Research On Land Use Classification Of High-resolution Remote Sensing Images Based On Deep Learning

Unmanned Aerial Vehicles(UAV)remote sensing has become one of the important means to acquire high-resolution remote sensing images due to its low cost,fast data acquisition,high resolution and strong real-time performance.At present,high-resolution remote sensing images are widely used in military and civilian fields,and play an important role in urban planning,land and resources planning,etc.At this stage how to make full use of high resolution remote sensing image to observation becomes especially important to intelligence,in recent years,based on in-depth study of semantic segmentation technology is gradually applied in remote sensing land classification,building extraction,etc,and made some achievements,but there are still lack of data sets,segmentation problem such as low precision,large amount of calculation model.Aiming at the above problems in semantic segmentation of high-resolution remote sensing images,this thesis has done the following research:(1)Establish high-resolution UAV remote sensing image annotation data set.Aiming at the shortage of semantic segmentation data sets at present,the UAV remote sensing technology was used to obtain the high-resolution UAV ortho image in the research area.On the basis of slicing it by QGIS,the UAV data set needed for the research was made by EISeg annotation software.Finally,the UAV data set is enhanced from two aspects:image geometry transformation and image pixel transformation.(2)In order to solve the problems of low segmentation accuracy and detail information loss in the model,this thesis proposes Deep Lab V3+＿CMASPP model in Deep Lab V3+ model,which uses Res Net101 as the backbone feature extraction network.CBAM attention mechanism was added to the first and last convolution of Res Net101 network.At last,the common convolution was replaced by deep separable convolution while the ASPP cavity rate was changed to 1,5,7 and 9.The experimental results show that the improved model has made great progress compared with other classical network models in UAV data set,m Io U and MPA are 68.03% and 80.10%,respectively.(3)In order to solve the problems of complex model structure,large amount of computation and low segmentation efficiency,the lightweight Deep Lab V3+＿MGCA model is proposed.Based on Deep Lab V3+＿CMASPP model,the lightweight network Mobile Net V2 is replaced by Res Net101 network first.In order to ensure the segmentation accuracy of the model,the attention mechanism is added to the trunk feature extraction network and group normalization is used to replace the batch normalization operation in the network.The experimental results show that the m Io U of the improved model is 67.69%and the number of references is only 13.33 MB on the UAV data set,indicating that the model has achieved an effective balance between segmentation efficiency and segmentation accuracy to a certain extent.In order to verify the generalization ability of the improved model,this thesis compares and verifies the above improved model and other classical models on the public remote sensing dataset WHDLD.Experimental results show that compared with other models,the improved model proposed in this thesis has certain advantages in remote sensing image land classification.