Explainable SVM for feature selection in Crop Monitoring

Abstract

Reliable crop monitoring from Earth-observation imagery needs models that agronomists can trust. We present an explainable SVM pipeline that converts multispectral Sentinel-2 (and optional UAV) data into decisions and reasons. Spectral bands and vegetation indices are stacked; weak labels are seeded with a lightweight GMM when ground truth is scarce; a polynomial-kernel SVM is trained; and the decision function is decomposed into monomials to deliver (i) global feature rankings and (ii) per-pixel attributions. This exposes both main effects and band–band interactions (e.g., red-edge × NIR) that drive class separation, enabling targeted feature selection and simpler sensor configurations. The approach achieves competitive accuracy with a compact, interpretable feature set and produces maps that explain why each pixel is classified. The workflow is simple, reproducible, and ready for operational crop scouting.