| Remote sensing technology plays an important role in agricultural monitoring,geological exploration,disaster prevention and mitigation,and military investigation.With the development of satellite technology,a large amount of high-and mediumresolution remote sensing image data can be obtained,providing important data support for ground cover classification research.At the same time,with the development of deep learning,algorithm support is provided for the rapid and accurate implementation of ground cover classification.How to combine deep learning methods to improve the accuracy of land cover classification is an important issue in remote sensing classification in recent years.In this study,the central part of Guangxi Zhuang Autonomous Region is selected as the research area,and multi-temporal LANDSAT 8remote sensing image is used as the data source to propose two improved multi-spectral remote sensing image classification methods based on the improved U-Net algorithm.The specific content is as follows.(1)The former is an improved U-Net remote sensing classification algorithm based on multi-feature fusion perception.Firstly,a new improved U-Net network framework based on multi-feature fusion perception is proposed.This framework includes a channel attention module(CAM-UNet)that is added to the original U-Net framework,and cascades shallow features with deep semantic features.Additionally,a support vector machine(SVM)replaces the original classification layer of the U-Net network,and the majority voting game theory algorithm is used to fuse the results of multiple feature classifications to obtain the final classification result.Secondly,the proposed method is used to classify the 2010,2015,and 2019 remote sensing images of the study area.Finally,dynamic changes in the three-phase images are monitored.The experimental results show that the improved algorithm can improve the classification accuracy,with the overall segmentation accuracy increasing from 90.50%before the improvement to 92.82%,and the segmentation accuracy of forest land increasing from 95.66% before the improvement to 97.16%.(2)In response to the problem of long training time in the former algorithm,an improved U-Net remote sensing classification algorithm incorporating attention and multi-scale features is proposed.Firstly,a new improved U-Net network framework called SA-UNet is proposed,which combines the atrous spatial pyramid pooling(ASPP)module with the convolutional units of the original U-Net encoder in a residual form.The ASPP module can enlarge the receptive field,fuse multi-scale features,and enhance the expression ability of shallow features.By fusing the residual module,shallow features and deep semantic features are deeply fused,further utilizing the characteristics of both.Meanwhile,the spatial attention mechanism is utilized to combine spatial information with semantic information,enabling the decoder to recover more spatial information.Secondly,the proposed method is used to classify the 2015,2017,2019,and 2021 remote sensing images of the study area,and dynamic changes of the four-period images are monitored.The experimental results show that the improved algorithm can enhance the classification accuracy,with the overall segmentation accuracy increasing from 92.66% to 95.72%,and the segmentation accuracy of sugarcane,rice,and other arable land increasing from 97.08%,64.6%,and75.31% to 97.33%,81.62%,and 94.93%,respectively. |
