This is like having a smart weather and crop advisor that looks at satellite images and sensor data over time and then predicts how much maize you will harvest in each field, learning which colors and patterns in the images matter most at different growth stages.
Manual, late, and inaccurate estimation of maize yield at field or regional scale, which limits planning for input use, logistics, pricing, and risk management. The system uses remote sensing and machine learning (support vector regression) to forecast yields earlier and more accurately.
Domain-specific agronomic feature engineering (e.g., crop phenology-aware spectral indices over time), access to high-quality yield ground truth and multi-year satellite data, and model calibration for specific geographies create a data and expertise moat that is hard to replicate quickly.
Classical-ML (Scikit/XGBoost)
Structured SQL
Medium (Integration logic)
Access to consistent, cloud-free remote sensing time series and representative ground-truth yield data; model retraining and recalibration for new regions and seasons may be required for robustness.
Early Majority
Compared with generic yield models, this work explicitly analyzes how the model’s sensitivity to different spectral bands and spatial/temporal features evolves over the growing season, improving interpretability and potentially enabling more robust feature selection and timing of data acquisition.