Remote Sensing Image Land Cover Classification Based On Convolutional Neural Network

Along with the development of remote sensing technology,remote sensing data plays a positive role in monitoring the global resources dynamic changes and land coverage and remote sensing image land cover classification is an important directioin of the application of remote sensing data.How to further improve the accuracy of remote sensing image land cover classification is very import to the application of remote sensing data.The traditional remote sensing image supervised classification algorithm classifies the image mainly based on pixels according to the spectrum of the ground object.However,due to the complexity of the image environment and the dynamic changes of the imaging environment,traditional method based on spectral data is always cannot achieve satisfactory classification results.In addition,the researchers also proposed a method based on the texture neighborhood information to make up for spectral features.The traditional classification method based on spectral features have solved the the problem of big workload by eyes interpretation in some means,the accuracy of classification is still not satisfactory due to the limitation of the semantic expression ability of the designed features with poor generalization ability.In recent years,the convolution neural network has made a breakthrough comparing with traditional methods in image classification and identification of the ImageNet Large Scale Visual Recognition Challenge(ILSVRC)and apply the powerful feature learning ability of the convolutional neural network to land cover classification becomes the focus of current research.Different from ILSVRC image classification,land cover classification usually collects neighborhood information around a sampling point for training data and has the characteristics of fewer samples,smaller sample size and fewer classification categories.To AlexNet,VGG and other classical convolutional neural network model,sample size and network input and output requirements in the solution of pixel level remote sensing images land cover classification exists not matching problem,the input layer of the model design too small to construct the network structure with a certain depth,the sample size too large to overflow the information of center point.Therefore,this study reference to the AlexNet ILSVRC model designed for land cover classification on the input size of 27×27 pixel size with three convolution layer?two pooling layer and two fully connection layer model named LCNet-27(Land-Cover convolutional neural Network-27)and an input size of 13×13 pixel size with two convolution layer?one pooling layer and two fully connected layer model named LCNet-13.TM medium resolution image and QuickBird high resolution image are utilized as the experimental data.The different depth models based on different input sizes,the sample sizes and different resolutions on the classification result are analyzed and compared with the traditional methods based on spectral features and spectral textures features.Experimental results show that LCNet-27 three convolution layer two pool layer compared with the LCNet-13 two convolution layer of a pool layer model of the model classification effect is better;due to the AlexNet model finetune with model adaptability,the oversampling is too serious to add too much noise,and the classification result is slightly worse than the classification method of spectral and texture;for LCNet-27 the training sample size of 5×5 pixels is the best for LCNet-27 in TM medium resolution image and overall classification accuracy is 96.6%.Bigger sample size will yield obvious filtering effect to classification results and reducing the detail information;however,smaller size might cause serious misclassification because of containing too fewer pixel information.Besides,the classification results of the proposed LCNet-27 show good consistency with less broken patches,effectively reducing the post-classification process.For QuickBird high resolution image the training sample size of 7×7 pixels is the best for LCNet-27 and overall classification accuracy is 97.4%;comparing with TM image has better visual and spatial features,the filtering effect is relieved;the detail information is better preserved and the precision is significantly increased,since LCNet-27 becomes more robust to the size of training samples.The classification accuracy of LCNet-27 is higher than those of the traditional methods based on spectral features and spectral + texture features,which indicates that the proposed method can be well applied to deal with land cover classification tasks.