Hybrid Ensemble-LSTM Framework for Crop Yield Prediction Using Ground-Based Sensor Data

Abstract

Horticultural crop production prediction prior to harvest is challenging due to a lack of ground-based data. Utilising a machine learning architecture that integrates soil moisture dynamics, meteorological causes, and management approaches, the yields of tomatoes and turmeric are anticipated. A thousand samples were gathered over the course of three growth seasons in southern Karnataka using industrial-grade sensors. Four baseline models—Random Forest, XGBoost, Extra Trees, and LSTM— were evaluated using time-series cross-validation. For tomato and turmeric, XGBoost achieved the best baseline performance (R 2 =0.79, RMSE=3.94t/ha and R 2 =0.84, RMSE=1.31t/ha, respectively). For tomatoes (R 2 =0.85, RMSE=3.41t/ha) and turmeric (R 2 =0.88, RMSE=1.09t/ha), the best results were obtained by a hybrid stacking model that included en- semble and LSTM. SHAP identified soil moisture variability and rainfall timing as the primary culprits. The 77 farmers' feedback showed 8.