A Machine Learning-based Surrogate Model for Optimization of Cropping Systems in Denmark

Process-based agricultural system models (PBMs) are pivotal tools for evaluating the environmental impacts of agricultural practices. However, their large-scale application is constrained by significant computational demands, extensive time requirements, and data availability. These challenges hinder policymakers and land managers in implementing sustainable agricultural practices at scales meaningful for decision-making. Recent advancements in machine learning (ML) offer a promising solution by providing computationally efficient alternatives, yet the lack of interpretability regarding agro-environmental processes remains a critical barrier.

In this study, we address this challenge by developing a machine learning-based surrogate model for LandscapeDNDC (LDNDC) framework. The surrogate model predicts key agro-environmental variables, including yield, nitrous oxide (N₂O) emissions, nitrate leaching (NO₃^-), and soil organic carbon (SOC), at a national scale for Denmark. Synthetic data were generated using a factorial design based on observed crop practices in Denmark, utilizing field-level data collected across six Danish catchments between 2013 and 2019 as part of the National Monitoring Program for Water Environment and Nature (NOVANA; LOOP-program). This approach incorporated crop rotations as well as spatially disaggregated information on soils and weather, resulting in a dataset comprising approximately 2 billion rows. To enhance the dataset's versatility and account for potential future scenarios, factors like manure amount and synthetic fertilizer amount were extrapolated beyond its current observed ranges. The synthetic dataset was subsequently simulated using the LDNDC modelling framework, and the resulting outputs were employed to train a variety of machine learning algorithms utilizing multi-task learning, optimizing predictions for multiple agro-environmental variables of interest.

Our results demonstrate that the ML-based surrogate model not only significantly reduces computational cost and processing time but also enables the exploration of multiple cropping scenarios with greater efficiency. This approach facilitates rapid scenario testing and optimization, making it accessible to policymakers and farmers without the constraints imposed by traditional PBM frameworks. We propose this methodology as a scalable and practical tool for advancing sustainable agricultural decision-making.