A research team from Wuhan University has developed a new artificial intelligence framework that can reconstruct hidden or partially obscured objects in satellite imagery with greater accuracy than existing methods. The framework, called Remote Sensing Amodal Completion (RSAC), uses diffusion-based generation combined with remote-sensing-specific structural guidance to infer complete object shape, surface texture, and semantic identity from incomplete observations. This advance addresses a critical challenge in geospatial artificial intelligence, where objects in satellite images are often obscured by cloud cover, overlapping structures, or imaging angles.
Published in the Journal of Remote Sensing on April 7, 2026, the study proposes a Dual-Adaptive Diffusion-Based Framework that shifts from scene-level inpainting to object-level reasoning. The method adapts Stable Diffusion to the remote sensing domain using Low-Rank Adaptation (LoRA) and a four-channel ControlNet that uses image and mask information to guide structural completion. A prior-enhanced initialization strategy preserves low-frequency information from the visible part of the object, improving physical consistency over approaches that start from random noise.
The researchers built a dedicated RSAC dataset containing 1,770 annotated instances across 10 categories of typical remote sensing objects, including planes, ships, large vehicles, storage tanks, roundabouts, tennis courts, basketball courts, baseball diamonds, soccer ball fields, and ground track fields. The dataset included 1,235 training images and 535 testing images. In comparative experiments against methods such as Stable Diffusion Inpainting, LaMa, BrushNet, and Open-World Amodal Appearance Completion, the proposed approach achieved 100% valid-output coverage, an Intersection over Union (IoU) of 0.853, an amodal completion IoU (ACIoU) of 0.688, a mean squared error (MSE) of 11.822, a peak signal-to-noise ratio (PSNR) of 24.799 dB, and a structural similarity index (SSIM) of 0.930. These results indicate more accurate object geometry, clearer boundaries, and more realistic texture continuity.
The research team emphasized that the goal was not simply to make incomplete satellite images look visually complete, but to help machines infer what an object is and how it should be structured. By integrating generative models with remote-sensing-specific constraints, the framework points toward more reliable object-level reasoning for geospatial AI under real-world occlusion. The technology could support more reliable geospatial intelligence in scenarios such as post-disaster assessment, infrastructure mapping, automated cartography, facility reconstruction, and urban monitoring. By restoring complete object morphology from partial observations, RSAC may also improve training data for detection models and help AI systems interpret satellite imagery more like human analysts.
The team first constructed a high-quality object dataset from remote sensing instance segmentation resources, using blind image quality assessment and expert screening. Complete objects were paired with simulated incomplete versions generated by random masks. The model then combined LoRA-based domain adaptation, ControlNet-based task conditioning, and prior-enhanced diffusion initialization. Its performance was evaluated with both structural metrics, including IoU and ACIoU, and texture metrics, including MSE, PSNR, and SSIM. Future studies may extend the framework to more object categories, dynamic drone perspectives, full three-dimensional reconstruction, and multitemporal or multimodal remote sensing data. The study is available at https://doi.org/10.34133/remotesensing.1035.
