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

Hyperspectral imaging technology captures fine-grained spectral information from the Earth's surface, offering transformative potential in fields such as environmental monitoring, agriculture, and defense. Hyperspectral unmixing (HU), which decomposes each pixel into pure spectral signatures (endmembers) and quantifies their fractional abundances, enables the detection of target materials by estimating the contribution of each endmember. This study applies hyperspectral unmixing techniques to map and classify invasive alien plant species (Ambrosia trifida, Humulus japonicus, and Sicyos angulatus) in the Geumgang Gomanaru region of South Korea. Endmembers were constructed and used for abundance estimation at the pixel level, overcoming the mixed-pixel problem inherent in hyperspectral data. The experimental pipeline consists of: (1) loading and preprocessing airborne hyperspectral data and endmembers (e.g., Savitzky-Golay filtering), (2) training diverse 1D spectral networks including CNN, CBAM, MLPMixer-1D, SpectralFormer, ViT-1D, and Swin-1D, (3) systematic hyperparameter optimization using Optuna with the proposed spectral loss function (a weighted combination of SAD, SID, and MSE), (4) ablation studies to quantify the individual contributions of attention mechanisms and mixing strategies across CNN, Mixer, and Transformer architectures, (5) model comparison through conventional and spectral metrics (SAD, SAM, SID, RMSE, PSNR) combined with a custom weighted evaluation, and (6) interpretability analysis using SHAP and model assessment. By combining airborne hyperspectral imagery with advanced 1D deep learning architectures-CNN, CBAM, MLPMixer-1D, SpectralFormer, ViT-1D, and Swin-1D-this study captures critical spectral variations across the visible, NIR, and SWIR regions reflecting vegetation physiology and structure. MLPMixer-1D achieved the highest performance through effective inter-band channel mixing. Endmember-specific analysis showed water and grassland were most accurately mapped, whereas invasive species exhibited moderate accuracy due to spectral overlap. SHAP analysis identified UV-A, red-edge, and SWIR bands as most informative for invasive species discrimination, highlighting the ecological and physiological relevance of the spectral features. Overall, integrating hyperspectral sensing with tailored deep learning models offers a powerful framework for resolving mixed pixels and supporting large-scale ecological monitoring of invasive plant species. Our data and code are available at https://drive.google.com/file/d/111g4Sbt7NcmYBcUVK_cbiPSt3h1meLHs/view?usp=drive_link.