Crop Yield Prediction in Sub-Saharan Africa Using Satellite Imagery
We develop a multi-modal ML model combining MODIS satellite data with weather station readings to predict maize yield 60 days before harvest. Tested across Kenya, Tanzania and Ethiopia.
Executive Summary
Predicting crop yields early and accurately is crucial for food security, supply chain management, and policy decisions in Sub-Saharan Africa.
This report presents a deep learning framework designed to predict crop yield using free satellite imagery (MODIS) combined with temperature and precipitation anomalies.
Methodology
We trained a Convolutional Neural Network (CNN) combined with a Long Short-Term Memory (LSTM) network to capture both spatial information (satellite indices like NDVI) and temporal information (weather sequences).
# Model architecture overview
# CNN extractor + LSTM sequential layer
Results
Our model achieved an $R^2$ score of 0.76 on the test set, predicting maize yield 60 days before harvest.