Segmentation Method Of GF-1 Remote Sensing Image Based On Full Convolutional Network

Accurate information for land,woodland,grassland and water area plays an important role in agricultural production.Using GF-1 remote sensing image as the data source and deep learning technology,high precision fields,woodlands,grasslands and waters can be obtained,which can promote the development of agricultural production.With the proposed convolution network,the segmented image can be input to any size of the image,and it improves the processing speed compared with the convolutional neural network.The full convolution network has a significant effect in dealing with natural images,but it is not suitable for processing remote sensing images.The main reason is that the full convolution network removes the full connection layer of the convolution neural network,and there is still too much pool layer.Although the pool increases the receptive field of the upper convolution core,the partial location information is discarded at the same time that the aggregation background is discarded.The more the pool layer is,the more information is lost.Another reason is that the resolution of high score 1 remote sensing image can not reach the standard of natural image,that is,the expression of high score 1 remote sensing image can not reach the standard of natural image.However,the semantic segmentation method needs to adjust the category map precisely,so we need to retain the location information which is discarded in the pool layer.This paper presents a new network structure high score 1 remote sensing image segmentation network based on full convolution network based on full convolution network.The main work is as follows:1.Data set production.After fusion,the whole range of GF-1 remote sensing image is very large.If direct input network training,the video card will have memory overflow problem.In this paper,ENVI 5.2 is used to cut out the image of JPG format with a resolution of 300*400 on the fusion GF-1 remote sensing image,and then annotate the remote sensing image with the annotation software.The gray value is marked as a value from 0 to N,that is,pixels belonging to different objects are marked 1,2,3,respectively.N-1,N,and the pixels of other objects are marked 0.2.Improved bilateral filtering algorithm.The segmentation results of different terrain types obtained by using the network model have high overall segmentation accuracy,but in the investigation of each object type,the edge of the ground objects is noisy,because the edge of the surface of the remote sensing image is complex.In order to improve the time efficiency,this paper improves the bilateral filtering algorithm,separates the two-dimensional convolution operation into two one-dimensional convolution operations.At the end of the neural network model,a layer of edge correction de-noising layer is added to the neural network,and the results of the original neural network segmentation are further processed.3.Based on full convolution network,a segmentation method(Remote Sensing Segmentation Full Convolutional Network,RSSFCN)is designed.The segmentation method can retain the spatial location information of GF-1 remote sensing images,and can directly input any size image without distortion to the network.Partition and output segmentation results.Experimental results show that the method proposed in this paper not only improves the time efficiency of processing GF-1 remote sensing images,but also has a total segmentation accuracy of 90.3%.