Comprehensive study on the performance optimization of hyperspectral unmixing algorithms: A focus on airborne...

Area of Science:

• Remote Sensing
• Ecology
• Computer Science

Background:

• Hyperspectral imaging provides detailed spectral data for Earth observation.
• Hyperspectral unmixing (HU) resolves mixed pixels into endmembers and abundances.
• Invasive alien plant species pose significant ecological threats.

Purpose of the Study:

• To apply hyperspectral unmixing and deep learning for mapping invasive alien plant species (Ambrosia trifida, Humulus japonicus, Sicyos angulatus).
• To evaluate various 1D deep learning architectures for hyperspectral unmixing.
• To identify informative spectral bands for invasive species discrimination.

Main Methods:

• Airborne hyperspectral data and endmembers were preprocessed.
• 1D deep learning models (CNN, CBAM, MLPMixer-1D, SpectralFormer, ViT-1D, Swin-1D) were trained.
• Hyperparameter optimization used Optuna with a novel spectral loss function.
• Ablation studies and SHAP analysis assessed model components and feature importance.

Main Results:

• MLPMixer-1D demonstrated superior performance in hyperspectral unmixing.
• Water and grassland mapping were highly accurate; invasive species mapping showed moderate accuracy due to spectral overlap.
• SHAP analysis highlighted UV-A, red-edge, and SWIR bands as critical for invasive species identification.

Conclusions:

• Integrating hyperspectral sensing with deep learning models effectively resolves mixed pixels.
• The study provides a powerful framework for large-scale ecological monitoring of invasive plant species.
• The findings underscore the importance of specific spectral regions for distinguishing invasive flora.