Relative Radiometric Correction via Virtual Low-Resolution Image Reconstructing With Spatio-Temporal Feature Fusion for Remote Sensing Images

Relative radiometric correction (RRC) of multiple remote sensing images is crucial for large-scale applications like orthophoto production, panoramic image generation, and 3D reconstruction. However, many existing RRC methods inadequately utilize detailed image information, often leading to local detail loss and degraded quality. To address this limitation, a novel RRC method based on virtual low-resolution image (VLRI) reconstruction with spatio-temporal feature fusion is proposed. The method generates high-quality correction results by integrating high-resolution detail features from the original image with the spectral features of a low-resolution reference image. The key innovation is a strategy that reconstructs the VLRI using spectral information from the original image as auxiliary data. This facilitates the fusion of spatio-temporal information across images of varying resolutions. Additionally, we explicitly quantify the temporal correlation, spatial proximity, and spectral similarity between the reconstructed image, original image, and reference image, incorporating these factors into the correction calculation. Experimental results demonstrate that the proposed method effectively mitigates radiation differences across multiple remote sensing images, minimizes detail loss during correction, and outperforms other competitive methods. Importantly, our approach robustly leverages reference information even with insufficient geo-registration, thereby enhancing the utility of historical low-resolution data.