Data-Driven Approach for Dehazing of High-Resolution Multispectral Remote Sensing Images

Abstract

Haze is caused due to presence of dust, light vapors, or smoke, causing a lack of transparency in the air. This creates a significant issue for satellite images as the image regions affected by haze suffer a lack of contrast and definition, resulting in difficulty interpreting the scene. Traditionally, this issue was solved by using atmospheric correction methods, a tedious process requiring estimating several geophysical quantities at once to give reliable results. A set of algorithms to recover the clarity in hazed images, called dehazing algorithms, are becoming popular in practice for their simplicity and efficacy. This paper introduces a Convolution Neural Network based solution in which using a compound loss function to prioritize the clarity and similarity to the original has improved performance to solve the dehazing problem for high-resolution multispectral images.