Digital image processing combined with machine learning: A novel approach for bee pollen classification

The classification of bee pollen is crucial for ensuring product authenticity, quality control, and fraud prevention, particularly given the high commercial value of stingless bee pot-pollen. Although traditional pollen analysis methods are available, they are often time-consuming and complex. To address this challenge, this study integrates digital image processing and machine learning to efficiently classify pollen loads produced by Apis mellifera and pot-pollen from stingless bee species based on their color patterns. A total of 246 pollen loads and pot-pollen samples from five bee species (Apis mellifera, Melipona marginata, Melipona quadrifasciata quadrifasciata, Scaptotrigona bipunctata, and Tetragona clavipes) were collected, and high-resolution images were captured using a smartphone. Color parameters extracted from images such as R, B, H, and V were analyzed, and classification models employing CatBoost, XGBoost, Random Forest, and k-Nearest Neighbors (kNN) algorithms were tested. Among these models, CatBoost achieved the highest performance, with accuracies of 100 % in the training phase and 98.9 % in the testing phase. Linear Discriminant Analysis (LDA) further validated the classification by grouping the pollen samples into five distinct clusters corresponding to the bee species studied. XGBoost and Random Forest achieved testing accuracies of 77 % and 78 %, respectively, while kNN also reached 78 %. These results demonstrate that combining digital image processing with machine learning offers an effective approach to classifying pollen from different bee species, with promising applications in apiculture to ensure product quality and authenticity.