Triple-path bidirectional Mamba enables linear-complexity temporal hyperspectral plant stress classification
Abstract
Temporal hyperspectral imaging enables non-destructive monitoring of agricultural stress through spectral signatures that evolve over extended observation periods, yet processing high-dimensional spatial-spectral-temporal sequences remains computationally demanding. Traditional machine learning methods compress temporal information through dimensionality reduction, while hybrid deep learning architectures combining convolutional and recurrent networks face optimization difficulties at module boundaries. We introduce PSUMamba, a triple-path bidirectional Mamba architecture that processes 204-band hyperspectral sequences across eight timepoints through linear-complexity state space models. The architecture separates spectral biochemical correlations, temporal physiological degradation, and multi-scale spatial structure into three dedicated pathways combined through cross-path self-attention, learnable softmax weights, and a logit ensemble. Evaluated under stratified source-based splitting across five random seeds, PSUMamba achieves 94.47% ± 2.17% accuracy, 99.36% ± 0.40% AUC-ROC, and 98.98% ± 1.25% sensitivity using 2.35 million parameters. Compared with Vision Transformer, PSUMamba uses approximately 2.5× fewer parameters and 10× fewer FLOPs while reaching accuracy that is at least comparable under the present evaluation, and achieves statistically significant improvements over PLS-DA, 3D CNN, 1D CNN-LSTM, and the recent MambaHSI state-space baseline with large effect sizes. Ablation experiments under matched block configurations show that the spectral, temporal, and spatial pathways encode complementary information, and that the full triple-path configuration improves over the strongest single pathway at the same parameter count. The learned fusion weights converge to a near-balanced distribution across the three pathways, indicating that each pathway contributes meaningfully to stress detection. These results position PSUMamba as a computationally efficient alternative to transformer-based models for temporal hyperspectral classification in agricultural monitoring.
Versions
Related
- CSwinFK: a bidirectional cross-attention CNN-Swin Transformer fusion model for fungal keratitis classification in in vivo confocal microscopy images
- 787 Synapse-localized γ-chain cytokine presentation enables single-dose, soluble cytokine–free T cell expansion in plate and G-Rex bioreactor formats with reduced manufacturing complexity and cost
- Editorial: Unveiling the complexity of neuroendocrine neoplasms: biomarkers, classification, and translational perspectives
- KT-CBHNET: a spatio-temporal deep learning framework for interpretable ECG signal classification for cardiovascular disease prediction
- Temporal imprecision of the self in drug-free psychosis: Abnormal infraslow phase dynamics and brain complexity
BibTeX
@article{wang2026triplepath,
title = {{Triple-path bidirectional Mamba enables linear-complexity temporal hyperspectral plant stress classification}},
author = {Weiqun Wang and Shirin Ghatrehsamani},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-74809-5},
url = {https://www.nature.com/articles/s41598-026-74809-5},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Triple-path bidirectional Mamba enables linear-complexity temporal hyperspectral plant stress classification
AU - Weiqun Wang
AU - Shirin Ghatrehsamani
PY - 2026
DA - 2026/10/07
JO - Scientific Reports
DO - 10.1038/s41598-026-74809-5
UR - https://www.nature.com/articles/s41598-026-74809-5
AB - Temporal hyperspectral imaging enables non-destructive monitoring of agricultural stress through spectral signatures that evolve over extended observation periods, yet processing high-dimensional spatial-spectral-temporal sequences remains computationally demanding. Traditional machine learning methods compress temporal information through dimensionality reduction, while hybrid deep learning architectures combining convolutional and recurrent networks face optimization difficulties at module boundaries. We introduce PSUMamba, a triple-path bidirectional Mamba architecture that processes 204-band hyperspectral sequences across eight timepoints through linear-complexity state space models. The architecture separates spectral biochemical correlations, temporal physiological degradation, and multi-scale spatial structure into three dedicated pathways combined through cross-path self-attention, learnable softmax weights, and a logit ensemble. Evaluated under stratified source-based splitting across five random seeds, PSUMamba achieves 94.47% ± 2.17% accuracy, 99.36% ± 0.40% AUC-ROC, and 98.98% ± 1.25% sensitivity using 2.35 million parameters. Compared with Vision Transformer, PSUMamba uses approximately 2.5× fewer parameters and 10× fewer FLOPs while reaching accuracy that is at least comparable under the present evaluation, and achieves statistically significant improvements over PLS-DA, 3D CNN, 1D CNN-LSTM, and the recent MambaHSI state-space baseline with large effect sizes. Ablation experiments under matched block configurations show that the spectral, temporal, and spatial pathways encode complementary information, and that the full triple-path configuration improves over the strongest single pathway at the same parameter count. The learned fusion weights converge to a near-balanced distribution across the three pathways, indicating that each pathway contributes meaningfully to stress detection. These results position PSUMamba as a computationally efficient alternative to transformer-based models for temporal hyperspectral classification in agricultural monitoring.
SN - 2045-2322
ER -