Relative Radiation Correction Based on CycleGAN for Visual Perception Improvement in High-Resolution Remote Sensing Images

Abstract

The differences between the imaging environments of sensors lead to great differences in remote sensing images of the same area in different seasons. Relative radiation correction has high practical value as the main method to reduce such differences. However, the differences in vegetation radiation caused by seasonal changes are difficult to correct by traditional radiation correction methods. The corrected results also have difficulty achieving better results at the level of human eye visual perception. Moreover, the traditional measurement of the relative radiation correction result image quality index is not consistent with the human eye visual perception effect. To address the above two problems, this paper performs seasonal relative radiation correction on high-resolution remote sensing images by CycleGAN based on a convolutional neural network, including two transformations: 1) the transformation of remote sensing images from autumn-winter to spring-summer and 2) the transformation of remote sensing images from spring-summer to autumn-winter. The similarity between the relative radiation-corrected image and the reference image is measured by the convolutional neural network model with the ability to discriminate distances. The results show that the visual effect of this method is significantly better than that of other relative radiation correction methods, and the visual perception distance is consistent with the human eye visual perception judgment. The changed area still retains its original feature characteristics. The visual perception distance of the conversion from autumn-winter to spring-summer images is improved by 9% compared with other state-of-the-art methods. The visual perception distance of spring-summer images to autumn-winter images is improved by 3%. We expect that the method in this paper can be used for preprocessing to improve the accuracy of algorithms for remote sensing image classification, image change detection, etc.

Learning Color Distributions from Bitemporal Remote Sensing Images to Update Existing Building Footprints

“…Image color translation involves the translation of the color style of the source domain to the color style of the target domain. Early color translation methods included linear and nonlinear methods [32]. The most commonly used nonlinear method is histogram matching (HM) [33] and linear methods include the image regression (IR) method [34], Reinhard method [35], and pseudo-invariant feature (PIF) method [36].…”