Identification and Classification of Snack-Type Watermelon (Citrullus lanatus) Genotypes Using Seed Morphology and Machine Learning Techniques

This study examines the effectiveness of machine learning approaches for the automatic identification of watermelon genotypes from the seeds of watermelon, for the snack-type watermelon (Citrullus lanatus). Nine genotypes with red, white, and black seed coats were assessed in total. For each genotype, 200 seeds were analyzed using high-resolution imaging and digital measurement techniques for the extraction of morphological characteristics (length, width, thickness, area, perimeter, equivalent diameter, etc., and physical (weight) and colorimetric attributes of the (L, a, b). The resulting dataset was modeled using Artificial Neural Network (ANN), Random Forest (RF) and Extra Tree (ET) algorithms and performance was validated by a 10-fold cross-validation. The primary objective of the study was to match (identify) each seed accurately with its respective genotype by using the morphological, physical, and colorimetric characteristics of the seed and thus to perform genotypic classification. The comparative results showed that the RF model had the highest genotypic performance (accuracy 92.22%, F1-score 91.87%, Cohen’s Kappa 0.9118), followed by the ET (accuracy, 90.00%) and ANN models with a relatively lower precision (86.11%). Statistical analysis using the Wilcoxon signed-rank test confirmed that both RF and ET significantly outperformed ANN, with RF providing superior balance and stability over ET. The findings highlight that machine learning-based frameworks enable rapid, reliable, and non-destructive classification (identification) of snack-type watermelon seeds according to their genotypes. Such approaches hold strong potential for enhancing varietal traceability in breeding programs, improving quality control in commercial seed production, and meeting the high-throughput demands of seed processing industries.