A Deep Learning Model for Corn Yield Prediction Using Spatial and Temporal Features

Abstract

Maize is a staple food crop in Zambia, making accurate yield prediction essential for food security and agricultural planning. This study presents an advanced deep learning approach for maize yield prediction that integrates Sentinel-2 satellite imagery with climate data. We develop a scalable and interpretable hybrid CNN-LSTM model to capture both spatial and temporal patterns of crop growth. The CNN component extracts spatial features from Sentinel-2 multispectral images (including vegetation indices such as NDVI and EVI), while the LSTM component learns temporal dynamics from time-series climate variables (rainfall, temperature, humidity). The model is trained and validated using historical yield records from major maize-growing regions in Zambia, demonstrating high predictive accuracy and outperforming traditional yield estimation methods. Accurate yield forecasts from this model enable early warnings of potential crop shortfalls, allowing farmers to take timely action to mitigate losses. Additionally, the predictions provide policymakers with insights for managing grain reserves, market supply, and food security strategies. By leveraging deep learning and remote sensing, this work offers a decision-support tool that contributes to sustainable agricultural practices and climate resilience in SSA, bridging the gap between academic and practical applications.