Research And Application Of Ground Object Information Extraction In Visible Light Remote Sensing Image

Extracting information about roads and water bodies from visible light remote sensing images is of great significance for urban basic data updating,water resource monitoring,and intelligent transportation construction.However,the application software is not mature at low cost,which is based on deep learning model to extract road and water information from visible remote sensing image quickly and accurately,and then makes vectorized water body and road maps by post-processing algorithms.To this end,this thesis studies the methods of extracting urban road information and water body information,and describes developing the processing software to provide data for digital city computing.The main work is as follows:(1)Based on U-Net-like DNN architecture,this thesis realizes intelligent recognition of water contours in visible light remote sensing images.In design of water segmentation model,considering the importance of spatial position information to recognition results,the spatial-channel attention mechanism is introduced to enhance the feature expression of spatial position information,and the feature maps of each stage of encoder is fully utilized by the hyper-column module.This thesis studies how to design appropriate loss function to overcome the imbalance of the data distribution,and improves the results based on the conditional random field(CRF).Limited by the size of remote sensing image,there are a large number of contour fragments which are cut off by the boundary of image.And therefore,an effective algorithm is proposed for large-scale automatic contour splicing.(2)Compared with the images containing water,the size of the images containing roads is larger.On the basis of the segmentation model of water body,the recognition performance of road segmentation model is effectively improved by introducing Cascade Dilated Convolution Blcok(CDCB).The algorithms in the post-processing stage are studied and the related modules are developed.The algorithms ensure the practical application value of the system with the help of the filling of holes in the recognition results,the elimination of isolated pixel clusters,the extraction of road skeletons,the vectorization of skeletons,and the removal of residual roads and the large-scale broken roads.(3)In practical applications,due to the large size of the remote sensing image,the amount of calculation and parameters of the segmentation model to process the images is large.In order to reduce resource consumption,a compression acceleration method for the segmentation model architecture is studied,and an improved deep-separation convolution module is proposed.This method guarantees that,under the condition that the recognition performance is almost not reduced,the amount of calculation is effectively reduced by more than 4 times,and the amount of parameter is more than 20 times,which is beneficial to the deployment and application of the segmentation model in the actual environment.Finally,the software is designed and implemented for extracting water and road information from visible light remote sensing image,and it is fast and accurate for water and road recognition.The validity of the above work is verified by using the visible light remote sensing images of a southern urban area in China.