Dynamic range compression method for high radiometric resolution remote sensing images using contrastive learning
Tengda Zhang, Jiguang Dai, Qian Hu · Expert Systems with Applications · 2025
Dynamic range compression of high radiometric resolution remote sensing images involves compressing the grayscale value range to 0∼255. This process is a necessary preprocessing step for the visualization, storage, and analysis of remote sensing images. This task falls under tone mapping, but due to data heterogeneity and variations in semantic representation, existing tone mapping methods often result in color distortion and visual artifacts when applied to remote sensing images. We propose an unsupervised dynamic range compression method that leverages the characteristics of remote sensing images and tasks. In the construction of the generator, we drew an analogy between the pixel values during the compression process and the particle motion within a closed thermal field. Utilizing the thermodynamic difference equation, we developed a third-order finite difference residual module to explicitly guide the model’s feature extraction. Considering the limitations of the existing contrastive loss in terms of attention range, we propose a multi-granularity contrastive loss that operates at both the patch and semantic levels. Additionally, based on the similarity of histogram shapes before and after dynamic range compression, we introduce a histogram shape context similarity loss to regulate the image color distribution. Due to the limited number of existing studies, we have constructed a dataset and conducted extensive experimental verifications. The results indicate that the proposed method yields superior outcomes and is applicable to downstream tasks. The relevant code and dataset can be accessed via the following link: https://github.com/ZzzTD/RS_DRC .