Research On Extraction Of Mangrove From Remote Sensing Image With U-Net Network Of Transformer

Mangroves are an important component of wetland ecosystems,playing an important role in regulating climate and protecting ecological balance.Due to the advantages of low cost and short cycle in obtaining high-resolution remote sensing images,they are widely used in mangrove extraction tasks.Remote sensing images can provide high-precision terrain information and complex details,and can also be used for large-scale monitoring of mangroves.Pixel classification and object-oriented methods have the phenomenon of unreasonable segmentation for complex background environments.Although traditional machine learning methods have certain effectiveness in mangrove extraction tasks,they can only rely on low-level features of remote sensing images for extraction and cannot meet the requirements of high-quality mangrove extraction.With the rapid development of deep learning,it provides a new research direction for the field of remote sensing image segmentation.Deep learning can extract feature information from different levels of remote sensing images for learning.This article is based on deep learning methods to automatically,quickly,and efficiently extract mangroves from remote sensing images.The main research contents are as follows:(1)Mangrove in remote sensing images is susceptible to interference from other ground object information and the scattered distribution of mangroves causes poor extraction results.In this paper,a ST-UNet network model based on Shuffle Transformer is constructed based on the symmetry of the U-Net network structure.The model uses Shuffle Transformer as the main component of the network model to extract features from remote sensing images,expand the receptive field of the network model,combine information features that are beneficial to mangroves,suppress other unrelated semantic information.Experiments show that the joint intersection union of ST-UNet network model in two datasets is 91.70 % and89.89 % respectively,which is 5.27 % and 3.95 % higher than that of U-Net network.Compared with other networks,the indicators have also been improved,which enhances the performance of the mangrove information extraction model of remote sensing images.(2)To address the issues of intra class and inter class similarity in mangrove extraction in remote sensing images,and to further improve the performance of mangrove extraction in remote sensing images,a continuous convolutional layer consisting of a hollow convolutional pooling pyramid and a small convolutional kernel is first used to segment the remote sensing image.At the same time,the pixel encoded spatial information of the mangrove is obtained,weakening the background information.Secondly,high-level and low-level feature information are fused through skip attention fusion,Accelerate the convergence of the network model and avoid gradient vanishing.Experimental results have shown that the joint intersection and union sets of the model on two datasets are 92.88% and92.78%,respectively,which are 1.18% and 2.89% higher than the ST-UNet network,thus improving the performance of the model for mangrove extraction.