Constructing a pollination network by identifying pollen on insect bodies: Consistency between human recognition and an AI model

doi:10.17520/biods.2024088

Abstract

Abstract:

Aims: Pollination plays a crucial role within ecosystems. Accurate identification and analysis of the pollen loads carried by pollinators are essential for understanding the plant‒pollinator interactions and assessing pollination activity. Traditional pollen identification methods, which are often reliant on microscopic observation, are time-consuming and require specialized expertise, limiting their utility in large-scale applications and in the assessment of pollination efficiency and rare plant‒pollinator interactions. To address this issue, we developed an Artificial Intelligence (AI) pollen identification model using a public platform.
Method: The AI model was based on pollen from 14 co-flowering plant species from Tianchi Mountain National Forest Park in Luoyang, China. We identified the pollen composition of 142 pollinators using both the traditional microscopic observation method and the AI model to explore and compare for the first time the structural differences in the plant‒pollinator interaction networks constructed by the two identification methods.
Results: The results demonstrated that the AI model achieved an overall accuracy rate of 96%. While there were differences between human recognition and the AI model in the number of identified links, quantity of pollen, and consistency rate of photo identification, both methods showed a high degree of consistency. The AI model slightly outperformed human methods in link identification (6.5%) and pollen quantity (0.8%). Moreover, in third-party consistency checks, the majority of the cases favored the results from the AI model. Despite some differences regarding unique links, the quantitative networks constructed by human recognition and the AI model showed a high degree of structural similarity.
Conclusion: This study reveals the potential of AI image recognition technology to enhance the efficiency and accuracy of pollen analysis, and its relevance to plant‒pollinator interaction research. This advancement could facilitate a more efficient advancement of large-scale studies of pollination networks, providing new tools and perspectives for pollination ecology research.

Key words: AI model, pollen identification, image recognition, plant-pollinator interaction, pollinator pollen load

Suyan Ba, Chunyan Zhao, Yuan Liu, Qiang Fang. Constructing a pollination network by identifying pollen on insect bodies: Consistency between human recognition and an AI model[J]. Biodiv Sci, 2024, 32(6): 24088.

Add to citation manager EndNote|Ris|BibTeX

URL: https://www.biodiversity-science.net/EN/10.17520/biods.2024088

https://www.biodiversity-science.net/EN/Y2024/V32/I6/24088

Figures/Tables 7

Table 1 Comparison of the training results of AI models for image segmentation and object detection based on the pollen identification bank of Tianchi Mountain. mAP (mean Average Precision) measures the average precision across all categories, with a range from 0 to 1, higher values indicate better model performance. Precision is the ratio of correctly predicted objects to the total number of predicted objects. Recall is the ratio of correctly predicted objects to the total number of actual objects.

模型选择 Model selection	平均精度均值 mAP	精确率 Precision	召回率 Recall
图像分割 Image segmentation	95.52%	91.52%	91.45%
物体检测 Object detection	99.25%	97.20%	93.92%

Fig. 1 AI model training effect. (a) Model accuracy under different thresholds, the highest accuracy of the model is 96% when the threshold is 0.8. F1-score refers to the harmonic average of accuracy and recall for a category, where the average of F1-score for each category is shown. (b)‒(e) The recognition accuracy of the AI model for pollen images of different species: (b) Alangium platanifolium, 98.2%; (c) Erigeron annuus, 87.5%; (d) Stellaria media, 96.4%; (e) Dontostemon dentatus, 91.4%.

Table 2 Comparison of the parameters of weighted plant- pollinator interaction networks quantified by AI models and human recognition. There were no significant differences in network metrics such as connectivity, nestedness, specialization level, clustering coefficient, and Shannon index. The significance of parameters was evaluated by comparing H2° and weighted NODF (wNODF) with randomly constructed null model networks. **P < 0.01.

网络指标 Network parameter	AI 模型 AI model	人工识别 Human recognition
连接度 Connectance	0.188	0.181
加权连接度 Weighted connectance	0.045	0.045
物种连接量 Links per species	2.100	2.014
特化值H₂° Specialization value H₂°	0.643^**	0.641^**
加权嵌套度 wNODF	13.847^**	14.298^**
聚类系数 Cluster coefficient	0.143	0.143
连接密度 Linkage density	3.185	3.132
Shannon指数 Shannon diversity	3.350	3.298

Table 3 Pearson correlation of the positions of the plant and pollinator species in AI model and human recognition based on weighted pollination networks. **P < 0.01.

物种水平指数 Species-level index	Pearson相关性 Pearson correlation
传粉者 Pollinator
介数中心度 Betweenness centrality	0.881^**
接近中心度 Closeness centrality	0.878^**
特化值d° Specialization value d°	0.943^**
植物 Plant
介数中心度 Betweenness centrality	0.814^**
接近中心度 Closeness centrality	0.875^**
特化值d° Specialisztion value d°	0.980^**

Fig. 2 Differences between AI models and human recognition results. (a) The similarities and differences between AI models and human recognition of pollen images. The interaction difference between the AI model and the human recognition is then manually identified by a third party. (b) The similarities and differences in the number of pollen grains carried by insects as identified by AI model and human recognition. The pollen load difference between the AI model and the human identification is then manually identified by a third party.

Fig. 3 Qualitative plant‒pollinator network by human recognition (a) and by AI model (b). The top of the graph shows pollinator species (Brown: Lepidoptera; Red: Hymenoptera; Yellow: Coleoptera; Purple: Diptera; Black, Others). The bottom represents plant species and the middle line represents plant‒pollinator interaction. Grey lines represent interactions identified as the same by both human recognition and AI model, while red lines represent interactions unique to that network. For a detailed list of species abbreviations, see Appendix 1.

Fig. 4 Quantitative plant‒pollinator network by human recognition (a) and by AI model (b). The middle lines represent the interaction between plants and pollinators, with wider lines indicating a greater quantity of pollen carried by the pollinator from that plant. The top of the graph shows pollinator species (Brown, Lepidoptera; Red, Hymenoptera; Yellow, Coleoptera; Purple, Diptera; Black, Others), the bottom represents plant species, and for a detailed list of species abbreviations, see Appendix 1.

References 48

[1]	Alarcón R (2010) Congruence between visitation and pollen-transport networks in a California plant-pollinator community. Oikos, 119, 35-44.
[2]	Alarcón R, Waser NM, Ollerton J (2008) Year-to-year variation in the topology of a plant-pollinator interaction network. Oikos, 117, 1796-1807.
[3]	Baldock KCR, Memmott J, Ruiz-Guajardo JC, Roze D, Stone GN (2011) Daily temporal structure in African savanna flower visitation networks and consequences for network sampling. Ecology, 92, 687-698.
[4]	Barker DA, Arceo-Gómez G (2021) Pollen transport networks reveal highly diverse and temporally stable plant-pollinator interactions in an Appalachian floral community. AoB Plants, 13, plab062.
[5]	Beil M, Horn H, Schwabe A (2008) Analysis of pollen loads in a wild bee community (Hymenoptera: Apidae)—A method for elucidating habitat use and foraging distances. Apidologie, 39, 456-467.
[6]	Bosch J, González AM, Rodrigo A, Navarro D (2009) Plant- pollinator networks: Adding the pollinator’s perspective. Ecology Letters, 12, 409-419.
[7]	Bourel B, Marchant R, de Garidel-Thoron T, Tetard M, Barboni D, Gally Y, Beaufort L (2020) Automated recognition by multiple convolutional neural networks of modern, fossil, intact and damaged pollen grains. Computers & Geosciences, 140, 104498.
[8]	Burkle LA, Alarcón R (2011) The future of plant-pollinator diversity: Understanding interaction networks across time, space, and global change. American Journal of Botany, 98, 528-538. DOI PMID
[9]	Cai XY, Hou YQ, Muhammad UJ, Wang H, Xu YC, Zheng J, Wang YH, Liu F, Zhou ZL, Hua JP, Wang KB (2023) The morphological diversity of pollen in the genus Gossypium. Journal of Cotton Research, 6, 6.
[10]	CaraDonna PJ, Burkle LA, Schwarz B, Resasco J, Knight TM, Benadi G, Blüthgen N, Dormann CF, Fang Q, Fründ J, Gauzens B, Kaiser-Bunbury CN, Winfree R, Vázquez DP (2021) Seeing through the static: The temporal dimension of plant-animal mutualistic interactions. Ecology Letters, 24, 149-161.
[11]	de Manincor N, Hautekèete N, Mazoyer C, Moreau P, Piquot Y, Schatz B, Schmitt E, Zélazny M, Massol F (2020) How biased is our perception of plant-pollinator networks? A comparison of visit- and pollen-based representations of the same networks. Acta Oecologica, 105, 103551.
[12]	de Santiago-Hernández MH, Martén-Rodríguez S, Lopezaraiza-Mikel M, Oyama K, González-Rodríguez A, Quesada M (2019) The role of pollination effectiveness on the attributes of interaction networks: From floral visitation to plant fitness. Ecology, 100, e02803.
[13]	Dormann CF, Fründ J, Blüthgen N, Gruber B (2009) Indices, graphs and null models: Analyzing bipartite ecological networks. The Open Ecology Journal, 2, 7-24.
[14]	Dunker S, Motivans E, Rakosy D, Boho D, Mäder P, Hornick T, Knight TM (2021) Pollen analysis using multispectral imaging flow cytometry and deep learning. New Phytologist, 229, 593-606.
[15]	Encinas-Viso F, Bovill J, Albrecht DE, Florez-Fernandez J, Lessard B, Lumbers J, Rodriguez J, Schmidt-Lebuhn A, Zwick A, Milla L (2023) Pollen DNA metabarcoding reveals cryptic diversity and high spatial turnover in alpine plant- pollinator networks. Molecular Ecology, 32, 6377-6393.
[16]	Fan W, Liu HT, Zhang Y, Chen XL, Huang MG, Xu BQ (2024) Diagnostic value of artificial intelligence based on computed tomography (CT) density in benign and malignant pulmonary nodules: A retrospective investigation. PeerJ, 12, e16577.
[17]	Fang Q, Huang SQ (2016) Plant-pollinator interactions in a biodiverse meadow are rather stable and tight for 3 consecutive years. Integrative Zoology, 11, 199-206. DOI PMID
[18]	Holt KA, Bennett KD (2014) Principles and methods for automated palynology. New Phytologist, 203, 735-742. PMID
[19]	Jędrzejewska-Szmek K, Zych M (2013) Flower-visitor and pollen transport networks in a large city: Structure and properties. Arthropod-Plant Interactions, 7, 503-516.
[20]	Kaneko M, Tsuji K, Masuda K, Ueno K, Henmi K, Nakagawa S, Fujita R, Suzuki K, Inoue Y, Teramukai S, Konishi E, Takamatsu T, Ukimura O (2022) Urine cell image recognition using a deep-learning model for an automated slide evaluation system. BJU International, 130, 235-243.
[21]	King C, Ballantyne G, Willmer PG (2013) Why flower visitation is a poor proxy for pollination: Measuring single-visit pollen deposition, with implications for pollination networks and conservation. Methods in Ecology and Evolution, 4, 811-818.
[22]	Klumpers SGT, Stang M, Klinkhamer PGL (2019) Foraging efficiency and size matching in a plant-pollinator community: The importance of sugar content and tongue length. Ecology Letters, 22, 469-479. DOI PMID
[23]	Lázaro A, Hegland SJ, Totland Ø (2008) The relationships between floral traits and specificity of pollination systems in three Scandinavian plant communities. Oecologia, 157, 249-257. DOI PMID
[24]	Li X, Prebble JG, de Lange PJ, Raine JI, Newstrom-Lloyd L (2022) Discrimination of pollen of New Zealand mānuka (Leptospermum scoparium agg.) and Kānuka (Kunzea spp.) (Myrtaceae). PLoS ONE, 17, e0269361.
[25]	Lomáscolo SB, Giannini N, Chacoff NP, Castro-Urgal R, Vázquez DP (2019) Inferring coevolution in a plant- pollinator network. Oikos, 128, 775-789.
[26]	Marcos JV, Nava R, Cristóbal G, Redondo R, Escalante- Ramírez B, Bueno G, Déniz Ó, González-Porto A, Pardo C, Chung F, Rodríguez T (2015) Automated pollen identification using microscopic imaging and texture analysis. Micron, 68, 36-46. DOI PMID
[27]	Miele V, Ramos-Jiliberto R, Vázquez DP (2020) Core-periphery dynamics in a plant-pollinator network. Journal of Animal Ecology, 89, 1670-1677. DOI PMID
[28]	Ollerton J (2017) Pollinator diversity: Distribution, ecological function, and conservation. Annual Review of Ecology, Evolution, and Systematics, 48, 353-376.
[29]	Olsson O, Karlsson M, Persson AS, Smith HG, Varadarajan V, Yourstone J, Stjernman M (2021) Efficient, automated and robust pollen analysis using deep learning. Methods in Ecology and Evolution, 12, 850-862.
[30]	Parra-Tabla V, Arceo-Gómez G (2021) Impacts of plant invasions in native plant-pollinator networks. New Phytologist, 230, 2117-2128. DOI PMID
[31]	Persson AS, Mazier F, Smith HG (2018) When beggars are choosers—How nesting of a solitary bee is affected by temporal dynamics of pollen plants in the landscape. Ecology and Evolution, 8, 5777-5791. DOI PMID
[32]	Pichler M, Hartig F (2023) Machine learning and deep learning—A review for ecologists. Methods in Ecology and Evolution, 14, 994-1016.
[33]	Pospiech M, Javůrková Z, Hrabec P, Štarha P, Ljasovská S, Bednář J, Tremlová B (2021) Identification of pollen taxa by different microscopy techniques. PLoS ONE, 16, e0256808.
[34]	Schwarz B, Vázquez DP, CaraDonna PJ, Knight TM, Benadi G, Dormann CF, Gauzens B, Motivans E, Resasco J, Blüthgen N, Burkle LA, Fang Q, Kaiser-Bunbury CN, Alarcón R, Bain JA, Chacoff NP, Huang SQ, LeBuhn G, MacLeod M, Petanidou T, Rasmussen C, Simanonok MP, Thompson AH, Fründ J (2020) Temporal scale-dependence of plant-pollinator networks. Oikos, 129, 1289-1302.
[35]	Sevillano V, Aznarte JL (2018) Improving classification of pollen grain images of the POLEN23E dataset through three different applications of deep learning convolutional neural networks. PLoS ONE, 13, e0201807.
[36]	Sevillano V, Holt K, Aznarte JL (2020) Precise automatic classification of 46 different pollen types with convolutional neural networks. PLoS ONE, 15, e0229751.
[37]	Thomson J (2021) How worthwhile are pollination networks? Journal of Pollination Ecology, 28, i-vi.
[38]	Tong ZY, Wu LY, Feng HH, Zhang M, Armbruster WS, Renner SS, Huang SQ (2023) New calculations indicate that 90% of flowering plant species are animal-pollinated. National Science Review, 10, nwad219.
[39]	Tourbez C, Gómez-Martínez C, González-Estévez MÁ, Lázaro A (2024) Pollen analysis reveals the effects of uncovered interactions, pollen-carrying structures, and pollinator sex on the structure of wild bee-plant networks. Insect Science, 31, 971-988.
[40]	Tsymbalyuk ZM, Celenk S, Mosyakin SL, Nitsenko LM (2021) Pollen morphology of some species of the genus Cephalaria Schrad. (Caprifoliaceae) and its significance for taxonomy. Microscopy Research and Technique, 84, 682-694. DOI PMID
[41]	Tu YL, Wang LP, Wang XL, Wang LL, Duan YW (2019) Status of invasive plants on local pollination networks: A case study of Tagetes minuta in Tibet based on pollen grains from pollinators. Biodiversity Science, 27, 306-313. (in Chinese with English abstract)
	[土艳丽, 王力平, 王喜龙, 王林林, 段元文 (2019) 利用昆虫携带的花粉初探西藏入侵植物印加孔雀草在当地传粉网络中的地位. 生物多样性, 27, 306-313.] DOI
[42]	Vázquez DP, Blüthgen N, Cagnolo L, Chacoff NP (2009) Uniting pattern and process in plant-animal mutualistic networks: A review. Annals of Botany, 103, 1445-1457. DOI PMID
[43]	Vázquez DP, Lomáscolo SB, Maldonado MB, Chacoff NP, Dorado J, Stevani EL, Vitale NL (2012) The strength of plant-pollinator interactions. Ecology, 93, 719-725. PMID
[44]	Wood TJ, Holland JM, Goulson D (2017) Providing foraging resources for solitary bees on farmland: Current schemes for pollinators benefit a limited suite of species. Journal of Applied Ecology, 54, 323-333.
[45]	Xiong XH, Zhou XM, Li M, Xu B, Deng HN, Yu Q, Gao XF (2019) Pollen morphology in Rubus (Rosaceae) and its taxonomic implications. Plant Systematics and Evolution, 305, 705-716.
[46]	Xu Q, Zhang M, Gu ZH, Pan G (2019) Overfitting remedy by sparsifying regularization on fully-connected layers of CNNs. Neurocomputing, 328, 69-74.
[47]	Xu XL, Zheng YZ, Chen XC, Zhu FL, Miao XQ (2018) Identification of cattail pollen, pine pollen and bee pollen by Fourier transform infrared spectroscopy and two-dimensional correlation infrared spectroscopy. Journal of Molecular Structure, 1167, 78-81.
[48]	Zhao YH, Lázaro A, Ren ZX, Zhou W, Li HD, Tao ZB, Xu K, Wu ZK, Wolfe LM, Li DZ, Wang H (2019) The topological differences between visitation and pollen transport networks: A comparison in species rich communities of the Himalaya-Hengduan Mountains. Oikos, 128, 551-562.