From Analytical Detection to Spatial Prediction: LC–MS and Machine Learning Approaches for Glyphosate Monitoring in Interconnected Land–Soil–Water Systems

The widespread application of glyphosate—the world’s most used herbicide—presents a significant environmental challenge due to its persistence and mobility within interconnected land–soil–water systems. This study addresses the limitations of traditional, discrete water monitoring by developing a predictive framework for glyphosate and its primary metabolite, aminomethylphosphonic acid (AMPA), in the agricultural context of Apulia, Southern Italy. The methodology integrates high-sensitivity analytical chemistry with advanced spatial intelligence. Water samples were analyzed using an optimized UHPLC–MS/MS framework with pre-column derivatization (FMOC-Cl), achieving an ultra-trace Limit of Quantification (LOQ) of 0.025 μg/L. To transition from point data to continuous spatial profiles, a hybrid Machine Learning (ML) architecture was implemented. The model utilized a suite of geospatial predictors, including land use (Corine Land Cover), Digital Elevation Models (DEMs), and slope characteristics extracted from river offset lines. A dual-modeling strategy was employed: Global Models (Random Forest, Gradient Boosting, and KNN) for regional trends and Individual Models for river segments exhibiting sufficient internal variability. Analytical findings (2018–2024) revealed that AMPA consistently exhibited higher mean concentrations than glyphosate, reaching peaks of 9.27 μg/L. This trend is primarily attributed to its superior environmental persistence and a half-life of up to 240 days, compared to the parent compound. Spatiotemporal analysis identified critical peaks in the second quarter for glyphosate and extreme surges in the fourth quarter for AMPA, particularly in the Cervaro basin. The Random Forest Regressor emerged as the most robust predictive tool, achieving a coefficient of determination (R2) of approximately 0.68 at the global scale and up to 0.75 for localized models where data density was sufficient. The integration of ML frameworks allows for the identification of contamination “micro-hotspots” and the mapping of probabilistic pollutant distribution along entire river reaches without additional sampling costs. This high-fidelity diagnostic tool provides a cost-effective strategy for environmental agencies to implement targeted mitigation and proactive water resource protection in Mediterranean agroecosystems.