RRCGAN: Unsupervised Compression of Radiometric Resolution of Remote Sensing Images Using Contrastive Learning

The majority of current remote sensing images possess high-radiometric resolution exceeding 10 bits. Precisely compressing this radiometric resolution to 8 bits is crucial for visualization and subsequent deep learning tasks. Previously, radiometric resolution compression required extensive parameter adjustments of traditional tone-mapping operators. Deep learning is gradually replacing this high manual dependency method. However, existing deep learning tone-mapping techniques are primarily designed for natural scene images captured by digital cameras, making direct application to remote sensing images challenging. This limitation stems from disparities in data formats and the complexity of semantic representation in remote sensing images. Moreover, the block prediction inherent in deep learning models often results in tiling artifacts postsplicing, failing to satisfy the scale dependency of remote sensing images. To tackle these challenges, we propose leveraging contrastive learning methods to compress the radiometric resolution of remote sensing images. Given the rich detail information and complex spatial distribution of objects in remote sensing images, we develop a CNN-Transformer hybrid generator capable of capturing both local details and long-range dependencies. Building upon this, we introduce nonlocal self-similarity contrastive loss and histogram similarity loss to enhance feature expression and regulate image color distribution. Additionally, we present a postprocessing technique based on hybrid histogram matching (HHM) to enhance image quality and seamlessly generate whole-scene images. Through experiments and comparisons on our dataset, our method demonstrates superior performance. The dataset and code can be obtained online at https://github.com/ZzzTD/RRCGAN.

I. Introduction

At present, 8-bit remote sensing images are commonly used for visualization [1] and various downstream tasks such as semantic segmentation [2], target detection [3], algorithm acceleration [4], and change monitoring [5]. However, with the advancement in sensor and storage technologies, most remote sensing images available today are high-radiometric-resolution remote sensing images (HRRs) with a bit depth exceeding 10 bits [6]. Consequently, there is a need to convert these images into low-radiometric-resolution remote sensing images (LRRs). Radiometric resolution compression is a crucial preprocessing step in interpreting remote sensing images, and traditional methods such as linear stretching [7], percentage truncation [8], and exponential transformation [9] are often used. While these methods offer simplicity and efficiency, they require manual parameter adjustments based on specific scenarios, leading to potential issues like low contrast and loss of details [10]. This can hinder meeting the real-time and accuracy requirements of remote sensing image processing. Therefore, there is an urgent need to explore more efficient radiometric resolution compression technologies to ensure timely and accurate remote sensing image processing [11].