Regularized Deep Convolutional Encoding-Decoding Generative Network for Remote Sensing Cloud Image Compression
Yu Wang, Jing Yang, Sijia Wang, Xingli Yang · 2025
Typically, meteorological cloud observations are often presented in the form of high-quality and high-resolution images, such as satellite remote sensing cloud images and ground-based remote sensing cloud images. However, faced with such a large number of high-resolution images, even for the existing GPU servers and parallel computing with high-speed data processing capabilities, the computational cost of image recognition and analysis is very large, or even unacceptable. A widely used method is to resize the original image to a version with a decreased resolution and to obtain cloud information from this decreased image. Clearly, image compression is crucial. Universal interpolation based image compression methods compress the image size by directly clipping unimportant pixels at the pixel level, which may fail to capture important image feature information such as image structure, resulting in low image recognition performance. In particular, the popular deep encoding-decoding generative network can perfectly reconstruct the image almost identical to the original image by self-supervised learning the useful high-resolution image feature representation. Thus, in this paper, a remote sensing cloud image compression method based on regularized deep convolutional encoding-decoding generation network is proposed by constraining encoding-decoding network and the interpolation method. The proposed method extracts useful feature representations from the original image through an encoding process with a convolutional neural network model, and then reconstructs the compressed image at a specified compression ratio through a decoding process with a deconvolutional neural network model. In the proposed method, a Kullback-Leibler (KL) divergence constraint is introduced to align the feature representation distribution with a standard normal distribution, enhancing the accuracy of the feature representation. Additionally, the regularization constraint between the feature maps of the convolutional and deconvolutional layers is applied to minimize reconstruction error. The proposed method can effectively capture both global and local image feature with different resolutions while preserving pixel level image information. Experimental results on three remote sensing cloud image datasets showed that the proposed method is superior to the existing image compression methods in SS (Sementic Score) and FRR (Feature Remain Ratio) performance indicators.