Domain Adaptation For High Resolution Remote Sensing Image Classification

With the development of aerospace technology,a large amount of high-resolution ground observation data has become easy to obtain.How to accurately interpret these data as useful information under the condition of limited manpower and material resources is one of the key points for the development of geographic information industry in China.In the task of remote sensing information interpretation,traditional supervised methods need some high-quality labels from the image to be interpreted(target domain)in advance.It is not only time-consuming and laborious,but also impossible to be real-time.In practical applications,a large number of labeled remote sensing data(source domain)are easily available without any additional cost.The main research content of this thesis is how to make full use of the source domain information to classify the target domain which does not have any label.Most existing machine learning algorithms assume that the training and testing datasets share the same feature distribution.However,there are significant differences between the source and target domains since the different data acquisition methods,atmospheric conditions,object patterns,building styles,etc.These differences make the joint probability density distribution of the source and target domains significantly different.Such a difference in feature distribution makes the assumptions in the above machine learning algorithm difficult to hold,resulting in the classifier directly trained by the source domain cannot perform well in the target domain.To reduce the demand on time-consuming and labor-intensive manual labeling work,how to effectively minimize the differences between the source and target domains has become the core technical bottleneck.By addressing this issue,Domain Adaptation(DA)algorithm has shown some ability to improve the performance of the classifier trained by the source domain on the target domain data and has become an important method to reduce the cost of labeling data.This thesis conducts in-depth research on domain adaptation algorithms for different remote sensing image classification application scenarios.To improve the classification ability of the remote sensing image and scene of the target domain without relying on any target domain labels,this thesis proposes different domain adaptation algorithms according to the difference between the source and target remote sensing images and the amount of data.The reminder of this thesis is summarized as the following four folds:The first part studies the domain adaptation algorithm for classification when the source domain and target domain remote sensing images come from the same area.The remote sensing images in the same area often have small differences and strong correlations.This thesis aims to use the domain adaptation algorithm with less calculation and less source domain data to classify the target domain image from the same area as the source domain.This type of application contains two specific scenarios: the source and target domains are remote sensing images at the same location and different time phases or remote sensing images at different locations in the same area.First,for the premise that the images in the source and target domains come from the same area,this thesis assumes that the difference between the two domains is relatively small,and it is possible to generate relatively accurate pseudo-labels to estimate the conditional probability density distribution of the target domain data.Based on this,a novel DA method named dispersion-optimized joint distribution adaptation is proposed.Based on the proposed algorithm,this thesis further proposes two transfer learning frameworks for two specific application scenarios: multi-temporal remote sensing images and remote sensing images of adjacent locations in the same area.Finally,experiments show the effectiveness of the algorithm.The second part studies the domain adaptation algorithm for classification when the source domain and target domain are multi-modal remote sensing data in different regions,aiming at the situation where the source and target domains have large differences and the amount of data is small,but auxiliary information such as elevation information can be obtained.Here,the multi-modal data is used to improve the classification accuracy of the target domain data.Considering this situation alone,on the one hand,although the domain adaptation performance of the non-deep learning algorithm is not as good as the deep learning algorithm,but its calculation amount is small and the data volume requirement is low,so it still has full research significance.On the other hand,the feature extract ability of non-deep learning algorithms and the ability to fit complex domain adaptation mappings are relatively weak.To achieve a more ideal classification effect,this part mainly studies multi-modal data which with more information,and proposes a multi-kernel joint domain matching algorithm.The algorithm utilizes the multi-kernel learning method and instance reweighting to align the distributions between the source and target domains.The proposed algorithm realizes transfer learning between multi-modal remote sensing images in different regions.At the same time,in single-modal scenarios,the multi-core joint domain adaptation algorithm is still effective,but the improved classification accuracy is still low.Even so,in the case where the amount of data or computing power is not sufficient to support deep learning training,the algorithm proposed in this section is still an effective solution for transfer learning between remote sensing images in different regions.The third part studies the domain adaptation algorithm for semantic segmentation when the source domain and target domain are single-modal(three-channel)remote sensing images in different regions.The purpose of this part is that when the source and target domains are quite different and only three-channel optical images are available,the domain adaptation method is still able to greatly improve the semantic segmentation performance of the target domain images.For such scenarios,this thesis proposes a bi-space alignment network domain adaptation algorithm.Based on the basic idea of adversarial learning,this paper proposes a bi-space adversarial learning strategy while reducing the difference between the source domain and the target domain in the feature space and output structure space.In addition,the proposed method is able to simultaneously align the feature distribution of the source domain and the target domain in the image representation and the wavelet representation.Experiments show that this strategy is able to further improve the domain adaptation performance of the network.The fourth part studies the domain adaptation algorithm for remote sensing scene classification,aiming to utilize the existing remote sensing scene classification dataset to effectively classify scenes that are different from it.Since scene classification and semantic segmentation are different,this thesis proposes an adversarial domain adaptation network for scene classification.The proposed network is able to fully utilize the label information of the source domain and reduce the difference between the source and target domains based on the idea of generative adversarial network with an auxiliary classifier.At the same time,this thesis proposes a new concept,domain confused network.Through the domain confused network,the proposed method is able to further improve the classification accuracy of remote sensing scenes in the target domain.