Class-Wise Subspace Alignment-Based Unsupervised Adaptive Land Cover Classification in Scene-Level Using Deep Siamese...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Scene-level land cover classification from remotely sensed images faces challenges due to domain shift.
• Unsupervised domain adaptation is crucial for leveraging labeled source data with unlabeled target data.

Purpose of the Study:

• To develop an unsupervised domain adaptation strategy for scene-level land cover classification.
• To improve the performance of deep learning models in classifying land cover across different remote sensing datasets.

Main Methods:

• Utilized a deep Siamese neural network for unsupervised domain adaptation.
• Employed a pretrained deep convolutional neural network to generate soft class labels and probability scores for target samples.
• Implemented a semiautomatic threshold selection algorithm and a graph-based approach to identify confident target samples.
• Trained the Siamese network with source and confident target samples to create a common subspace.

Main Results:

• The proposed framework demonstrated encouraging results in land cover classification.
• The method showed improved performance compared to existing state-of-the-art techniques.
• Experiments were validated using three distinct aerial image datasets.

Conclusions:

• The unsupervised domain adaptation strategy using a deep Siamese network is effective for scene-level land cover classification.
• The approach successfully bridges the domain gap between different remote sensing image datasets.
• This method offers a promising solution for accurate land cover mapping without extensive target domain labeling.