Nighttime light remote sensing image haze removal based on a deep learning model

Haze contamination is a quite common issue in nighttime light remote sensing (NTLRS) images. It significantly limits the application of NTLRS images, especially in human activity monitoring and socio-economic studies. Furthermore, NTLRS images usually struggle with noise. Although many remote sensing image haze removal methods have been developed, to the best of our knowledge, very few studies have been conducted on haze removal of NTLRS images. In this study, to address haze in NTLRS images, particularly the challenging issue of concomitant noise contamination, we developed a nighttime light haze removal network (NTLHR-Net). Specifically, to capture effective spatial structural information (dominated by sparse or spot-like shapes) and eliminate joint haze and noise contamination, an encoder-decoder structure coupled with a mixture attention block was developed. Moreover, a multiscale convolutional block was employed iteratively in the middle of the encoder-decoder structure to distill the spatial structural information in high-dimensional spaces. In the experiments, the NTLHR-Net method was compared with seven state-of-the-art haze removal methods for both simulated and real hazy NTLRS images with different spatial structures. The results demonstrate the feasibility of the proposed NTLHR-Net method in cases with various haze and noise contamination. This study provides a new solution for increasing the quality of the observed NTLRS images for downstream applications.