Leaf and Seed Hyperspectral Signatures Enable Early and Accurate Prediction of Soybean Seed Quality

High-quality soybean seeds possess genetic, physical, and physiological characteristics that directly influence crop yield. The use of hyperspectral sensors combined with machine learning (ML) can streamline and accelerate seed germination testing. Therefore, the objectives of this study were: (i) to evaluate whether leaf and seed reflectance can effectively predict the physiological quality of soybean seeds using ML algorithms, and (ii) to identify which algorithm provides the highest prediction accuracy. Thirty-two soybean genotypes were evaluated in a controlled experiment. Leaves and seeds were analyzed using a hyperspectral sensor capable of measuring reflectance across the 350 to 2500 nm range. The resulting data were subjected to ML analysis with two types of input: spectral variables from leaves and seeds. The output variables predicted included germination test (GERM), electrical conductivity (EC), first germination count (FGC), vigorous tetrazolium test (VIG-TZ), and viable tetrazolium test (VIAB). Predictions were performed using stratified 10-fold cross-validation with ten repetitions (100 runs per model). All model parameters were set to the default configuration in Weka version 3.8.5. The ML models used for prediction included artificial neural networks (ANN), REPTree and M5P decision trees, random forest (RF), support vector machine (SVM), and ZeroR, with the latter serving as a control algorithm. The models showed consistent performance in predicting physiological variations in seeds, with better results when seed reflectance was used as input. For germination (GERM), the M5P, RF, and SVM algorithms obtained the highest correlations (r = 0.565–0.575). In predicting electrical conductivity (EC), M5P showed greater accuracy with leaf data (r = 0.506), while SVM performed best with seed data (r = 0.658). For first germination count (CPG), M5P was the most accurate with leaf data (r = 0.720), while M5P, RF, and SVM showed r between approximately 0.735 and 0.777 with seed data. In tetrazolium vigor (TZVG), RF showed the best performance (MAE 0.25), again highlighting seed reflection, which resulted in the lowest errors and highest correlations. Overall, the M5P, RF, and SVM algorithms achieved the most robust results, especially when used with seed spectral data. The highest germination prediction accuracy was achieved by the M5P, SVM, and RF algorithms for both input types. Seed reflectance yielded the best accuracy and the lowest MAE and RMSE values. Leaf reflectance also enabled accurate predictions, indicating that this input can serve as an early, in-field strategy for predicting soybean seed physiological quality.