Evaluating low-cost hyperspectral and RGB fusion with PSNet for presymptomatic detection of Albugo candida in Arabidopsis thaliana

Plant pathogens cause substantial yield losses worldwide, threatening food security and livelihoods. Because early infection is difficult to diagnose, management often relies on prophylactic pesticide use, increasing costs and environmental impacts. Here we evaluate PSNet, a multimodal framework that fuses hyperspectral and RGB imaging for presymptomatic plant disease detection, together with a low-cost Raspberry Pi hyperspectral camera. We tested the system using Arabidopsis thaliana inoculated with the oomycete Albugo candida in a controlled experiment comprising 96 plants imaged at 2, 4 and 6 days post inoculation (dpi). PSNet achieved 90.00% overall accuracy on the four-class task comprising Healthy, 2 dpi, 4 dpi and 6 dpi. Collapsing the same true and predicted labels into healthy versus inoculated yielded 97.50% binary accuracy across all dates. Supplementary date-specific binary analyses compared inoculated and mock-inoculated plants separately, using plant-disjoint five-fold cross-validation, three training seeds and within-plant probability aggregation. PSNet achieved balanced accuracies of 55.79%, 83.56% and 95.37% at 2, 4 and 6 dpi, respectively. The 4 dpi observations preceded visible symptoms, whereas at 6 dpi abaxial sporulation was present despite the imaged adaxial surface remaining symptom-free. RGB-only models also performed strongly at 4 and 6 dpi, while HSI-only discrimination weakened after 2 dpi, and no consistent advantage of hyperspectral fusion over RGB alone was demonstrated. Whole-cube spectral analysis identified significant treatment-group differences across 269 of 313 spectral positions at 2 dpi, but none at 4 or 6 dpi after within-time-point Benjamini-Hochberg correction. Because treatment was confounded with tray and acquisition order, these spectral differences cannot be attributed specifically to infection. Together, these results support the use of low-cost multimodal imaging for controlled-condition plant disease classification, while highlighting the need for independent validation before diagnostic deployment.