Meta-modeling of a physically-based pesticide runoff model with a Long-Short term Memory approach

Surface water contamination by pesticides is widespread across the European Union (European Environment Agency, 2024). A primary pathway for pesticide transfer from agricultural fields to surface waters is surface runoff (Wauchope et al., 1995 [https://doi.org/10.1162/neco.1997.9.8.1735]; Louchart et al., 2001 [https://doi.org/10.2134/jeq2001.303982x]; Reichenberger et al., 2007 [https://doi.org/10.1016/j.scitotenv.2007.04.046]). This process is influenced by various spatial and temporal factors, including compound properties, topography, application date and methods, climate, soil properties, and agricultural practices (Shipitalo and Owens, 2003 [https://doi.org/doi: 10.1021/es020870b]). Richards-based models are valuable for predicting the temporal variability of pesticide runoff (Métayer et al., 2024 [https://doi.org/10.1016/j.scitotenv.2023.167357]), especially in regions with high rainfall intensity variability, such as the Mediterranean. However, their operational application is constrained by substantial computational demands and extensive data requirements. Meta-modeling approaches provide a means to reduce the computational time of an initial physically-based model. Among these, the Long Short-Term Memory (LSTM; Hochreiter and Schmidhuber, 1997 [https://doi.org/10.1162/neco.1997.9.8.1735]) model has demonstrated high efficiency in replicating hydrological (Kratzert et al., 2018 [https://doi.org/10.5194/hess-22-6005-2018]) and hydrochemical time series (Pyo et al., 2023 [https://doi.org/10.1016/j.wroa.2023.100207]), making them a promising meta-modeling strategy for pesticide runoff models. This study aimed to develop and evaluate a meta-modeling approach using LSTM models for a Richards-based model to simulate hourly variations in water and pesticide runoff over an entire year while minimizing computation times. The proposed approach was applied to a field-scale pesticide runoff model implemented in the fully spatially distributed hydrological model MHYDAS-Pesticide 1.0 (Crevoisier et al., 2021 [https://hal.inrae.fr/hal-04090048v1]) that integrates Richards and convection-dispersion equations, the uniform mixing cell concept, and an overland flow routine. This represents a challenge for at least the following three reasons: i) the time series contains mainly zero values of runoff discharge, ii) the prediction of pesticide runoff requires an efficient prediction of water runoff and ii) the order of magnitude of the targeted non-zero values of runoff concentration varies by several orders of magnitude. The LSTM meta-model was trained and validated using 560,560 annual time series simulations generated by the initial physically-based model. The training dataset comprised 70% of the simulations, with the remaining 30% reserved for validation. The resulting meta-model accounted for meteorological conditions, compound properties, and pesticide application date and rate. It demonstrated high accuracy in simulating hourly runoff and pesticide concentrations, achieving significant reductions in computation time. However, challenges remain, such as improving the precision of runoff occurrence simulation and enhancing the meta-model's generalizability by incorporating additional static parameters as inputs.

The poster will focus on the methodology for the meta-model’s development and the results of its evaluation. The meta-model has been implemented within a fully spatially distributed physically-based hydrological model, MHYDAS Pesticide, to form an hybrid version.