A hybrid model for agricultural pollution load forecasting by integrating meteorological data and soil properties: an XGBoost-LSTM approach
TL;DR
A hybrid forecasting model that fuses eXtreme Gradient Boosting for spatial feature importance evaluation with Long Short-Term Memory (LSTM) networks for sequential load prediction is proposed that provides a robust tool for proactive nutrient runoff management in data-sparse agricultural contexts.
Abstract
Accurate forecasting of agricultural non-point source pollution is pivotal for sustainable land management and environmental risk mitigation. However, the complex interplay between meteorological factors and heterogeneous soil properties introduces significant temporal and spatial variability into pollutant load estimation. This paper proposes a hybrid forecasting model that fuses eXtreme Gradient Boosting (XGBoost) for spatial feature importance evaluation with Long Short-Term Memory (LSTM) networks for sequential load prediction. Meteorological data—including rainfall, temperature, and evapotranspiration—are temporally aligned with high-resolution soil datasets, comprising texture, organic matter content, and infiltration rate. XGBoost is first employed to rank dominant environmental predictors, which are then dynamically fed into a time-aware LSTM model to capture non-linear temporal dependencies. Experiments conducted on a multi-year agricultural catchment dataset demonstrate improved prediction accuracy over standalone statistical and deep learning baselines, reducing RMSE by 17.6%. The framework provides a robust tool for proactive nutrient runoff management in data-sparse agricultural contexts.