Convolutional Neural Network Based On Remote Sensing Image Classification

The accuracy and efficiency of remote sensing image classification is one of the important research contents of remote sensing image analysis.As artificial intelligence gradually displays the advantages of processing information,using neural networks to process remote sensing images becomes a more efficient method.Compared with traditional neural networks,deep neural networks have more computational levels and can apply statistical learning methods to massive data,extracting remote sensing image information from the perspective of computer vision,and tapping potential information of massive data more fully and accurately,Improving the use value of information.Remote sensing images use different pixel values or brightness differences and spatial layout changes to distinguish different objects,which is also the physical basis of classifying remote sensing images.By extracting the spectral information of various objects contained in the remote sensing information as a classification basis,the features in the image are identified for classification processing.Based on the application of desertification monitoring in Ningxia region,this paper uses the MODIS remote sensing image datas from 2000 to 2015,analyzing the normalized difference vegetation index(NDVI)in Ningxia,and analyzes the spatial and temporal changes of vegetation cover in Ningxia during the past 16 years,exploring the relationship between vegetation cover change to population and economy.This paper focuses on the convolutional neural network model in deep neural network.Convolutional neural network shows strong feature extraction ability for image classification.Its essence is the input-to-output mapping,which contains two layers of basic structure.One layer is the feature extraction layer,and the other layer is the feature mapping layer.It has the characteristics of local receptive fields and shared weights,which can effectively reduce the number of parameters in the training process and reduce the training difficulty.The experiment categorizes remote sensing images based on the convolutional neural network model of UNet,SegNet and DeconvNet with three symmetrical structures.The jaccard coefficients of the three models are calculated as the accuracy evaluation index.This experiment builds neural network models based on the deep learning development framework of Keras and Theano.The data adopts WorldView satellite multispectral remote sensing image.From the experimental results can be drawn:(1)The UNet network model is superior to the SegNet and DeconvNet network models in the classification of remote sensing images.The average jaccard coefficient of the Net network is 0.765,which is higher than 0.732 of the SegNet network and 0.706 of the DeconvNet network,and the UNet network training time is the shortest.The classification accuracy of the SegNet network model is higher than that of the DeconvNet network model,but the training time of the SegNet network is much shorter than that of the DeconvNet network.The three models have higher classification accuracy for water bodies and vehicles,accuracy is around 0.95.The classification result of buildings is obviously higher than that of misc.manmade structures.The classification result of roads is better than that of tracks.(2)Utilizing three convolutional neural networks to classify remote sensing images of Ningxia region and extracting trees and crop regions in Ningxia respectively.The classification results show that UNet convolutional neural network has the best classification effect on vegetation coverage areas.Compared with NDVI,the edge of the vegetation area in the convolutional neural network is more refined.