Kohonen self-organizing maps to identify cercosporiosis and orange rust in coffee crops through digital imaging

This work uses images to compare different computational intelligence techniques to identify rust (Hemileia vastatrix) and cercosporiosis (Cercospora coffeicola) in Arabica and Conilon coffee leaves. A total of 428 images with different severities of rust (Hemileia vastatrix) and cercosporiosis (Cercospora coffeicola) were obtained. Each image generated information from 32 biometric traits of the leaves used to identify diseases through Kohonen's self-organizing map (SOM), principal component analysis (PCA), and K-means. The first PCA explained 65.20% of the variability, indicating that most of the data variance was concentrated in the first components. PCA allowed images to be classified into two distinct groups according to rust and cercosporiosis symptoms. Using K-means, it was possible to observe two distinct groups, in which the closer the points are to the center of the cluster, the smaller the dispersion, which represents greater homogeneity within the cluster. The SOM method identified five distinct groups, showing a greater ability to detect similarity patterns between images than PCA. The filling of neurons in the map indicated a greater concentration of images in certain groups. The SOM technique preserved spatial relationships between data and captured nonlinear characteristics, being more efficient than PCA in classifying images. Among the different techniques evaluated, Kohonen self-organizing maps can be used as an auxiliary tool in genetic improvement programs to identify rust and cercosporiosis symptoms in coffee leaves through digital images.