3D-CNN detection of systemic symptoms induced by different Potexvirus infections in four Nicotiana benthamiana genotypes using leaf hyperspectral imaging

The proposed patch-based method demonstrated key advantages: (a) exploiting both spectral and textural information, (b) deriving a large training dataset from few hyperspectral images, (c) providing localized classification explainability within leaf regions, and (d) achieving high accuracy for early detection of viral infections.

Four genotypes of Nicotiana benthamiana L. (wild-type, DCL2/4, AGO2, and NahG) were inoculated with different potexviruses (PepMV mild or severe strain, PVX, BaMV). Viral infection was verified via northern blot analysis at 5 and 10 days post inoculation (DPI). Hyperspectral images were captured over 10 days following inoculation, focusing on the top 3 leaves where symptoms typically appear. The dataset was carefully processed to remove errors, and raster masks were generated to isolate only the leaf pixels. The Extremely Randomized Trees algorithm was used for Effective Wavelength selection, and a novel 3D-CNN architecture was developed to classify 16×16×1616×16×16<math><mrow><mn>16</mn> <mo>×</mo> <mn>16</mn> <mo>×</mo> <mn>16</mn></mrow> </math> nonoverlapping cubes extracted from the unmasked leaf surfaces. The aim was to classify each cube into healthy or diseased for each of the four viruses at different time points.

Hyperspectral imaging combined with machine learning offers a promising, cost-effective alternative to invasive chemical analysis for early plant disease detection. In this study, the use of 3D Convolutional Neural Networks (3D-CNNs) was explored to detect presymptomatic viral infections in the model plant Nicotiana benthamiana L. and assess the generalization of these models across different plant genotypes.

Accuracies of 0.780.78<math><mrow><mn>0.78</mn></mrow> </math> - 0.870.87<math><mrow><mn>0.87</mn></mrow> </math> were achieved for AGO2 mutants at the cube level, and overall plant-level accuracies of 0.680.68<math><mrow><mn>0.68</mn></mrow> </math> - 0.890.89<math><mrow><mn>0.89</mn></mrow> </math> . The model's generalization capabilities were tested across other genotypes, yielding accuracies of up to 0.750.75<math><mrow><mn>0.75</mn></mrow> </math> for DCL2/4, 0.830.83<math><mrow><mn>0.83</mn></mrow> </math> for NahG, and 0.780.78<math><mrow><mn>0.78</mn></mrow> </math> for the wild-type. The timing of disease detection was also assessed, finding that accuracies approached 0.8 as early as 66<math><mrow><mn>6</mn></mrow> </math> - 88<math><mrow><mn>8</mn></mrow> </math> DPI depending on the virus. The results were validated against northern blot analyses and benchmarked against another state-of-the-art methodology for Nicotiana benthamiana viral infections, achieving superior overall classification accuracies.