Band-based interpretability with SHAP for hyperspectral classification

Official URL: http://dx.doi.org/10.1109/SIU59756.2023.10223795

Hyperspectral data enable high classification accuracies due to the large amount of spectral information they contain. However, the complete black box or at least partially opaque nature of machine learning approaches often make classification processes challenging to interpret and low in explainability. In this work, SHAP, which is an explainable artificial intelligence (XAI) method, is used for the interpretability of hyperspectral classification. For each class, band-based SHAP values are obtained over the trained classification model, providing a quantitative evaluation of the contribution of the spectral bands to the classification process. Preliminary experimental results are provided in this paper, and shed light on the proposed method's potential.