Research On Semantic Segmentation Of High-resolution Remote Sensing Image Based On Deep Learning

Semantic segmentation of high-resolution remote sensing images has great application prospects in territorial planning,smart cities and other fields.With the continuous improvement of remote sensing image resolution and resolution,various fields have also put forward higher requirements for the accuracy of remote sensing image segmentation and extraction.However,how to improve the fineness of the target boundary division and the accuracy of the recognition of the types of ground objects in remote sensing applications has always been a difficult point in the process.In addition,the number of remote sensing images is gradually increasing,and the research field lacks efficient automatic segmentation methods.The in-depth application of remote sensing images in various fields is greatly restricted.In recent years,the deep learning method based on neural network has become a research hotspot of remote sensing image processing as a breakthrough to solve the above problems.The mainstream deep learning methods in the current research are not designed for the characteristics of high-resolution remote sensing images,so the ability to extract features of different scales in the images and the accuracy of the segmentation results need to be further improved.Aiming at the above problems and combining the characteristics of high-resolution remote sensing images,this paper improves and optimizes the existing models and methods to obtain higher semantic segmentation accuracy.The main research work of this paper includes the following aspects:(1)In view of the large size of remote sensing images,a suitable preprocessing scheme is designed.Through experiments,the effect of the size of image blocks and the coverage between blocks on the segmentation results is analyzed,and appropriate preprocessing parameters are selected,which effectively improves the segmentation accuracy of the model.(2)In order to improve the segmentation accuracy of high-resolution remote sensing images,this paper designs and implements a multi-scale feature fusion model MFFU-Net based on the U-Net encoder-decoder structure.This model can effectively solve the problems of complex feature information,uneven distribution of categories,large scale gaps,and indistinguishable boundaries in high-resolution remote sensing images.The model first uses the multi-scale jump connection algorithm to make full use of the multi-scale characteristics of the remote sensing image,and then designs the channel attention algorithm to optimize the fused feature map,strengthen the expression of key features to achieve the improvement of segmentation accuracy,and finally embed the improved design in the network.The hollow space pyramid pooling model is further integrated with the semantic information of the image.This paper selects the Vaihingen,Potsdam and GID data sets released by ISPRS and the GID data set released by Wuhan University to conduct experiments.The results show that the improved algorithm obtains all three data sets with their own characteristics.With good results,the F1 score reached 90.2%,90.4%,and 84.3%,respectively.The performance is better than the more popular general segmentation network,and it has a significant improvement compared to the basic network U-Net.(3)In order to further enhance the effect of image segmentation and enhance the application value in actual engineering,this paper studies the multi-scale feature fusion model and proposes the segmentation network MFFU-Net+for the Gaofen-2 image.The network first improves the channel attention algorithm in MFFU-Net,calculates multi-dimensional attention information in parallel,and effectively utilizes the spatial information of the feature,and then designs an upsampling algorithm based on the overall pixel mapping,which reduces linear interpolation.At the same time of information loss,the global information of the feature is effectively used.Finally,a hybrid loss function is designed according to the characteristics of remote sensing images to solve the problem of category imbalance and enhance the classification effect.Experiments with the GID data set show that MFFU-Net+can effectively improve the classification accuracy of various objects,especially for difficult-to-classify samples.The average F1 score has increased by 2.3%.The experiment selected the South-toNorth Water Diversion outside the data set.In the project,the Gaofen-2 remote sensing image in Xingyang City has a significantly improved prediction result compared with the basic network,which further proves the excellent performance and practicability of the model.