Learnable manifold alignment (LeMA): A semi-supervised cross-modality learning framework for land cover and land use...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-modality feature learning is crucial in remote sensing.
• Limited hyperspectral data and abundant multispectral data present a challenge for classification tasks.
• Traditional semi-supervised manifold alignment methods struggle with this data imbalance.

Purpose of the Study:

• To develop a novel semi-supervised cross-modality learning framework to leverage limited hyperspectral data for improving multispectral data classification.
• To address the limitations of existing methods in handling the cost and efficiency trade-offs associated with hyperspectral data collection.

Main Methods:

• Proposed a novel framework named Learnable Manifold Alignment (LeMA).
• LeMA learns a joint graph structure directly from data, unlike fixed graph methods.
• Employed graph-based label propagation for accurate data distribution capture and decision boundary identification.
• Utilized an alternating direction method of multipliers (ADMM) for optimization.

Main Results:

• Demonstrated superior performance on hyperspectral-multispectral datasets.
• Showcased the effectiveness of LeMA in cross-modality feature learning.
• Achieved more accurate classification boundaries compared to state-of-the-art methods.

Conclusions:

• LeMA offers a significant advancement in semi-supervised cross-modality learning for remote sensing.
• The method effectively utilizes limited high-quality data to enhance analysis of larger, lower-quality datasets.
• LeMA provides a robust and effective solution for remote sensing image classification challenges.