Research On Water Extraction Method And Change Monitoring In Jilin Province Based On Deep Learning

The spatial-temporal distribution of water is very important for the application and monitoring of water resources.With the rapid development of satellite remote sensing technology in China,it has shown strong application prospects in water extraction and change monitoring research due to its low cost,high efficiency,and strong operability,and has become an important data source.This paper chooses Landsat remote sensing image as the experimental data source and takes Jingyuan and Chagan Lake in Jilin Province as typical study areas to carry out water extraction methods and change monitoring research.During the actual work,the main research results include:(1)Water extraction method in Jilin Province based on the spectral index and deep learning network.This paper,based on Landsat single-phase remote sensing image,uses the spectral index-based water extraction method and deep learning network-based water extraction method to study the water extraction method in Jilin Province.The results show that compared with the spectral index-based water extraction method,the deep learning network-based water extraction method can extract the water body in study areas more effectively.Among them,the Deeplabv3+network water extraction method has the best results,and its overall accuracy and Kappa coefficient are above 93% and 0.86,respectively.(2)Water extraction method in Jilin Province based on improved deep learning network.In this paper,the convolution neural network(CNN),U-Net network,and Deeplabv3+ network water body extraction method based on deep learning network are improved respectively.First,a water extraction method based on the convolution neural network(CNN)is established,which combines multilayer perception(MLP)and is based on a joint probability model with spectral information.Secondly,the F-DIo U loss function is constructed,the loss function in the U-Net network is replaced by the F-DIo U loss function,and a water extraction method based on an improved U-Net network is established.Finally,based on the Deeplabv3+ network,a water extraction method based on the improved Deeplabv3+ network is established by optimizing the backbone network and ASPP structure.The experimental results show that the improved deep learning network method proposed in this paper can not only improve the water body extraction accuracy but also can be applied to long-time series water body monitoring.Among them,the results based on the improved Deeplabv3+ network are the best,with overall accuracy and Kappa coefficient of 95%and 0.9 above,respectively.(3)Water extraction method based on improved deep learning network and fused information.In this paper,a water extraction method based on an improved Deeplabv3+ network with multi-feature and multi-temporal fusion information is proposed by fully mining water features in remote sensing images through HIS transformation,Tasseled Cap transformation(K-T)and principal component analysis(PCA),and combining Landsat summer and autumn multi-temporal remote sensing images.The results show that the accuracy of water extraction can be significantly improved by combining multi-feature and multi-temporal remote sensing images with the improved deep learning network.The overall accuracy and Kappa coefficients reach 97% and 0.94,respectively.(4)Spatial-temporal monitoring and change analysis of water bodies in Jilin Province based on the GEE platform.This paper applies the water body extraction method based on the improved Deeplabv3+ network with multi-feature and multi-temporal fusion information to the GEE platform,obtains the water body extraction results in Jilin Province from 2000 to 2022,and makes a correlation analysis on the spatial and temporal changes of water body in Jilin Province.