Semantic Segmentation Of Remote Sensing Image Based On Deep Learning

Deep learning is a significant subdomain of machine learning.With the high speed of development in recent years,it has greatly promoted the application of artificial intelligence in various fields.Deep learning is based on artificial neural network and driven by data.With the arrival of big data and the improvement of computing ability of computer hardware,deep learning can apply deeper network structures to solve more complex problems.In this paper,the topic is the semantic segmentation task of remote sensing images,and we focus on the problems which exist in the deep learning semantic segmentation algorithm when applied to the remote sensing images.Semantic segmentation is of great significance in remote sensing images,and it is the basis of land usage and environmental change monitoring.Traditional machine learning methods need to design feature extractors manually,it is difficult to extract abstract semantic features effectively.While the deep learning networks can extract more abstract,sparse and invariant features based on massive data,which have become the primary method to solve the problem of semantic segmentation.Current research mainly focuses on two aspects:applying different network structures to improve the segmentation accuracy;reducing network parameters and computational overheads to meet the real-time requirements with a relatively high segmentation accuracy.In the view of the above problems,the specific research contents of this topic are as follows:For the high-resolution remote sensing images,a high precision semantic segmentation network based on deep learning is proposed.The network combines the advantages of the encoder-decoder structure and the atrous spatial pyramid pooling(ASPP).It takes the deep residual network(ResNet)followed by the atrous spatial pyramid pooling as the encoder and constructs two scales of feature fusion as the decoder in the upsampling stage.Moreover,a multi-scale loss function is developed to enhance the training procedure.In the post-processing,a novel superpixel-based dense conditional random field is employed to refine the predictions.Experimental results on the open access Potsdam and Vaihingen datasets demonstrate that the proposed network performs better than other machine learning or deep learning methods.Compared with the state-of-the-art DeepLab_v3+network,the proposed network gains 0.4%and 0.6%improvements in overall accuracy on these two datasets respectively.For the high-resolution remote sensing images,a lightweight semantic segmentation network based on two branch structure is proposed,which includes the semantic branch and the spatial branch.A simple structure of three convolution layers is used to extract spatial features in the spatial branch,and the lightweight network MFNet and the pyramid pooling module(PPM)are used to extract semantic features in the semantic branch.In the feature fusion stage,the channel attention module is used to optimize the extracted features.At the same time,the weighted multi-scale loss function is applied to enhance the training process of the network.Experimental results on the open access Potsdam and Vaihingen datasets demonstrate that proposed network greatly reduces the network parameters and computational overheads with high segmentation accuracy.Compared with the widely used U-Net,the network parameters are reduced by 61%,and the computational overheads are reduced 96%,while the segmentation accuracy is improved by 1.5%.For the hyperspectral remote sensing images,a high precision deep learning network based on dual branch structure is proposed,which includes the spectral branch and the spatial branch.The spectral branch adopts one dimensional convolution to extract the spectral features,and the spatial branch uses two dimensional convolution to extract spatial features after the PCA preprocessing.At the same time,in each branch,the residual connection and the dense connection are utilized to enhance the feature extraction ability further.Experimental results on the open access datasets Indian Pines,Pavia University,Pavia Center and Salinas show that the proposed network achieves higher segmentation accuracy.