CNN-based Image Color Consistency for Multi-source Satellite Images: A Case Study in China’s Water Diversion Project Application

Color consistency processing is important for optical satellite image mosaicking to obtain necessary geographic information, then to serve land and water resource monitoring and many other public fields. In recent years, convolutional neural networks (CNN) have been successfully used in image processing applications such as image classification, change detection, and fusion. However, there are a few studies on image color consistency, especially based on Chinese high-resolution satellite images. In this study, we proposed a new CNN-based image color consistency method that used high-resolution Google images as a reference. Combined with a CNN-based method of image fusion, we produced large-area color-consistent mosaics with high resolution based on multi-source satellite images. The loss functions and network parameters were optimized based on our dataset. We successfully applied the proposed method in one of China's mega water diversion projects ‘Yangtze-Huai River Diversion Project’ and found that it had better performance when compared with the existing methods. The proposed method does not only provide a basic tool in image mosaicking for remote sensing applications but also acts as a precedent for subsequent automated image change detection and object recognition based on deep learning.