Semi-supervised bidirectional alignment for Remote Sensing cross-domain scene classification.

Area of Science:

• Computer Vision
• Machine Learning
• Geospatial Analysis

Background:

• Remote sensing (RS) image scene classification is crucial but hindered by the need for extensive labeled data.
• Conventional supervised methods struggle with large volumes of unlabeled RS images and domain shift across datasets.
• Semi-supervised learning (SSL) offers a solution by leveraging unlabeled data, but cross-domain challenges remain.

Purpose of the Study:

• To address the domain shift problem in remote sensing cross-domain scene classification.
• To propose a novel semi-supervised domain adaptation (SSDA) method for effective knowledge transfer between RS datasets.
• To improve the utilization of limited labeled data in target domains.

Main Methods:

• Proposed a novel SSDA method named bidirectional sample-class alignment (BSCA).
• BSCA employs unsupervised alignment (UA) to reduce Maximum Mean Discrepancy (MMD) across domains.
• BSCA utilizes supervised alignment (SA) for sample-to-class center distribution alignment in both source and target domains.

Main Results:

• BSCA demonstrated superior cross-domain classification performance compared to state-of-the-art methods on a benchmark dataset.
• The method achieved compact feature representation and a low-entropy classification boundary.
• Extensive experiments validated the effectiveness of UA and SA components.

Conclusions:

• BSCA effectively mitigates domain shift in remote sensing cross-domain scene classification.
• The proposed method offers a robust solution for leveraging limited labeled data in target domains.
• BSCA advances the field of semi-supervised domain adaptation for geospatial imagery.