Research And Implementation Of Satellite Image Semantic Segmentation Technology Based On Deep Learning

Semantic segmentation of satellite image plays an important role in geographic information exploration and regional architectural planning.At present,the semantic segmentation algorithms based on deep learning perform well in the scenes such as street scene and indoor.However,due to the interference of shadow occlusion and tree coverage in satellite images,the segmentation results generated by these algorithms are unsmooth and inaccurate.In this thesis,two segmentation algorithms are proposed to address the problem of the unsmooth and incorrect results in the following two tasks: buildings segmentation and roads segmentation,respectively.For the task of building segmentation,since the buildings in the satellite image are affected by shadows and trees,the segmentation results,which are obtained by known neural network based semantic segmentation algorithms,have the following problems: unsmooth edges and incomplete segmentation of large-area building.For dealing with the problem of unsmooth edges,this thesis proposes a building semantic segmentation algorithm based on the Generative Adversarial Networks,which makes the segmentation model to generate smoother,more accurate edge details.For addressing the problem of incomplete segmentation of large-area building,the dilated convolution is used to provide a larger receptive field to optimize the characteristics of the interior area of large-area buildings.The experimental results show that the proposed algorithm effectively solves the two above problems.For the task of road segmentation,the known neural network based semantic segmentation algorithms will produce the results of discontinuous road segmentation because the roads in the satellite image are affected by trees and the road features that are not obvious.Unlike the building segmentation,there is less road area in the local space in satellite imagery,and dilated convolution cannot be used to recover the interrupted roads.For addressing the problem,this thesis proposes a road segmentation algorithm based on Variational Auto-Encoder.By pre-training the Variational Auto-Encoder,it will obtain the distribution of the road data.At the same time,the Variational Auto-Encoder is embedded into the neural network-based segmentation model to repair the interrupted roads.The experimental results show that the algorithm can extract the connected road information.