Research On The Classification Of Objects In The Three Gorges Reservoir Area Based On Landsat And GAN Super-resolutio

Land use and land cover(LULC)are closely related to socio-economic development,ecological and environmental changes,surface material cycle and energy balance.As an important part of the Yangtze River basin,it is important to monitor and study the LULC changes in the Three Gorges reservoir area.With the development of sensor technology,the resolution of satellite images is getting higher and higher,and the resolution requirement of LULC data is rising,but the time span of new remote sensing satellite high-resolution images is short,and the spatial resolution of early remote sensing satellites is low,so it is impossible to obtain long time series high-resolution feature classification.Super Resolution(SR)based on Generative Adversarial Network(GAN)can effectively improve the spatial resolution of images without supervision.In this thesis,we use Landsat satellite data from 1985-2021 to obtain 10 m resolution images based on super-resolution technique,and use the 2021 World Cover data released by ESA as the sample training to obtain long-time 10 m resolution feature coverage data of the Three Gorges reservoir area,which is of great significance to study the changes of natural resources and human environment in the area.The main works and results are summarized as follows.(1)To address the problem that GAN is difficult to learn the mapping between high-and low-resolution images in SR tasks,Deep Back-Projection Networks are combined to improve the generator part,and Deep Back-Projection Networks-based Remote Sensing Super-Resolution Generative Adversarial Network(D-RS-SRGAN)is designed,which learns the mapping between high and low resolutions without supervision to better realize the unsupervised super-resolution of remote sensing images.The PSNR value is 34.01 d B and SSIM value is 0.949 on RGB band of Sentinel-2 satellite,and the PSNR is 34.28 d B/33.99 d B and SSIM is 0.956/0.951 on bands 1,2,5 and 7 of Landsat8 and bands 3,4 and 5 of Landsat5,respectively,which are better than the current several commonly used unsupervised SR algorithms.(2)Given the effectiveness of D-RS-SRGAN for SR tasks,it was used to super-resolve Landsat images to obtain 10 m resolution images.The World Cover 2021 data were trained using the Swin-Unet model,and the SR results were fed into the trained model to obtain 10 m resolution feature classification data.The results showed that the classification accuracy of Landsat8 super-resolution images was 83.7%,and the accuracy of Landsat5 super-resolution images was above 70%,indicating that the super-resolution results in the segmentation model output is effective.(3)The Landsat and D-RS-SRGAN methods were used to obtain 10-m resolution feature classification datasets of the Three Gorges reservoir area at each five-year interval from 1985 to 2020 and the changes of feature cover in the Three Gorges reservoir area are analyzed.The effect of returning farmland to forest in the reservoir area is obvious.The area of water bodies increases significantly,the area of grassland and bare land decreases rapidly,urban expansion is obvious,and the floor area rises nearly four times.