Urban Green Space Classification Based On GF-2 Remote Sensing Image With Model Res2UNet-M-ECA+

Urban green space is an important part of urban ecosystem,and it has the important functions including improving environmental quality,maintaining diversity,beautifying landscape,recreation,cultural creation,disaster prevention and green space economy.Urban green space classification based on high-resolution remote sensing images is an important basic data in many scientific researches of urban green space,such as the development level of urban green space,ecological benefits of urban green space,urban heat island and so on.However,the efficiency of traditional classification of urban green space based on high-resolution remote sensing image is low.Therefore,it is necessary to carry out intelligent urban green space classification research based on high-resolution remote sensing images.The deep learning semantic segmentation infers the label of each pixel for fine-grained reasoning and intensive prediction,which can effectively and accurately classify urban green space based on high-resolution remote sensing images.But the current researches still have the following shortcomings: the deep learning model for urban green space classification does not make full use of multi-scale context information,so that it is difficult to cope with the challenges brought by the rich spatial information of high-resolution remote sensing images and some sporadic and irregular urban green spaces,which makes it difficult to meet the data requirements of a variety of urban green space subtypes(trees,shrubs,grasslands,etc.)in some related researches.To solve the shortcomings,the following researches are carried out:(1)A new convolutional neural network model Res2UNet-Modified(Res2UNet-M)was constructed.Res2UNet-M was formed by using Res2 Net as the U-Net-like encoder,reconstructing the U-Net-like decoder,and introducing the group normalization(GN)method and Mish activation function.Res2UNet-M used skip connection between the encoder and the decoder,shortcut connection and residual-like connection of the bottle2 neck module to enhance the ability of extracting and utilizing multi-scale context information.(2)A new attention ECA+(Improved Efficient Channel Attention)was constructed.A parallel feature extraction branch based on global maximum pooling was added in the ECA(Efficient Channel Attention)attention block,and the ECA+ block with better feature recalibration effect was formed.From the global perspective,the important features in the model were paid more attention with the help of ECA+,indirectly enhancing the extraction and utilization of multi-scale context information,and restraining the influence of redundant and irrelevant information in high-resolution remote sensing images.(3)By integrating Res2UNet-M model and ECA+ block,Res2UNet-M-ECA+ model was formed,which realized the high-precision classification of eight urban green space subtypes including arbor woodland,shrub land,garden plot,natural grassland,artificial grassland,paddy field,irrigated land and dry cropland.The comparison experiment between Res2UNet-M,U-Net,Res UNet and Deep Lab V3+ model shows that Res2UNet-M model achieves better classification effect,and the results also provide references for the selection of depth,width and scale of the model.The comparative experiment between ECA+ block,channel domain attention SE(Squeeze & Excitation)and ECA block shows that the feature recalibration effect of ECA+ block is better than that of SE and ECA block,which can help to improve the classification effect of deep learning models.The Res2UNet-M-ECA+ model,which integrates Res2UNet-M model and ECA+ block,achieves good results in fine classification of urban green space.This study improves the classification effect of deep learning model from the local view(different dimensions of the model)and the global view(attention mechanism),and proposes a Res2UNet-M-ECA+ model applying to urban green space classification based on highresolution remote sensing images,serving the related researches of urban green space,which may have certain references to improve the precision of deep learning semantic segmentation based on high-resolution remote sensing images.