Land Use Classification Of High Resolution Remote Sensing Images Based On Convolutional Neural Networks

Land use classification of high-resolution remote sensing images is a basic issue in the field of land resource management.Due to the influence from both natural and social factors,there are great mutual interference among different land use types.How to further improve the accuracy of classification is a complex problem that has been confronted with in current research.And in this thesis,two scales,scene-level and pixel-level,of classification have been studied.The research of scene-level classification can provide fast and accurate scene judgment function,and has potential application value in scene retrieval and land resource investigation,etc.Moreover,The land use classification of pixel-level can accomplish interpretation of remote sensing images automatically and it is helpful to grasp the state of land resources and ecological environment.The related work in this thesis mainly focuses on the classification tasks in scene-level and pixellevel.In the scene-level land use classification task,the research work focuses on enhancing the expression ability of image features and more effective classification.The UC Merced land use dataset is used to fine-tune the parameters of the fully-connected layers of CaffeNet,VGG-S and VGG-F.Then the fine-tuned networks are used as the feature extractors to extract the image features,and cascading the outputs extracted from second fully-connected layer of fine-tuned CNNs,as the final expression of the image.Finally,the cascaded features are input into the mcODM classifier to obtain the classification results.Through the experimental verification,the following conclusions are drawn:(1)The method of multi-structure convolutional neural network features cascading can effectively remedy the shortcomings of single CNN in image information extraction.Compared with the single CNN model,the classification accuracy of proposed method in UC Merced land use dataset reaches 97.55%,which is 2%-5% higher than single CNN.(2)Fine-tuning CNNs fully-connected layer can effectively improve the classification performance of the model.In the experiments of fine-tuning CaffeNet,VGG-S and VGG-F,the classification accuracy of the fine-tuned CNNs is improved by 3%-5%.In the classification at the pixel-level,in order to further improve the classification accuracy,based on the idea of ensemble learning,several U-Nets are combined as one model.Each land type data set(only labeling the information of target and non-target land types)is trained separately for a U-Net,and the results of U-Net extraction after training are optimized by mathematical morphology.The extraction results of each U-Net model are evaluated according to F-measure,and the fusion priority is determined according to the evaluation results.Finally,the experimental results show that the optimization method based on mathematical morphology can improve the unreasonable phenomenon after UNet classification.The F-measure of three test images increases by 1.3%-1.5%,achieving 92.29%?93.66%? 87.96%,which proves the effectiveness and robustness of the method.