A research team from Wuhan University has developed a new artificial intelligence framework that significantly improves the restoration of partially hidden objects in satellite imagery. The method, detailed in the Journal of Remote Sensing (DOI: 10.34133/remotesensing.1035), addresses a critical challenge in geospatial AI: how to infer complete objects from fragmented observations caused by cloud cover, overlapping objects, or imaging angles.
Rather than simply filling missing pixels, the framework integrates diffusion-based generation with structural guidance specific to remote sensing. It adapts Stable Diffusion using Low-Rank Adaptation (LoRA) and employs a four-channel ControlNet to maintain geometric integrity. A prior-enhanced initialization strategy preserves low-frequency information from visible parts, avoiding the randomness of starting from noise. In tests against methods like Stable Diffusion Inpainting and LaMa, the proposed approach achieved superior results: 100% valid-output coverage, an Intersection over Union (IoU) of 0.853, and a structural similarity index (SSIM) of 0.930.
The study introduces Remote Sensing Amodal Completion (RSAC) as a dedicated task and builds a dataset of 1,770 annotated instances across ten object categories, including planes, ships, and sports fields. The framework not only restores visual appearance but also aids downstream tasks like object detection and vision-language model interpretation. As noted by the research team, the goal is to help machines infer what an object is and how it should be structured, enabling more reliable geospatial intelligence in disaster response, urban planning, and automated mapping.
Supported by the National Natural Science Foundation of China (grants 42422109 and 42371366), the work points toward future extensions to more object categories, dynamic drone perspectives, and 3D reconstruction. The Journal of Remote Sensing, an open-access publication associated with AIR-CAS, promotes interdisciplinary research in earth and information science.
