Remote Sensing Image Semantic Segmentation Based On Convolutional Neural Network

Remote sensing images contain a wealth of geospatial information which has important practical value for urban planning,making and updating of topographic maps,monitoring and management of forestry resources,etc.Over the last few years,semantic segmentation of remote sensing images has received widespread attention.However,due to reasons such as distance,illumination,ground objects,and environment in the imaging process.Objects of different categories in remote sensing images may have similar visual features,while objects of the same type may have large differences.This leads to uncertain phenomena in semantic segmentation of remote sensing images.As one of the representative deep learning algorithms,convolutional neural networks can have been extensively used in remote sensing field due to their superior nonlinear representation capabilities,which can learn deeper and more essential features from massive sample data.Therefore,this paper focuses on the application of convolutional neural networks in the semantic segmentation of remote sensing images,overcomes the problems of intra-class heterogeneity and inter-class complexity in remote sensing image segmentation.The main research work is summarized as follows:(1)In view of the shadow interference,inaccurate segmentation,and complex background information in the remote sensing image segmentation process,a fuzzy neighborhood convolutional neural network is proposed.First,the idea of fuzzy learning is introduced into deep learning,and the fuzzy neighborhood module is used to calculate the fuzzy similarity between samples,which effectively reduces the influence of noise on image segmentation and improves the segmentation accuracy.In addition,a set of attention gating modules are added,which uses low-level features to provide guidance information for highlevel features to highlight the target object in the feature map through weighting indices and separate the target object from the complex background information to achieve fine segmentation of remote sensing images.Finally,experimental results on three different segmentation datasets suggest that the proposed fuzzy neighborhood convolutional neural network has high segmentation accuracy.(2)In view of the complex scale information and object texture similarity in the remote sensing image segmentation process,a fuzzy multi-scale convolutional neural network is proposed to explore in depth the remote sensing image segmentation method based on convolutional neural network.Firstly,to increase sensitivity to objects of different sizes and shapes,multi-level semantic information is obtained using residual parallel branching and fused to the backbone network.And then the fuzzy similarity between pixel samples is calculated in order to reduce the interference of noise and mixed pixels on the segmentation accuracy of remote sensing images.Subsequently,a multi-scale feature extraction module is embedded in the network structure to effectively encode the high-level semantic feature maps and enrich the final feature representation capability by controlling the size of the receptive domain.Finally,experimental results on two different segmentation datasets suggest that the proposed network can achieve category classification well.