Assisting monofloral honey classification by automated pollen identification based on convolutional neural networks
Categorising monoflorality of honey by human experts is labour intensive and time-consuming, resulting in the urgent need for automated techniques to overcome the constraints of traditional method which involves manually sorting of pollen grains. This study constructs a comprehensive dataset of poll...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125003498 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Categorising monoflorality of honey by human experts is labour intensive and time-consuming, resulting in the urgent need for automated techniques to overcome the constraints of traditional method which involves manually sorting of pollen grains. This study constructs a comprehensive dataset of pollen images, labelling them with an enhanced version of HoneyApp application. This Ground Truth termed POLLEN24_SP, comprises 32,285 pollen/particle images (captured by an expert using optical microscopy), covering the 24 most prevalent types of pollen grains found in Spanish honeys. Twelve different pre-existing Convolutional Neural Networks (CNN) were evaluated, achieving an accuracy rate of up to 98.03 % with EfficientNetV2M. The new version of in-house networks (PolleNetV2, PolleNetV2.mobile) was also introduced. The pre-existing CNN models' results were satisfactory in discerning the relative frequencies of different pollen types; nevertheless, our new proposals are promising due to their straightforward architecture as well as their scalability for integration into the HoneyApp ecosystem. Studying the classification errors made by the CNN models confirmed their strong generalization ability. To minimise errors in types such as Leguminoseae, segmenting this family is recommended. In addition, it is vital to continue expanding POLLEN24_SP dataset focusing on the underrepresented types to balance them and include more pollen types to better represent Spanish honeys. Finally, the results of this work will enable the creation of a public dataset to be diffused among the scientific community, following open science principles. The development of a computer application incorporating CNN models, makes these methodologies suitable for efficient classification of monofloral honeys. |
|---|---|
| ISSN: | 1574-9541 |