Image Based High throughput Phenotyping for Fusarium Wilt Resistance in Pigeon Pea (Cajanus cajan)

• Published in: Phytoparasitica 2022-11, Vol.50 (5), p.1075-1090
• Main Authors: Bannihatti, Rudrappa K., Sinha, Parimal, Raju, Dhandapani, Das, Shubhajyoti, Mandal, S. N., Raje, R. S., Viswanathan, C., Kumar, Sudhir, Gaikwad, K., Aggarwal, R.
• Format: Article
• Language: English
• Subjects: Accuracy; Agriculture; Algorithms; Biomedical and Life Sciences; Canopies; Chlorophyll; Color imagery; Disease resistance; Ecology; Feature extraction; Fluorescence; Genotype & phenotype; Genotypes; Image classification; Imaging techniques; Inoculation; Leaf area; Life Sciences; Machine learning; Neural networks; Original Article; Phenotyping; Photosynthesis; Pigeonpeas; Pixels; Plant Pathology; Plant Sciences; Quantum efficiency; Texture; Variation; Wilt
• ISSN: 0334-2123; 1876-7184
• DOI: 10.1007/s12600-022-00993-5

In pigeonpea, resistance against vascular wilt disease was assessed based on leaf images captured throughred-green–blue (RGB) and chlorophyll fluorescence imaging sensors. At leaf level, wilt response in RGB images was characterized by changes in pixel intensities in red, green, and blue channels leading to variation in texture. Texture analysis based on gray level co-occurrence matrix (GLCM) was able to explain variation pattern between resistance and susceptible genotypes. Extracted texture features particularly contrast and energy were significantly different between the two genotype groups. Training of a neural network model for contrast and energy feature enabled genotype prediction with 79–98% accuracy. Healthy leaf area estimated based on photosynthetic or quantum efficiency (F v /F m  > 0.75 as healthy) in chlorophyll fluorescence images, indicated significant variation ( p   0.75) remained unaffected between 10-25dpi.At canopy level, although differences in pixel intensity (Fv/Fm > 0.75) were noted between inoculated and healthy (mock) particularly in susceptible types but differences between inoculated susceptible and resistant type were non-significant ( p  > 0.05). Although trained ML algorithms for leaf and canopy level images resulted low accuracy (41–54%) in genotype classification but with large number of images captured later than 15 dpi expected to increase in accuracy. A protocol to facilitate non-invasive imaging techniques in association with machine learning tools is proposed over the tedious, time consuming and error-prone conventional screening method.