Satellite-Informed Crop Yield Prediction System

The Challenge

An agricultural cooperative serving 3,000+ smallholder farmers had no reliable method for predicting crop yields before harvest. Farmers made input decisions (fertiliser, irrigation, pesticide) based on tradition and intuition rather than data. Over-application of nitrogen fertiliser was degrading soil health and increasing costs, while under-application in some fields reduced yields by up to 30%.

Our Approach

We built a crop yield prediction system that combines satellite imagery, ground-based sensors, and weather data to produce field-level yield forecasts updated weekly throughout the growing season.

Data Sources: Sentinel-2 satellite imagery (10m resolution, 5-day revisit) provides vegetation index time series (NDVI, EVI, SAVI). Ground-based IoT sensors deployed in representative fields measure soil moisture, temperature, and electrical conductivity at 30cm depth. ERA5 reanalysis data provides historical and forecast weather variables (precipitation, temperature, solar radiation, humidity).

Model Architecture: A temporal convolutional network (TCN) processes the multi-source time series for each field. The model was trained on 5 years of historical yield data from 1,200 fields with known outcomes. Input features include vegetation index trajectories, accumulated growing degree days, cumulative precipitation, soil moisture trends, and crop type.

Farmer Interface: Yield predictions are delivered weekly via a lightweight mobile app (React Native, offline-capable) in the local language. Each prediction includes the expected yield range, a comparison to the field's historical average, and actionable recommendations (e.g., "Current nitrogen levels are sufficient — skip the planned top-dressing application").

Results