results at a much larger scale.
Take-away for data users
If you download Pl@ntNet occurrences from this portal to model distributions, account for observer effort. GBIF occurrence counts of all species in an area are a useful proxy for it — and the models described here are one principled way to use them.
Read the full interview on plantnet.org.