Explainable Feature Engineering for EO Data Analysis

Explainable feature engineering is essential for enhancing the interpretability of Earth Observation (EO) data analysis, particularly in hyperspectral imaging, where large volumes of spectral and spatial information are processed for material identification. Traditional spectral unmixing techniques, such as Nonnegative Matrix Factorization (NMF) and Non-Negative Least Squares (NNLS), effectively decompose hyperspectral images but fail to provide insights into the contribution of individual features. To address this, we propose an explainable spectral unmixing framework that integrates SHapley Additive Explanations (SHAP) to quantify feature importance in hyperspectral data analysis. The results confirm that the proposed approach effectively extracts end-member spectra and abundance maps, while offering deeper insights into the spectral and spatial dependencies influencing feature importance. In EO datasets, the framework enhances land-cover classification and environmental monitoring by distinguishing vegetation, soil, and water. In lunar datasets, it successfully identifies mineralogical compositions, supporting geological interpretation and planetary exploration. By integrating explainability into feature engineering, this study ensures scientific transparency, improves model interpretability, and enhances decision-making in hyperspectral remote sensing applications.