CNN-based binary pollen classification for monofloral honey authentication with conventional light microscopy

Monofloral honeys have distinct organoleptic and physicochemical characteristics depending on the predominant plant species from which their nectar is collected. These honeys command higher market prices than unlabeled blends, as informed consumers show a greater willingness to pay. The standard procedure for authenticity testing is the quantitative analysis of their pollen grains in suspension where minimum percentage thresholds vary according to the plant species involved and applicable jurisdictional legislation. Here, we assessed the capacity of the convolutional neural network (CNN) DenseNet-121 to use binary classification to discriminate and quantify 5 pollen types (Echium plantagineum, Helianthus annuus, Lotus sp., Scutia buxifolia and Trifolium spp. [T. repens + T. pratense]) associated with monofloral honeys in Argentina. Datasets were built with individual pollen grain images extracted from low-magnification bright-field microscopy images from 30 apicultural species, each with at least 3 independent samples. One of the samples was used for testing while the rest were used for training. For each species, the binary classifier was trained with 5 different seeds with specified hyperparameters following a balanced one-vs-rest scheme: each positive taxon against an aggregated class built as an equal mixture of all 29 other classes. Metrics (accuracy, F1, precision, recall, ROC-AUC and PR-AUC) were computed for each classifier. For Lotus, all metrics were above 99% with a near-perfect discrimination. Echium and Scutia were also very well classified with metrics above 95%. Helianthus was detectable with high confidence when the model predicted positive (precision ~99%), but let a significant fraction of positive grains escape (recall 72.4 ± 4.7%). Trifolium showed the worst metrics, in some cases barely above chance for positive detection. Our results show that this classifier is highly pollen-type dependent and it has potential to be integrated into a user-friendly software interface for routine validation of certain monofloral honeys with conventional laboratory equipment.