Domain Adaptation For Deep Feature-based High-resolution Remote Sensing Image Scene Classification

With the continuous development and improvement of global earth observation system,the acquisition of geospatial information has entered the era of "three highs"(high spatial resolution,hyperspectral resolution and high temporal resolution)and "three multi"(multi-platform,multisensor and multi-angle).A large number of remote sensing data has become an important source of information for human beings to understand the world.With the gradual improvement of spatial resolution of remote sensing image,remote sensing can provide complex surface structure information and provide data sources for higher-level semantic information extraction.However,the pixel-level and objected-oriented methods cannot describe the high-level semantic information.Therefore,high resolution remote sensing image scene classification has become an activate research topic in the remote sensing community.In recent years,several scene classification methods have been proposed for high-resolution remote sensing image based on deep learning,which can extract the semantic information effectively.However,these methods follow a basic assumption: training data and testing data should be sampled independently from an identical probability distribution,which is almost impossible in the real world.Therefore,when the source data set is far from the target data set(also known as data shift),transferring strategies for pretrained deep convolutional neural network(DCNN)is likely to yield unsatisfactory results.To overcome the existing problems,this thesis studied the domain adaptation method for deep feature-based high-resolution remote sensing image scene classification.The main contributions of this thesis are as follows:(1)An adversarial domain adaptation framework is proposed for deep feature-based highresolution remote sensing image scene classification.A DCNN is used to build feature representations to describe the semantic content of scenes before the adaptation process.Then,adversarial domain adaptation is used to align the feature distribution of the source and the target.(2)An adversarial domain adaptation method is proposed for deep feature-based highresolution remote sensing image scene classification.Specifically,a VGG16 pretrained DCNN is used to build feature representations to describe the semantic content of scenes before the adaptation process.Then,the discriminator is learned by minimizing the classification error associated with distinguishing the source from the target domains,while the generator learns transferable representations that are indistinguishable to confound the domain discriminator,and thus align features across domains.(3)A classifier-constrained adversarial domain adaptation method is proposed for high resolution remote sensing image scene classification.Specifically,two different land-cover classifiers are used as a discriminator to consider land-cover decision boundaries between classes and increase their distance to separate them from the original land-cover class boundaries.The generator then creates robust transferable features far from the original land-cover class boundaries under the classifier constraint.(4)The experimental results of six scenarios(AID ? Merced,AID ? RSI-CB,and Merced ? RSI-CB)built from three benchmark RS scene data sets(AID,Merced,and RSI-CB data sets)are reported and discussed.The experimental results show that this method can creates robust transferable features far from the original land-cover class boundaries under the classifier constraint.The average accuracy of this method for six cross-domain scene data sets is 92.17%,which is 14.75% higher than that of fine-tuned VGG16 without adaptation.The experimental results demonstrate the effectiveness of the proposed method.