- Research Article
6
- 10.1109/jstars.2024.3399741
Cross-Domain Urban Land Use Classification via Scenewise Unsupervised Multisource Domain Adaptation With Transformer
- Jan 01, 2024
- IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Mengmeng Li + 3 more +3
High Resolution (VHR) remote sensing images often suffer 2 from high sample dependency and poor transferability. To 3 address these challenges, we propose an unsupervised multi-4 source domain adaptation framework for cross-domain land use 5 classification that eliminates the need for repeatedly using source 6 domain data. Our method uses the Swin Transformer as the 7 backbone of the source domain model to extract features from 8 multiple source domain samples. The model trained on source 9 domain samples, and unlabeled target domain samples are then 10 used for target domain model training. To minimize the feature 11 discrepancies between the source and target domains, we use 12 a weighted information maximization loss and self-supervised 13 pseudo-labels to alleviate cross-domain classification noise. We 14 conducted experiments on four public scene datasets and four 15 new land use scene datasets created from different VHR images in 16 four Chinese cities. Results show that our method outperformed 17 three existing single-source cross-domain methods (i.e., DANN, 18 DeepCORAL, and DSAN) and four multi-source cross-domain 19 methods (i.e., M3SDA, PTMDA, MFSAN, and SHOT), achieving 20 the highest average classification accuracy and strong stability. 21 We conclude that our method has high potential for practical 22 applications in cross-domain land use classification using VHR 23 images. 24 Index Terms-land use classification, VHR remote sensing 25 images, cross-domain classification, multi-source domain adap-26 tation, Swin Transformer. 27 I. INTRODUCTION 28 A CCURATE mapping of urban land use is essential for 29 urban planning and management, as it reflects the socio-30 economic functions and human activities of cities [1]. Tradi-31 tional mapping techniques are often time-consuming, and can 32 not keep pace with rapid urban development. The advent of 33 Very High Resolution (VHR) remote sensing images, however, 34 has revolutionized this, enabling efficient, large-scale land use 35 mapping [2]. VHR images provide detailed spatial contextual 36 Recently, Unsupervised Domain Adaptation (UDA) methods have attracted significant interest for their potential to improve model performance and reduce data labeling costs when transferring from source (labeled) to target (unlabeled) domains [12], [13], [14]. In remote sensing, UDA-related studies mainly focuses on the single domain closed-set as-
Read more