Lightweight Deep Learning for Agricultural Loss Forecasting: GRU with Transfer Learning and Edge Deployment

Precise forecasting of post-harvest deterioration in perishable crops like cassava is essential for minimizing food waste, enhancing supply chain effectiveness, and aiding decision-making in agricultural systems. This study introduces an efficient, interpretable forecasting system founded on Gated Recurrent Units (GRUs), designed for implementation in low-resource settings defined by scarce data, operational noise, and restricted computational power. The suggested method identifies short- to medium-term time dependencies in multivariate sensor data and integrates interpretability techniques such as SHAP, saliency maps, and LIME to deliver feature attribution throughout time steps. A two-phase transfer learning approach is used to improve generalization from high-resource to low-resource settings, tackling data shortages in smallholder agricultural situations. Experimental assessments juxtapose the GRU with conventional (ARIMA, XGBoost), recurrent (LSTM, BiLSTM), and transformer-based (Temporal Fusion Transformer, Informer) benchmarks, where the GRU records the minimum MAE (4.26%), RMSE (7.88%), and the maximum R ² (0.884). On-device benchmarking validates real-time capability, achieving sub-10 ms latency on Raspberry Pi 4, under 100 ms latency on ESP32, and a model size below 512 KB post-quantization. The findings show that the suggested GRU provides an efficient balance among predictive accuracy, interpretability, and computational efficiency, facilitating practical field application for forecasting post-harvest losses in resource-limited agricultural settings.