Uzaktan Algilama Sahne Siniflandirmasinda CAM Tabanli A iklanabilirlik Y ntemlerinin Kapsamli Analizi

The interpretability of decision-making processes in deep learning models has increased interest in explainable artificial intelligence (XAI) methods, especially in domains where reliability is critical, such as remote sensing. However, a review of the literature reveals that the effectiveness of widely used Class Activation Mapping (CAM)-based XAI techniques in the context of remote sensing has not been systematically and comparatively examined. The main objective of this study is to comprehensively evaluate the reliability and explanation quality of CAM-based XAI methods across four remote sensing datasets (WHU-RS19, EuroSAT, SEN12MS, BigEarthNet) that feature varying scene types, resolution levels, and object complexities. A total of 11 different CAM methods were applied, and the quality of the resulting explanations was quantitatively assessed using 6 evaluation metrics. This study goes beyond traditional approaches that rely solely on visual attribution maps by offering a holistic evaluation that considers data types and content diversity, contributing to the development of more reliable XAI methods in the field of remote sensing.