Research On Object Classification In High-resolution Remote Sensing Images Based On Semantic Segmentatio

High-resolution remote sensing images provide rich information in many fields such as map mapping,agricultural management,environmental protection and disaster emergency response.Achieving fine classification of ground cover is one of the key tasks in remote sensing image interpretation.Unlike natural scenes,high-resolution remote sensing image feature classification usually faces a wide variety,complex small targets and incomplete edge segmentation.In this paper,we analyze remote sensing image data sources and use a semantic segmentation method based on deep learning to improve the above problems,and the main research contents are as follows:(1)The characteristics of high-resolution remote sensing image dataset are studied.In this paper,two publicly available datasets,GID and Love DA of HIS II,are selected for study.The remotely sensed sub-images after pre-processing are counted in terms of pixel points for feature sample categories,and penalty weights are calculated and used in the weighted cross-entropy loss function of joint Dice coefficients to weaken the impact of uneven distribution of feature samples on model training.In terms of dataset enhancement,random cropping and hybrid stitching methods,as well as augmentation strategies of geometric and color transformations are used to expand the remote sensing image dataset,which improves the robustness and generalization ability of the model.(2)Improved DeepLabV3+ semantic segmentation model design.In order to solve the phenomena that the segmentation of small target feature categories in remote sensing images is not fine enough and large size feature categories are prone to holes,this paper chooses DeepLabV3+ network combined with coordinate attention mechanism,respectively,in order to improve the model’s attention to important feature layers and the location of the region of interest,so as to pay more attention to feature boundaries.In addition,a multi-branch feature fusion unit is introduced in the decoder part,which can effectively extract contextual semantic information and help the whole model to better identify remote sensing image feature targets.(3)Comparative experiments on feature classification of high-resolution remote sensing images based on semantic segmentation model.Firstly,it is verified that the loss function proposed in this paper can improve the focus on weak samples compared with the traditional cross-entropy loss.Secondly,the reliability of the module introduced in this paper is verified in the ablation experiments.Finally,the comparison experiments of feature classification are conducted,and the overall accuracy(OA)and mean intersection over union(MIo U)of the improved method in this paper are improved by 5%,2.03% and2.66% and 1.2%,respectively,compared with the baseline model in the five and fifteen classifications of GID;in the six classifications of Love DA,the OA and MIo U of the improved model are improved by 4% and 1.62% compared with the baseline model.It is confirmed that the improved method in this paper can effectively improve the segmentation effect and accuracy.(4)This paper takes Taihu Lake and the surrounding area at the south edge of the middle and lower reaches of the Yangtze River as the research object,and selects Band2,Band3 and Band4 in visible wavelengths as the data sources,and produces the Sentinel 2remote sensing data set by itself.The GID five classification was used as the source domain task for migration learning.The experimental results show that the use of migration learning can effectively improve the convergence speed of the model in new tasks,reduce the training cost and confirm the portability of the method in the relevant feature classification tasks.