Advancing Precision Conservation through Machine Learning for Field-Level Prioritization of Agricultural Best Management Practices

Effective identification of opportunities for agricultural Best Management Practices (BMPs) is essential for reducing nonpoint source pollution in intensively managed agricultural watersheds globally. In the Great Lakes basin, this task is complicated by widespread tile-drained agriculture. Existing decision-support tools often generate an overwhelming number of potential intervention sites without clear prioritization, which limits practical implementation. The Agricultural Conservation Planning Framework (ACPF) systematically maps potential BMP locations but provides limited guidance for prioritizing sites. To address this limitation, the present study developed and applied a framework that integrates ACPF with machine learning and explicit uncertainty mapping to predict a continuous, field-level measure of conservation need. A survey of 138 agricultural fields in a 205 km 2 tile-drained watershed in Southern Ontario was conducted to derive a Composite BMP Need Score (CBNS), integrating field-observed evidence of BMP need, site severity, and land management indicators. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models were trained to predict CBNS using geospatial predictors. Both models exhibited comparable performance, with stable and consistent predictions across the independent test set (n = 35) and 100 repeated train-test splits (RMSE = 1.32 ± 0.12 CBNS units on a 0.5–6.5 scale). Learning curves indicated that performance was primarily limited by sample size rather than model capability. XGBoost was selected for spatial prediction due to its greater flexibility and enhanced ability to differentiate among fields, and was used to estimate CBNS for all 627 agricultural fields in the watershed. Model interpretation identified hydrologic connectivity and erosive force as the strongest drivers of conservation need. The framework identified the top 20% of agricultural fields requiring the most interventions. By combining machine learning with ACPF and incorporating uncertainty layers, this study advances precision conservation from conceptual understanding to prioritization guidance, offering a transparent and transferable framework applicable to agricultural watersheds worldwide, particularly in tile-drained landscapes.