PSUMamba: Dual-Path Bidirectional Mamba for Plant Stress Monitoring via Temporal Hyperspectral Imaging

Temporal hyperspectral imaging enables non-destructive monitoring of agricultural stress through spectral signatures evolving across extended observation periods, yet processing high-dimensional spatial-spectral-temporal sequences remains computationally prohibitive for real-time deployment. Traditional machine learning methods sacrifice temporal information through dimensionality reduction, while hybrid deep learning architectures combining convolutional and recurrent networks suffer from optimization pathologies at component boundaries. We introduce PSUMamba, a dual-path bidirectional Mamba architecture that processes 204-band hyperspectral sequences across eight timepoints through linear-complexity state space models, achieving 95.05% accuracy with 99.00% AUC-ROC using 153,268 parameters. The architecture maintains perfect specificity (100%) with 93.67% sensitivity while out-performing Vision Transformer with 39-fold fewer parameters and 37.5% reduced training time. Separate spectral and temporal pathways with adaptive fusion enable specialized biochemical and physiological feature extraction without quadratic attention overhead. Ablation studies confirm temporal features dominate classification under experimental conditions, with dual-path fusion providing superior probabilistic calibration (97.35% AUC) over single-path variants. Statistical comparisons demonstrate significant improvements over PLS-DA (∆=12.07%, p=0.0001), 3D CNN (∆=15.48%, p=0.0042) and 1D CNN-LSTM (∆=33.77%, p < 0.0001). The results establish PSUMamba as efficient alternatives to transformers for temporal hyperspectral classification in agricultural stress monitoring, with the dual-path framework providing a generalizable template for multimodal temporal problems requiring linear-complexity long-range dependency modeling.