From Incomplete Views to Complete Objects: Amodal Completion for Semantic and Structural Restoration in Remote Sensing

Occlusions in satellite imagery pose significant challenges for accurate interpretation of geospatial objects. While existing inpainting methods focus on visual plausibility, they often neglect structural fidelity and semantic reasoning. To address this, we formally define the task of Remote Sensing Amodal Completion (RSAC) and propose a Dual-Adaptive Diffusion-Based Framework. This framework integrates Low-Rank Adaptation for domain specialization, a 4-channel ControlNet for mask-guided structural inference, and a prior-enhanced initialization strategy to preserve physical consistency. A dedicated RSAC dataset with 1,770 annotated instances across 10 object categories is constructed to support this research. Comprehensive experiments demonstrate superior performance over state-of-the-art inpainting methods, achieving a high structural fidelity of 85.3% Intersection over Union and a textural quality of 24.8 dB peak signal-to-noise ratio, with robust generalization. The work bridges the gap between generative models and domain-specific reasoning. Beyond visual restoration, our method is shown to correctly restore the semantic identity of occluded objects for vision-language models, significantly improve downstream object detection, and enable a layered 2.5D understanding of complex scenes, offering tangible benefits for intelligent geospatial analysis.