Multi-Scale Temporal Feature Decoupling for Climate-Robust Agricultural Machine Learning Models

Abstract

Reliable yield prediction under changing climatic conditions continues to be a major bottleneck in precision agriculture because climatic variability directly influences crop productivity and food-security planning. Conventional machine learning models trained on raw meteorological time-series frequently conflate short-term synoptic weather anomalies, intra-annual seasonal fluctuations, and multi-decadal climatic trends into a single undifferentiated signal. This entanglement of temporal scales degrades model generalization across crop-years and geographies, producing inflated in-sample accuracy that masks poor out-of-distribution performance. This paper proposes a Multi-Scale Temporal Feature Decoupling (MSTFD) framework that explicitly separates each climate covariate into three orthogonal temporal components before any model training: a long-term trend estimated via LOESS smoothing, an intra-annual seasonal component extracted through Seasonal-Trend decomposition using LOESS (STL), and a short-term residual capturing synoptic anomalies. The framework is applied to 33 years of weekly temperature, precipitation, and humidity records from the NASA POWER Agroclimatology archive and FAO AQUASTAT, covering twelve agro-climatic zones and five staple crops: wheat, rice, maize, soybean, and sorghum. Per-component summary statistics and agroclimatic indicators are engineered and supplied to five predictive models: Random Forest, Support Vector Regression, XGBoost, LSTM, and a CNN-LSTM hybrid. A rigorous leave-one-year-out (LOYO) and leave-one-zone-out (LOZO) evaluation protocol simulates realistic deployment conditions. Results demonstrate that MSTFD features reduce RMSE by an average of 18.4% and improve $\mathbf{R}^{\mathbf{2}}$ by 0.09 compared to raw-feature baselines across all five models. The CNN-LSTM variant achieves $\mathbf{R}^{\mathbf{2}}=\mathbf{0. 8 7}$ in cross-year validation and $\mathbf{R}^{\mathbf{2}}=\mathbf{0. 8 3}$ in cross-region generalization. These findings establish multi-scale temporal decoupling as an effective, model-agnostic strategy that simultaneously improves out-of-distribution accuracy and provides agronomists with interpretable, component-level attribution of yield variance.