A Study On Single Remote Sensing Multispectral Image Dehazing

Remote sensing imaging is very vulnerable to cloudy or hazy weather,which makes the captured remote sensing multispectral images significantly degraded and cannot be used effectively.Therefore,it is necessary to study algorithms for efficient dehazing on single remote sensing multispectral images to improve image quality.With the advent of deep learning,researchers have tried to apply artificial neural network to remote sensing multispectral image dehazing task and have made remarkable progress.However,the researches on deep learningbased remote sensing multispectral image dehazing still face two challenges:(1)The mainstream deep learning-based remote sensing multispectral image dehazing methods are based on a supervised manner,which requires a large amount of labeled data to train the models.However,there is a lack of remote sensing multispectral image dehazing datasets with sufficient amount of data and containing pairs of hazy and haze-free images to promote research on supervised learning-based remote sensing multispectral image dehazing methods.(2)With the continuous development of deep learning,neural network architectures with better performance(e.g.,Transformer)have been emerging.However,few researchers have applied these state-of-the-art neural network architectures to remote sensing multispectral image dehazing tasks.In response to the above challenges,the following work is done in this thesis.(1)A new method for synthesizing hazy remote sensing multispectral images from haze-free remote sensing multispectral images is proposed.The method considers the non-uniformity of the spatial distribution of haze in remote sensing multispectral images,the variability and correlation of the effects on different channels,and simulates the situation where the information of the ground scene is completely attenuated under thick haze.In addition,the method is very stable in the estimation of multispectral atmospheric light,which is not easily affected by the scenery in the images.(2)A large-scale synthetic remote sensing multispectral image dehazing dataset RSHaze is proposed and open-sourced.The dataset contains 54,000 pairs of hazy and haze-free remote sensing multispectral images,which are classified into three classes according to the haze concentration: light,moderate and dense.RS-Haze is the largest open-sourced remote sensing multispectral image dehazing dataset in terms of data scale so far.(3)A remote sensing multispectral image dehazing model RSDehaze Former based on the multiscale transformer architecture is proposed.The model combines spatial self-attention and channel self-attention and has powerful spatial correlation and channel correlation characterization capability.It can model the non-uniformity and the correlation between different channels of haze in remote sensing multispectral images to improve dehazing performance.(4)Adequate experiments are conducted on RS-Haze and RSDehaze Former.Extensive experimental results on the RS-Haze dataset show that RSDehaze Former outperforms other advanced image dehazing methods in terms of quantitative and qualitative evaluation,and has an advantage in terms of the number of parameters and computational consumption.Meanwhile,the experimental results on real hazy remote sensing multispectral images show that the models trained on the RS-Haze dataset have certain generalization ability and can solve the dehazing problem on real hazy remote sensing multispectral images to a certain extent.