ColorSate: A Color Correction Network for Satellite Imagery with Channel-Specific Distortion Decoder

Color distortion in satellite imagery is a sensor-related hardware issue that manifests as color artifacts within the captured scenes. Existing color correction methods have been primarily developed for consumer-grade digital cameras, whose single-sensor arrays inherently produce interchannel correlations. In contrast, satellite imaging systems employ independent sensor arrays for each spectral band, leading to asymmetric and channel-specific distortion patterns. Therefore, specialized algorithms are required to address these unique distortions, which differ fundamentally from those observed in consumer digital cameras. To address these limitations, we propose ColorSate, a novel deep learning-based algorithm that captures unique characteristics of satellite color distortions by estimating channel-specific patterns using a dedicated distortion decoder. The core of this distortion decoder is a dual-stage correction block, which consists of a channel-decoupled distortion extractor (CDE) and an interchannel refinement module (IRM). The CDE separately extracts channel-specific distortion patterns using depth-wise convolutions, while the IRM subsequently refines these patterns using interchannel attention. In addition, we introduce a simulation process to generate paired datasets, addressing the lack of clean ground-truth satellite images. Extensive evaluations on both simulated and real-world datasets from the Korea Aerospace Research Institute (KARI) demonstrate the effectiveness and robustness of ColorSate in restoring color fidelity for satellite imagery.