Classification of Jilin-1 GP01 hyperspectral image using machine learning techniques with explainable artificial intelligence

Abstract

Machine learning (ML) techniques have been significant potential for the image classification; however, they behave as a black box because of the use of unknown descriptors in model construction. Thus, explainable Artificial Intelligence can assist with comprehending the prediction process of a model.  In this study, XgBoost and Random Forest were utilized to generate LULC maps for the Özbağ district of Krışehir, a highly forested through the valley, using Jilin-1 GP01 hyperspectral image. Accordingly, the overall accuracies of thematic maps produced by XgBoost and Random Forest were estimated as 93.17% and 91.98%, respectively. Moreover, the Shapley additive explanations (SHAP) technique is employed to understand the output of the models. After SHAP analysis of the ML models, the feature importance of each spectral band was determined. Therefore, given the trained by both algorithms, Band 7 was determined the most important of the hyperspectral bands used in this study. According to the Shapley values, band 5 in the Xgboost model and Band 7 in the random forest model are efficient in class-based evaluations for identifying the bare soil class with the highest F-score value.  Although the differences were obtained in the SHAP analysis according to some spectral bands since the working principles of the classification algorithms are different.