Optimization Of Urban High-resolution Remote Sensing Image Semantic Segmentation Model Based On Deep Learning

Due to the rapid development of satellite technology,the data volume of high-resolution remote sensing images is growing rapidly,and manual visual interpretation methods are inefficient,time-consuming and laborious.With time going,surface objects undergo a series of changes,such as buildings and roads in cities.This requires people to extract and update information from remote sensing images in time to provide it for use in other areas such as urban planning and transportation planning.These tasks are difficult to accomplish by hand alone.Therefore,how to use computers to replace part of the work and even the entire process manually has become an important research direction.In recent years,the vigorous development of deep learning has provided an avenue for the automatic extraction of remote sensing image information.In this paper,we perform semantic segmentation of buildings and roads in urban high-resolution remote sensing images based on deep learning.Based on the characteristics of buildings and roads in remote sensing images,UNet and HRNet are choosen to finish model optimization.The main research contents of this article are as follows:(1)Aiming at the semantic segmentation dataset of urban high-resolution remote sensing images,UNet is choosen to change from module improvement,feature concatenation method,and model rebuilding with limited computing resources.This article uses the Residual module in ResNet,ResNeXt,and Res2 Net to improve the UNet.These modules are applied in UNet to replace single convolution layer,and these models are named UNet V1-V3 respectively.Experiments are performed on the Massachusetts building dataset.After comprehensive comparison,UNet V3 is the best of the three.This paper summarizes the methods of feature concatenation and proposes a tandem concatenation structure with weights.Through mathematical derivation and experiments,the reliability of the structure in reducing the semantic gap between shallow features and deep features is proved.Fused UNet(F-UNet)in terms of model building based on the first two points is rebuilt.With limited parameters,it obtained relatively accurate semantic segmentation results in urban high-resolution remote sensing images semantic segmentation tasks.At the same time,it also verifies that the tandem integration structure with weights can be used in different models to improve the accuracy.It shows a certain wide applicability.(2)Because road and building information extraction requires relatively high boundary information,this paper selects parallel network HRNet to optimize from three perspectives of feature concatenation,loss function and normalization method.Firstly,based on the HRNet multi-level feature concatnation method,a gate mechanism is introduced to reduce the semantic gap between different levels of feature maps,and Gated HRNet(G-HRNet)is built.Secondly,based on the problem of imbalanced sample in remote sensing images semantic segmentation,Weighted CrossEntropy(WCE)and focal loss(FL)functions are compared in G-HRNet and the outcome shows that FL performs better.Finally,due to the limitation of computing resources,the batch size parameter setting of Batch Normalization(BN)in G-HRNet is too small and the effect is not good.Two normalization methods,Group Normalization(GN)and Filter Response Normalization(FRN),which are not related to the batch size parameter,have been experimentally compared and the result shows that FRN has the best effect.When using FRN in G-HRNet,the mIoU of test results reaches 82.03% in the Massachusetts road semantic segmentation dataset and 83.91% in the Massachusetts road semantic segmentation dataset.