This presentation introduces a deep learning framework for spatio-temporal modeling of airborne pesticide exposure in France. The work builds on the analysis of heterogeneous and partially censored environmental monitoring data, with a particular focus on the CNEP campaign.
We construct enriched input representations that combine agricultural context, pesticide purchase data, pedoclimatic variables, and spatial encodings designed to capture local variability. These representations are then used to develop joint deep estimators adapted to sparse and censored observations. The proposed models are evaluated through dedicated validation protocols, including an assessment of extrapolation risk.
Beyond predictive performance, the presentation addresses the reliability and trustworthiness of the resulting estimates. It investigates monotonicity violations caused by perturbations of input variables and introduces a conceptual interpretability pipeline to better understand model behavior.
Overall, this work argues for hybrid AI-for-Science systems that combine data-driven learning with expert knowledge, both in the design of representations and in the constraints imposed during training, in order to produce more robust and reliable exposure estimates.