Principal color extraction and decomposition for enhanced remote sensing image homogenization

Abstract Remote sensing image color correction faces challenges due to diverse acquisition conditions and sensor variations. Existing methods often require one-to-one training and large amounts of domain-specific annotated data, limiting scalability and generalization. To tackle these issues, we propose PCED-Match , a novel color layer decomposition and alignment framework. It decomposes RGB images into homogeneous color layers and performs layer-wise color style alignment under physical constraints. Trained once on a general RGB dataset, PCED-Match generalizes well to diverse remote sensing imagery without domain-specific retraining or large-scale annotations. Extensive experiments on the WHU Building and Inria Aerial Image Labeling datasets demonstrate the effectiveness of our method. Integrated into mainstream segmentation models (U-Net++, DeepLabv3, SegFormer), PCED-Match consistently outperforms existing approaches in mIoU, F1-score, and Pixel Accuracy, achieving superior color consistency and semantic preservation. Our method source code and related data collection of links are available on https://github.com/freesui1984/PCED-Match