Smart farming: Real-time rice yield forecasting on mobile devices using lightweight CNN-LSTM

This work presents a framework for accurate and punctual in-season crop yield estimation at high spatial resolution for Indian farmers through the utilisation of low-resource edge devices by reducing the CNN-LSTM model neural activations' memory requirements. We propose a new memory optimisation approach—Clustering and Compression (C²)—that is tailored to combating the large memory needs of CNN-LSTM architecture neural activations. Through combining spatial feature extraction and temporal learning, the model acquires efficient spatiotemporal representations. It is trained on high-resolution block-level yield data, satellite-delivered Normalized Difference Vegetation Index (NDVI) and Normalized Difference Moisture Index (NDMI), and Jammu region weather data. Optimized CNN-LSTM comprehensively surpasses performance of baseline CNN and LSTM models while minimising memory usage by orders of magnitude—especially in neural activations. This optimisation allows for cost-effective, cloud-independent on-device inference and routine model training, which are essential for handling the day-to-day environmental fluctuations in the dynamic climates. In summary, the proposed method allows for a novel neural activation memory optimisation technique that facilitates device-local high-resolution crop yield estimation, paving the way for sustainable and strong agriculture for smallholder farmers.