Sentinel-2 Image Dehazing using Visible and Infrared Band Correlation Based on NDVI Classification

Due to the penetration of the electromagnetic (EM) wave in infrared are higher than the visible bands. The satellite image details from these bands are obviously than others. Most of the visible bands directly varied with the infrared bands, especially the non-water area. We can employ these properties to generate synthetic visible bands for decreasing the haze effect. However, some area has a more complex relation between visible and infrared bands. To overcome this problem, we proposed the dehazing technique for Sentinel-2 imagery by using visible and infrared band correlation based on the Normalized Difference Vegetation Index (NDVI) classification. The synthetic visible bands are constructed from the linear combination of the infrared bands. The multiple linear regression is applied to determine the linear coefficients of each formula. The NDVI is employed to classify the group of considered samples. The haze-free images with the nearest sensing date were employed to compare probably realistic images, both visual and metric comparisons. The results show that our proposed techniques can be well applied to reduce the haze effects, especially the uniform thin haze. Moreover, there give both visual and metrics results superior to the traditional method. In the case of thick haze, our proposed methods give more obvious vision. However, our methods have a disadvantage in the water areas. It had more artifact results due to its NDVI failure in the corrected classes.