Remote Sensing Image Tampering Dataset

Traditional image forgery localization methods struggle when applied to remote sensing imagery due to insufficient feature extraction, low accuracy, high computational cost, and poor performance near blurred tampered boundaries. To address these challenges, we propose FASDNet, a multi-module detection framework that synergistically processes RGB images and DWT-enhanced frequency features using a dual-encoder architecture with ResNet18 and MobileNetV2 to extract and cascadedly fuse semantic and texture features. Spatial attention is introduced to highlight suspicious regions, while a refined ASPP module integrates multi-scale context for high-resolution mask generation. Experiments show that FASDNet outperforms existing methods, with gains of 2.65% in F1-score and 4.29% increase in IoU, and demonstrates strong robustness against JPEG compression, filtering, and scaling attacks. Ablation studies validate each component\u2019s contribution.