Research On Remote Sensing Image Water Body Extraction And Change Detection Model Based On Deep Learning

With the convenience of acquiring remote sensing data,especially high-resolution data,quickly obtaining high-precision water information from large data volumes and detect changes in water bodies is very important for water monitoring,disaster protection and agricultural production.In order to extract water bodies more quickly and efficiently and improve the accuracy of water body change detection,this article applies semantic segmentation network and time series model to water element extraction and change detection of remote sensing image time series.The semantic segmentation network and the LSTM(Long Short Time Memory)model have carried out a series of studies on the problems involved in water extraction and change detection of remote sensing image time series.The main research contents and innovations of the paper are as follows:(1)U-net semantic segmentation model for multi-channel water extraction is constructed and a neural network training method is designed based on the combination and optimization of SGD and Adelta optimizer,and the input layer and upsampling layer structure in the U-net network is improved.Experimental results show that the proposed U-net has good water extraction performance,and the classification accuracy is generally higher than that of the support vector machine and the standard U-net.At the same time,the proposed U-net training time is significantly lower than the standard U-net.(2)By combining the characteristics of CNN to extract spatial features and LSTM to extract temporal features,CNN＿LSTM and Convolutional Seq2 Seq time series water change detection models are proposed to detect changes in water bodies on Landsat remote sensing images in the study area.Due to the large amount of time series remote sensing image data,the model is difficult to run smoothly with limited GPU resources.Therefore,the down-sampling and resolution recovery module is proposed to reduce the computing resource occupation without reducing the performance of the model.Experiments show that the proposed method can effectively improve the accuracy of time series water change monitoring and is better than methods such as LSTM.(3)In order to solve the problem of low performance of the semantic segmentation algorithm for change detection,we introduce the mask method into semantic segmentation,and a Mask local convolution method is proposed to improve the performance of semantic segmentation networks Mask Local U-net and Local Res U-net.Local Res U-net are smaller than Mask Local U-net networks.Experiments on three change detection datasets demonstrate higher semantic change detection performance for Mask Local U-net and Local Res U-net than DASNet,DTCDSCN,U-net and others.Combined with the water distribution,water change,local material elements change,meteorological data and other related analysis and sliding average analysis,the area change trend and causes of each water area were obtained.