Task-Guided Prompting for Unified Remote Sensing Image Restoration
A single model now handles denoising, cloud removal, and SAR despeckling across RGB, multispectral, and thermal data.
A research team has introduced TGPNet, a groundbreaking unified framework designed to overcome a major bottleneck in remote sensing. Traditional AI models for satellite and aerial imagery are typically built to handle only one type of degradation—like just removing clouds or just reducing noise—and are often limited to a single data type (e.g., only RGB). TGPNet shatters this paradigm by integrating five critical restoration tasks into a single, adaptable model. Its core innovation is a Task-Guided Prompting (TGP) mechanism, which uses learnable, task-specific embeddings to generate degradation-aware cues. These cues then hierarchically modulate features throughout the network's decoder, allowing it to precisely tailor its processing for each specific problem while maintaining a single, shared set of weights.
To validate their approach, the team constructed a comprehensive benchmark covering diverse sensor modalities, including RGB, multispectral, Synthetic Aperture Radar (SAR), and thermal infrared imagery. Experimental results show TGPNet achieves state-of-the-art performance not only in unified multi-task scenarios but also on challenging, unseen composite degradations. Remarkably, it even surpasses specialized models built for individual tasks, such as cloud removal. By successfully unifying the restoration of heterogeneous degradations, this work represents a significant leap toward practical, scalable AI for operational remote sensing pipelines, where data often arrives corrupted by multiple, overlapping issues. The code and benchmark are slated for public release.
- Unifies five distinct restoration tasks (denoising, cloud/shadow removal, deblurring, SAR despeckling) in one model.
- Uses a novel Task-Guided Prompting (TGP) strategy with learnable embeddings to adapt a single network to different degradations.
- Outperforms specialized models on individual tasks and handles unseen composite degradations across RGB, multispectral, SAR, and thermal data.
Why It Matters
Enables more robust, efficient analysis of satellite/aerial imagery for climate monitoring, agriculture, and defense by cleaning complex, real-world data.