Multimodal RGB and optical coherence tomography imaging for enhanced machine learning classification of crop diseases and stress phenotyping

Abstract

Plant diseases and abiotic stressors threaten global food security, necessitating rapid, noninvasive detection methods. Current single-modality imaging techniques, such as RGB (spectral) or Optical Coherence Tomography (OCT; structural), face limitations in misclassification and incomplete feature extraction. This study proposes a multimodal fusion approach integrating OCT and RGB imaging to enhance machine learning (ML)-based classification of rice leaf health. OCT provided high-resolution cross-sectional microstructural data (axial resolution: ∼4 µm, depth range: ∼3.2 mm), revealing statistically significant differences in optical attenuation coefficients (healthy: 3.65 ± 0.36 mm−1, dry: 4.75 ± 0.4 mm−1, diseased: 5.49 ± 0.52 mm−1; p < 0.0001) and leaf thickness (healthy: 437.4 ± 38.32 µm, dry: 331.4 ± 21.83 µm, diseased: 266 ± 27.91 µm). RGB imaging captured spectral reflectance variations across visible wavelengths (480–650 nm), achieving 96.7% classification accuracy. Neural networks trained on combined OCT-RGB data achieved 99.6% accuracy, outperforming single-modality models (OCT: 88.9%, RGB: 96.7%). The fusion strategy leveraged complementary features: OCT quantified subsurface structural degradation, while RGB detected surface-level biochemical changes. This dual-modality approach demonstrates superior diagnostic precision, offering a scalable, noninvasive solution for early disease detection and stress phenotyping in precision agriculture.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.