Research On Very High Spatial Resolution Remote Sensing Image Classification Method Based On Deep Learning

Regarding the classification problem of high spatial resolution remote sensing images,the researches on the methods of high spatial resolution satellite image classification has been conducted in this dissertation based on the principles of deep learning.The detailed research contents are given as follows:(1)An end-to-end and pixel-to-pixel semi-supervised high resolution remote sensing image classification theory is proposed in this dissertation.Considering that the traditional CNN based methods have difficulties in generating pixel-to-pixel dense classification results differently,a de-convolution network design is employed in this dissertation,realizing the real end-to-end and pixel-to-pixel classification of high resolution images and increasing the efficiency of the neural network on high resolution image pixel wise classification.Moreover,using deep learning for network supervised training requires a huge amount of labeled data as ground truth.However,acquiring such amount of labeled data is difficult in remote sensing classification domain.Regarding the problem of insufficient training samples,I propose using massive unlabeled high resolution remote sensing images for the unsupervised learning of structure features to assist the labeled high resolution remote sensing image for the better supervised feature extraction and classification training.The balance between general image structure feature learning and object class related distinctive learning is made to increase the network robustness of neural network,decrease the dependency of the supervised classification to labeled sample,and increase the classification accuracy of the network when the labeled samples are scarce.(2)Methods of feature extraction and classification on high resolution remote sensing image based on space-class information decomposition is proposed in this dissertation.The union of the remote sensing images from different classes can provide more abundant information,but the images from different sources have different merits,and using the same model can hardly take the full advantages of images in different classes.Regarding the utilization of the union of the images from different sources,an aggregation model of double line convolutional neural network is proposed in this dissertation.The model constructs the different networks based on the characteristics of color and multi-spectral images respectively,and increases the capability of network utilizing the low level features such as object boundaries and locations,and high level features such as semantic information jointly from the perspectives of “what” and “where”.Additionally,in order to increase the feature extraction capability,a reinforcing structure for feature extraction is constructed in the principle of generative adversarial network derived from “zero sum game”,to enhance the feature extraction capabilities of the network,ensure the effectiveness of the network,and increase the accuracy of the algorithms.(3)A pixel-wise classification method oriented to multi-scale local spatial structure information is proposed in this dissertation.Considering that receptive fields in different sizes can observe different local spatial structure features and influent the semantic understanding of high resolution images differently,I propose using a feature extraction method utilizing variant scales which unifies the spatial-spectral information and exploits the neighborhood environments of the ground objects in different scales by information mining in different local spatial structures to provide more abundant feature for ground object identification,in order to make the addictiveness of the ground objects in different scales stronger to increase the feature extraction ability on high resolution remote sensing images for better image classification.Additionally,regarding the network training problem based on deep learning,I propose constructing a feature reusing network by densely concatenation and configure the internal classifier to conduct companion supervision,which deepens the network while decreases the feature redundancy,decrease the transparency of the hidden layers,and facilitate the information feedback of the gradient information of the back propagation in order to increase the expressive ability of the network to some extent when the network goes deeper to achieve better high resolution remote sensing image classification results.(4)Experiments are conducted using the proposed methods on common Quickbird,Geoeye,BJ02,GF02 remote sensing image data and ISPRS Vaihingen public data collection,etc.to explore the influences of various elements such as network structure design and network parameters to the network's performance to validate the rationality of the network design.Meanwhile,three proposed methods are compared with other popular methods in the remote sensing domain The characteristics and advantages of these three proposed methods are clarified,and the superiority of the proposed methods to the other methods are proved.Based on the characteristics of the algorithms,we can choose more appropriate classification methods according to classification aims and emphases in the future tasks,in order to achieve better results inhigh resolution remote sensing image classification.