Research On Coastline Extraction And Land Classification Based On Deep Learning Model

The Guangdong-Hong Kong-Macao Greater Bay Area is one of the regions with the strongest economic vitality in China.Due to its rapid development,ecological pr oblems such as wetland destruction and coastal erosion are becoming more and more serious.The automatic extraction of coastline and the classification of coastal land h ave important practical significance for the construction of ecological civilization in t he Guangdong-Hong Kong-Macao Greater Bay Area.With the rapid development of deep neural network technology,the precision of edge detection algorithm based on d eep neural network has been greatly improved compared with the traditional edge det ection algorithm,and the performance of land and ocean boundary extraction based o n it has also been improved.Combined with the meteorological characteristics of the research area,a scheme was developed to extract the coastline using Sentinel-1A remote sensing satellite imag e data.Based on the richer convolutional features,a kind of water edge extraction al gorithm is proposed,and different strategies are adopted for different coast types to c omplete the coastline extraction.Meanwhile,the combination of morphological image processing technology and richer convolutional features can be better applied to Sent inel-1A images.Then,compared with the traditional method,the new water edge extr action method can get more continuous and more accurate water edge.The control p oints are used to measure the accuracy.Compared with the traditional Canny edge de tection algorithm with the best performance,this method improves the detection accur acy by 11.45%.In addition,six classification networks based on PSPNet and Link Net are studied and implemented to complete the classification of coastal land.The classification me thod is based on Terra SAR-X images.Based on the classification results and confusi on matrix,it is found that the PSPNet classification network of backbone VGG16 ca n achieve the optimal classification effect,and its PA,MPA,MIo U and Kappa coefficients are 87.15%,81.21%,66.36% and 81.56%,respectively,which are generally bet ter than the other five networks.In terms of ground objects category,the classificatio n accuracy of water body is the highest,and the accuracy of the six networks can r each more than 96%.Forest land and build-up area also have good classification acc uracy,and in different networks,the performance is relatively consistent.At the same time,due to the low sample proportion of bare soil and road,the classification erro r is slightly larger,and there is inconsistency in each network.