Research On Remote Sensing Image Segmentation Based On Convolutional Neural Network

Remote sensing images are an important data source for quickly obtaining large-scale ground information,and are widely used in many fields such as agriculture monitoring,urban planning,and map production and update.With the improvement of high-resolution satellite observation capabilities,accurate and rapid segmentation of high-resolution remote sensing images is of great significance for conducting academic research and guiding production practices.Image segmentation requires not only obtaining the semantic information represented by each pixel but also the segment to obtain the regional contour of the target.With the development of deep learning technology,the use of convolutional neural networks(CNNs)for remote sensing image segmentation has become a hot research topic.However,due to the complexity and various types of coverage in high-resolution remote sensing images,there were problems such as “same spectrum foreign matter” and “same object different spectrum” in surface cover and mutual obscuration of objects.CNNs gradually expand the receptive field and accumulate contextual information,resulting in the loss of high-frequency details of ground objects.At present,CNNs used for remote sensing image segmentation has some problems,such as fuzzy target boundary and missing small-scale target segmentation.In order to prevent the loss of high-frequency detail information of ground objects,CNNs have problems such as blurred ground object boundaries and missed classification of small-scale targets when segmenting remote sensing images.In response to the above problems,this paper designs two dual-branch CNNs around the collaborative modeling of edge prediction and semantic segmentation.The main work is as follows:1.Data set production.In this paper,128 Gaofen-2 images collected from Taian City,Shandong Province,and Ningxia Region were pre-processed.The pre-processed images were manually labeled,divided into data sets,and enhanced to get the data sets of Ningxia Wolfberry and Shandong Feicheng winter wheat.In addition,to ensure the fairness of the test model,the Gaofen Image Dataset(Gaofen Image Dataset)and the high-resolution water body Dataset are also collected in this paper.2.In order to solve the problem of poor edge segmentation of remote sensing image features,a multi-scale edge feature semantic feature fusion strategy is used to compensate for the loss of edge detail information to extract edge feature information of features in different areas.A dual-branch model architecture of semantic features and edge features is proposed.The EFFNet(Edge Fine Feature Network),a high-resolution remote sensing image segmentation method combining semantic features and edge features,is established.3.In order to optimise the edge feature extraction method and reduce human factor interference,a two-branch model architecture of semantic features and edge features is designed.The HED network(Holistically Nested Edge Detection Network)is combined with the semantic segmentation network of remote sensing images.The ESNet(Edge Feature and Semantic Feature Fusion Network)model is constructed to train the edge detection network using multi-scale semantic feature maps as input.This model uses the designed weighted edge loss to enhance the structure to adjust the two branches of the network,which can retain more boundary information in the network.The experimental results show that the overall accuracy of ESNet on the GID data set reaches 92.47%,and the average extraction accuracy index reaches 93.04%,which is better than the comparison model in the experiment.The results show that the introduction of edge features improves the segmentation accuracy of remote sensing images.Experiments are performed on two kinds of crop data sets and public data sets.The results show that the experimental method can reduce the feature difference between the edge pixel and the inner pixel of the object.As well as obtaining pixel-by-pixel feature vectors with high inter-class differentiation and intra-class consistency,the problem of high noise of boundary blurring results was effectively solved.Highly accurate experimental results were achieved.