Research On Remote Sensing Image Dehazing And Application

At present,remote sensing images have been widely used in military,agriculture,forestry and other fields.The fuzzy senses of haze in remote sensing images reduce their quality and bring severe challenges to object classification and target detection.The traditional dehazing algorithms can not effectively meet the needs of the actual production work.In this thesis,images of the operational land imager(OLI)of Landsat-8 satellite were taken as the experimental objects,and the atmospheric scattering model commonly used in the field of dehazing was taken as the theoretical basis.According to the characteristics of haze's distribution in remote sensing images,the parameters of atmospheric scattering model were converted,and the convolution neural network structure of remote sensing image was designed.The main research contents and results are as follows:(1)Considering the lack of training data in remote sensing image dehazing,based on the dark channel prior method,the haze masks were extracted from natural remote sensing images.And the haze masks were added to clear remote sensing images based on the atmospheric scattering model to create synthetic haze images who are similar to the real haze images.The true values of relevant parameters were saved at the same time.The training data set of remote sensing image dehazing with true values was obtained.(2)We studied the photo image dehazing based on deep learning,and built physics driven dehaze network(PDDN)which balanced the dehazing performance and the parameters quantity.The PDDN network achieved the highest peak signal to noise ratio(PSNR)and structure similarity(SSIM)accuracy of 31.11 d B and 0.96 respectively in the test data set with the same source as the training set.It also achieved the highest PSNR accuracy of 22.27 d B and SSIM accuracy of 0.83 in the test data set with a different source of the training set.Our remote sensing imagees dehazing network's structure was based on PDDN.(3)According to the characteristics of remote sensing image,such as large coverage,the features related to the distribution of haze were fused into one feature,and the PDDN network was modified to obtain a single feature dehaze network(SFDN).The experimental results showed that SFDN achieved the best PSNR and SSIM accuracy in the test set and verification set of synthetic data,the test set was 36.54 d B and 0.98,and the verification set was 27.69 d B and 0.87.After dehazed by our model,three real haze images' correlations between their related clear images had been improved,except their near-infrared bands.The land-use classification accuracy of these real haze images had been improved after doing dehazing with our model.The overall accuracy(OA)had been improved by 0.68%,0.72%,1.58% respectively,and the Kappa coefficient(KC)had been improved by 0.03,0.05,0.05 respectively.(4)To improve the practicability of the dehazing models,the dehazing toolbox with a graphical interface was created.This toolbox includes the functions of setting the read-write path,selecting the model type and image display.