Deep Manifold Embedding for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning is crucial for hyperspectral image classification.
• Existing methods often overlook intrinsic class structures, relying solely on sample-wise information.
• High spectral dimensions and channel redundancy in hyperspectral images limit traditional deep learning performance.

Purpose of the Study:

• To develop a novel Deep Manifold Embedding Method (DMEM) for hyperspectral image classification.
• To incorporate class-specific information into deep learning training losses.
• To achieve more discriminative representations for hyperspectral images.

Main Methods:

• Modeling each class as a nonlinear manifold with geodesic distance for sample correlation.
• Utilizing hierarchical clustering to capture data manifold structure and divide into subclasses.
• Developing DMEM as a training loss considering subclass distribution and inter-subclass correlation.

Main Results:

• DMEM effectively incorporates class-specific information into the training process.
• The proposed method demonstrates superior performance compared to general sample-based losses.
• Experiments on four real-world datasets confirm the effectiveness and superiority of DMEM over state-of-the-art methods.

Conclusions:

• DMEM offers a significant advancement in hyperspectral image classification by utilizing manifold learning.
• The method enhances deep learning representations by considering intrinsic data structures.
• DMEM provides a more discriminative and effective approach for analyzing hyperspectral imagery.