Abstract
Advancing our understanding of environmental interactions in rice crops contributes to food production in water-limited regions. This paper proposes an integrated crop modeling architecture, demonstrating how machine-learning (ML) models enhance classic Mechanistic Crop Modeling (MCM) estimations by learning directly from environmental data. Here, we quantify the impact of noise-induced uncertainty on the CERES-Rice crop growth model, particularly relevant for drought-tolerant varieties that exhibit complex adaptation mechanisms, such as Nerica 4. Environment characterization is achieved through a novel 3D Gaussian Mixture Model (GMM), offering enhanced precision and scalability when coupled with remote-sensing satellite-derived environmental data. By coupling both MCM and ML models, we achieved higher estimations for grain yield (R 2 = 0.99) and biomass (R 2 = 0.8) in the northwest Tambacounda region of Senegal in Africa, providing reliable estimates of 30% grain conversion efficiency and 2.18kg/hamm water use efficiency from an environment characterized by sandy soils with high saturated hydraulic conductivity (1.1cm/h) and the lowest regional precipitation (513mm, 49%).