A GAN-based Relative Radiometric Correction Model of Remote Sensing Data
Linglin Xie, Jianhao Miao, Xinghua Li, Xuechen Bai, Kaijun Yang · 2023
While there are many traditional methods for radiometric correction of multitemporal remote sensing images, deep learning methods are quite rare. Recently it has witnessed a rapid development of computer vision, many style transfer and domain transfer methods have given us great inspiration. However, traditional methods have problem in achieving uniform effect of radiometric correction, while style transfer methods struggle to realize control of local areas and transfer degree, especially when dealing with complicated remote sensing data. Thus, a generative adversarial network (GAN)-based relative radiometric correction method combined with deep style transfer (NormGAN) is proposed. A cycle-consistent structure is applied for bi-directional domain transfer of non-corresponding regions. VGG19 network is used to obtain feature map for the calculation of content loss and style loss. Meanwhile, rough masks of landcover enable transfer within each class and weight matrix is put forward to control transfer degree. Our results indicate that the proposed NormGAN is capable and effective, which has much superiority over other methods.