Spectral-Structural Collaborative Learning for Fine-Grained Hyperspectral Mineral Classification

Area of Science:

• Geoscience
• Remote Sensing
• Artificial Intelligence

Background:

• Fine-grained hyperspectral mineral classification faces challenges due to spectral homogeneity, mixing, and high dimensionality.
• Existing methods struggle with subtle spectral differences and spatial-structural ambiguity, limiting accuracy.

Purpose of the Study:

• To develop a spectral-structural collaborative feature learning framework for improved hyperspectral mineral classification.
• To address limitations of existing methods in distinguishing spectrally similar minerals with different morphologies.

Main Methods:

• Proposed S3AM-ECA-3DCNN framework utilizing a 3DCNN backbone for joint spectral-spatial feature learning.
• Incorporated spectral-similarity-based spatial attention module (S3AM) for spatial purification and efficient channel attention (ECA) for spectral band recalibration.
• Employed adaptive global pooling and a lightweight classification head to enhance generalization with limited training data.

Main Results:

• Achieved 93.424% overall accuracy, 91.099% average accuracy, and a Kappa coefficient of 93.368 on a 146-class mineral dataset.
• Outperformed mainstream methods in hyperspectral mineral classification tasks.
• Significantly reduced misclassification of spectrally similar but morphologically distinct minerals.

Conclusions:

• The S3AM-ECA-3DCNN framework demonstrates strong robustness and discriminative capability for large-scale fine-grained hyperspectral mineral classification.
• The spatial-first, channel-second collaborative optimization paradigm effectively handles spectral mixing and subtle spectral variations.
• The proposed method offers a significant advancement for mineral identification using hyperspectral data.