Research On Vegetation Extraction From Remote Sensing Images Based On DeepLabV3+ Network

With the advancement of urbanization and the universality of human activities,the ecological environment has been affected to a certain extent,and vegetation as an important factor in it,its accurate extraction has become particularly important,accurate acquisition of vegetation category information,promote the understanding of the ecological environment and help future urban greening planning.With the continuous improvement of image resolution and the gradual enrichment of acquisition methods,remote sensing images contain more obvious spectral,texture,and spatial information,and how to quickly and effectively obtain information from remote sensing images has become a key issue in current research.In recent years,the rapid development of deep learning technology has made it possible to automatically and efficiently extract shallow and deep semantic information from images,providing technical support for accurate and efficient extraction of high-resolution remote sensing image information.However,the multi-layer neural network structure of deep learning brings a lot of computation,and how to improve the network performance while reducing the number of parameters is a problem that must be faced.Combined with deep learning,the methods of vegetation classification extraction in remote sensing images are studied.The main work is as follows:(1)Based on high-resolution remote sensing images,a vegetation sample set is constructed,using a 2-meter resolution GF-1D image of the central urban area of Chongqing as the data source.After image preprocessing,a custom label is created by combining standard false color images with true color images through GIS vectorization,and then the overfitting phenomenon is avoided by using data enhancement to generate a vegetation sample set,which is used to achieve the extraction training and verification of different models.(2)Aiming at the problems of using semantic segmentation networks to extract vegetation,such as excessive parameters,low computational efficiency,and poor spatial scale adaptability,a lightweight DeepLabV3+ network model was proposed for vegetation extraction.Firstly,Mobile Net V2 is used as the backbone feature extraction network architecture,and the amount of model parameters is reduced through a deeply separable convolution and reverse residual module to ensure the basic performance of the model;Secondly,the convolution layer and expansion coefficient of the Atrous Spatial Pyramid Pooling(ASPP)are adjusted to enhance the extraction ability for different sizes of woodland and grassland.Based on sample 1 of true color images,vegetation extraction is conducted,the experiment shows that the improved fusion attention mechanism lightweight DeepLabV3+ network model performs better than PSPNet and the initial network in terms of vegetation extraction,compared to the DeepLabV3+ network model,the number of parameters is reduced from 54708931 to6617107,improving the computational efficiency of the network and enhancing the vegetation extraction effect on the premise of enhancing the classification effect for forest land and grassland of different sizes.(3)Based on the lightweight DeepLabV3+ network,attention mechanism structures of channel and spatial types are added to the network Encoder layer to obtain accurate vegetation and edge feature information.Experiments show that the vegetation extraction accuracy of the lightweight DeepLabV3+ model integrated with scSE attention mechanism reaches 78.64%,88.11% and 91.69%,respectively,in terms of mean intersection ratio(MIoU),mean pixel accuracy(MPA)and accuracy(ACC).respectively,with an increase of 3.62%,1.54%,and 1.39% compared to the DeepLabV3+ network model,and can better avoid the impact of other factors on forest land and grassland extraction.(4)According to the improved vegetation extraction model,three different sample data sets are trained,and the impact of different image enhancement methods on vegetation extraction is analyzed.The experiment shows that the extraction effect of sample 3 after vegetation enhancement is significantly enhanced,and compared to sample 1 of true color images,the three accuracy indicators reach 80.61%,88.97%,and 92.66%.Compared with the results of traditional vegetation extraction methods,the model obtains good vegetation extraction results when using vegetation enhanced images as sample base images.