Classification of soybean genotypes based on physiological clustering (PCA + k-means) integrated with VIS-NIR hyperspectral data and machine learning models

• Integration of physiological and hyperspectral data enables rapid and non-destructive classification of soybean genotypes, supporting high-throughput phenotyping in breeding programs. • Machine learning algorithms—especially Random Forest and Support Vector Machine—achieved superior accuracy in discriminating soybean genotypes using spectral and physiological traits. • Spectral signatures and physiological traits revealed genotype-specific responses to environmental conditions, providing robust tools for precision selection and sustainable soybean improvement. Given the wide variety of genotypes, classifying them based on physiological traits can aid breeding efforts. This study aimed to identify the most effective machine learning algorithms and input variables for classifying soybean genotypes according to physiological features. A field experiment was conducted using a random complete block design with three replications, testing 32 soybean genotypes. Multispectral and hyperspectral reflectance data were collected and grouped into 20 representative wavelength ranges. The physiological traits measured included net photosynthesis, internal CO 2 concentration, stomatal conductance, and transpiration rates. Genotype groups based on physiological performance were classified with the k-means algorithm and principal component analysis, resulting in two clusters. K-means is an algorithm that groups data into k similar groups. It does this by finding patterns and placing each sample in the group whose center (centroid) it is closest to. These clusters served as output variables in machine learning models, with input variables including wavelengths and spectral band averages. A hyperspectral sensor, was employed to record leaf reflectance at wavelengths ranging from 450 to 824 nm in the laboratory under artificial lighting. The algorithms tested were artificial neural networks, J48 decision trees, REPTree, support vector machines (SVM), random forest, and logistic regression (LR) as a control. Model accuracy was assessed using correct classification percentage and F-score. SVM and RL stood out with accuracies above 0.6 in classifications for CC and an F-score exceeding 0.75. When using spectral bands as predictors, both showed similar performance, but with wavelengths as predictors, WL becomes a robust input for the models because its complete spectrum information provides important data for group classifications.