Classification of soybean groups for grain yield and industrial traits using Vnir-Swir spectroscopy

This research aimed to evaluate the accuracy of machine learning techniques in distinguishing groups soybean genotypes according to grain industrial traits using hyperspectral reflectance of the leaves. A total of 32 soybean genotypes were evaluated and allocated in randomized blocks with four replications. At 60 days after emergence, spectral analysis was carried out on three leaf samples from each plot. The spectral analysis of the leaves was carried out with a hyperspectral sensor providing ranges from 350 to 2500 nm. Once the wavelengths were obtained, they were grouped into averages of representative intervals into bands. At the end of the crop cycle, grain yield was obtained, and subsequently the determination of carbohydrate, oil, and protein content. Initially, the genotypes were subjected to cluster analysis using the k-means algorithm and subsequently, the data was subjected to machine learning analysis, using six models: J48 Decision Trees (J48) and REPTree (DT), Random Forest (RF), Artificial Neural Networks (ANW), Logistic Regression (LR) and Support Vector Machine (SVM). Logistic regression (LR) was used as a reference point as it is a traditional regression algorithm. The clusters formed acted as the output of the models, while for the input of the models, two groups of data were used: the spectral variables (SV) obtained by the sensor (350-2500 nm) and the spectral averages of the bands selected (BS) (350-2200 nm). The use of machine learning techniques presented lower responses than the standard technique used in the work, that is, LR, which presented superiority in the classification of soybean genotypes in terms of industrial traits. The use of wavelengths provided better performance of the algorithms in the classification in relation to selected bands.