Research On The Extraction Method Of River In Complex Environment Based On Satellite Remote Sensing Image

Rivers are one of the most important natural resources on earth,not only providing essential water resources for human beings,but also providing important support in transportation,economic development and ecosystem.The remote sensing image has the characteristics of wide range,strong timeliness and large amount of information,etc.River extraction with the help of remote sensing image is of great theoretical and practical significance to grasp the hydrological characteristics of a certain area,environmental protection and development construction,etc.There are still some difficulties in the river extraction technology of high-resolution satellite remote sensing images.First,the accuracy of river extraction using a single feature is low,and the description of river features is not comprehensive enough;second,the segmentation effect of stroke width transform algorithm is poor when the river edge is influenced by highway,building and other features;third,the extraction accuracy of U-Net for small rivers is not high,and it is easy to have disconnection and Thirdly,the accuracy of U-Net for extracting small rivers is not high,and the phenomenon of disconnection and incomplete structure is easy to occur.These are the keys to affect the accuracy and completeness of river extraction from remote sensing images.In response to the above problems,the main work of this paper is as follows:(1)To remedy the problem of limited applicability of single-feature river extraction,a river extraction method that incorporates texture,spectral and shape features is proposed.A comprehensive and effective feature description is obtained by considering the multiple information contained in the river.The near-infrared band input,where the grayscale features of the river differ most from other features,is selected,and each statistic is calculated based on the texture features using a grayscale coeval matrix.The angular second-order moments can reflect the homogeneous homogeneity of the river,and the obtained angular second-order moments are automatically thresholded for segmentation,which can filter out part of the noise.Then,according to the slender shape feature of the river,the geometric feature filter is constructed to filter out the non-river features,and finally the high accuracy extraction of the river is realized.Through the comparison and verification with the classical algorithm,the extraction results show that the method can effectively improve the accuracy and completeness of river basin extraction.(2)Based on the feature that rivers have similar consistent widths in a certain range of remote sensing images,the stroke width transformation algorithm has a good effect in river extraction.However,this algorithm is easily influenced by the surrounding features of rivers.In order to obtain the river area information with high accuracy,a river extraction method combining relative total variation and stroke width transform is proposed.The relative total variance model is used to smooth the image,further enhance the feature description of key information in the river region of the image,smooth the texture information of the features,and suppress the interference of non-essential information.The experimental results show that the method can effectively improve the stroke width transformation algorithm in river extraction accuracy.The method was applied to the remote sensing images of Taishan City taken by High-Scale One,and the extraction of river basins in Taishan City with higher accuracy was achieved.(3)An improved U-Net network incorporating an attention mechanism is proposed to address the problem that the U-Net network does not have high extraction accuracy for fine rivers in high-resolution remote sensing satellite images.To improve the extraction of fine water bodies,the number of layers in the network is deepened by adding four convolutional layers,an up-sampling layer,a down-sampling layer and a skip link layer,and the depth of the filter is gradually increased from 16 to 512.the Attention Gate is incorporated to adjust the weights of the passed features when the skip link is performed to suppress irrelevant information in the image and highlight the important features in the target region The network learns the fine river features.The experimental results show that the method can not only effectively improve the accuracy and completeness of fine river extraction,but also shorten the training time to 1/3 of the original one.