Research On The Extraction Of County Water Bodies From GF-2 Remote Sensing Images Based On U-net

In order to protect water resources,establish the distribution and real-time monitoring of water resources,in order to better management and governance,it has become an effective means to prevent the river from drying up,cutting off and water pollution.With the development of remote sensing technology,the consumption of human and material resources in establishing the distribution and real-time monitoring of water resources is greatly reduced.Although the traditional recognition methods perform well in some specific scenes,with the increasing amount of data,people have higher requirements for universality and efficiency.However,the traditional recognition methods have the disadvantages of low accuracy,low efficiency and unable to extract deep features from remote sensing images.Deep learning can effectively solve these problems.Through comparison,it is found that u-net(semantic image segmentation network structure)is an improvement of FCN(fully volatile neural network).It has an end-to-end operation architecture,which can accurately classify complex distributed images,and can obtain high-precision results in the case of limited data sets.In this paper,the u-net semantic segmentation model based on fully connected convolutional neural network is used to extract the county water from gf-2 high-resolution remote sensing image,and compared with the semantic segmentation segnet model,because both of them are not good.At last,the convolution depth is reduced and the depth of the convolution model is improved.The results show that: the improved u-net makes full use of the rich spatial information of remote sensing images,and achieves a good segmentation effect.Compared with segnet model and original u-net,the accuracy,F1 value and overall accuracy of the improved u-net are improved by about10%.