Multi-scale optical scanning holography for robust manipulation detection in satellite images

Recent progress in generative AI has made it increasingly easy to produce highly realistic satellite imagery. While this advancement enables a wide range of beneficial applications, it also raises serious concerns regarding data reliability, misinformation, and security in critical remote sensing scenarios. Detecting manipulated satellite images remains a challenging task, as they lack human-related cues and instead exhibit complex spatial structures, where artifacts are often subtle and widely distributed. This paper introduces a novel multi-scale framework based on Optical Scanning Holography (OSH) to address this challenge. OSH is employed as a physics-inspired transformation that maps input images into a more informative feature space, where spatial, frequency, and phase characteristics are jointly represented within a unified domain. To further improve feature extraction, OSH is applied at multiple scales to generate complementary representations, which are combined into a multi-channel input. A convolutional neural network (CNN) is then utilized to learn discriminative patterns from this enhanced feature space. Experimental results on a large-scale dataset of real and AI-generated images demonstrate the effectiveness of the proposed framework, achieving an accuracy of 99.31%.