Class-independent domain adaptation for hyperspectral image classification
Hyperspectral image supervised classification is a crucial and challenging task in remote sensing,as its performance depends heavily on the quantity and quality of labeled samples.However,labeling hyperspectral data is a difficult and time-consuming procedure.This problem results in a limited number of labeled samples in real-world scenarios,rendering the supervised classifiers vulnerable to the issue of overfitting.To address this problem,researchers have sought solutions in the field of unsupervised domain adaptation,utilizing labeled samples from previous images(source domain)to classify new hyperspectral data(target domain).Most existing domain adaptation methods strive to learn domain-invariant features in a new space,but many of them focus on aligning the overall statistics of the two domains without considering the spectral shifts in each class.Other methods attempt to align every class of the source and target domains simultaneously but often overlook the issues of mixture of samples across classes and incorrect sample selection.This may lead to a negative transfer and reduced separability of data.The significant discrepancies across domains will further compound the problem.In this paper,we propose a novel class-independent domain adaptation algorithm that addresses these issues in hyperspectral image classification.Our method first creates an independent subspace for each class and then aligns the corresponding single-class samples of the two domains in those subspaces.The posterior probabilities are learned independently through the aligned samples in each subspace.Then,the posterior probabilities obtained from multiple subspaces are fused to produce the final classification labels,aiming at increasing the confidence of results.Additionally,we use smoothed classification labels as pseudo labels for further iteration and incorporate a strategy for selecting representative samples to enhance subspace performance.Experimental results on two real hyperspectral datasets demonstrate the high classification performance of our proposed method.Compared to the joint domain adaptation algorithm,our method with the nearest neighbor classifier improved the accuracy by 9.56％on the Honghu data and 18.45％on the Wen-County data.Compared to other competitors,our method also has the advantage of generating smoother classification maps with more distinct boundaries of ground objects.These remarkable results stem from the substantial improvement in data separability achieved by our approach,which has been validated through calculations in our experiments.In conclusion,our proposed class-independent domain adaptation algorithm is a promising solution for hyperspectral image classification,providing high performance with reduced risk of overfitting.By aligning samples in the class-independent subspaces and fusing posterior probabilities,our method leads to improved data separability and more accurate classifications.Furthermore,the use of representative sample selection helps mitigate the potential impact of mislabeled samples on class alignments.Thus,our algorithm is able to overcome the limitations of existing domain adaptation methods and achieve improved results.In future work,we plan to extend our method to more complex high-dimensional datasets and incorporate advanced deep learning models.We also intend to evaluate the applicability of our method for real-time hyperspectral image classification,which is a critical requirement for many remote sensing applications.Overall,our research represents a significant advancement in the field of hyperspectral image classification,offering a new approach for solving the challenge of insufficient labeled samples.
NETL
NSTL
万方数据
