부산대학교 도서관

Satellite imagery and machine learning (ML) techniques for land use land cover (LULC) mapping have become prevalent in remote sensing (RS). However, these developed models usually need comprehensive explanations of model functioning and limitations. Therefore, this study aims to establish a systematic ML model training process for LULC mapping, encompassing model interpretation and performance evaluation of the new data. The study compares three ML models - Random Forest, extreme gradient boosting (XGB), and light gradient boosting - for classifying a LISS IV satellite image. The optimal algorithm was selected based on random and spatial cross-validation (CV) outcomes, and model interpretation was achieved using Shapely (SHAP) values. Dissimilarity index explicated model predictions on new data. Results indicate XGB’s superior accuracy across both CV methods. SHAP values effectively explained data behaviour in XGB model predictions. Ultimately, this systematic workflow established a standard for LULC mapping with RS imagery using ML.