Precise Pollen Grain Detection in Bright Field Microscopy Using Deep Learning Techniques

Abstract

The determination of daily concentrations of atmospheric pollen is important in the medical and biological fields. Obtaining pollen concentrations is a complex and time-consuming task for specialized personnel. The automatic location of pollen grains is a handicap due to the high complexity of the images to be processed, with polymorphic and clumped pollen grains, dust, or debris. The purpose of this study is to analyze the feasibility of implementing a reliable pollen grain detection system based on a convolutional neural network architecture, which will be used later as a critical part of an automated pollen concentration estimation system. We used a training set of 251 videos to train our system. As the videos record the process of focusing the samples, this system makes use of the 3D information presented by several focal planes. Besides, a separate set of 135 videos (containing 1234 pollen grains of 11 pollen types) was used to evaluate detection performance. The results are promising in detection (98.54% of recall and 99.75% of precision) and location accuracy (0.89 IoU as the average value). These results suggest that this technique can provide a reliable basis for the development of an automated pollen counting system.

Using ITS2 metabarcoding and microscopy to analyse shifts in pollen diets of honey bees and bumble bees along a mass‐flowering crop gradient

“…New developments in microscopic pollen detection using deep learning techniques (Gallardo-Caballero et al, 2019) or in full-length amplicon or genome sequencing with, for example, nanopore sequencing techniques (Lang et al 2019;Leidenfrost et al, 2020;Peel et al, 2019) could improve the weaknesses of both approaches (e.g., time expenditure in microscopy or quantification accuracy in molecular methods). However, studies are needed to compare and evaluate the accuracy of those new developments.…”

Estimation of pollen counts from light scattering intensity when sampling multiple pollen taxa – Establishment of Automated Multi-taxa Pollen Counting Estimation System (AME System)

“…However, recent technological developments have made automatic pollen sampling and identification possible (Buters et al 2018), for example, with recognition systems using microscopic images of pollen grains (Boucher et al 2002;Ranzato et al 2007;Oteros et al 2015), pollen color patterns from pollen images (Landsmeer et al 2009), fluorescence emission signals, (Swanson and Huffman 2018;Mitsumoto et al 2009;Mitsumoto et al 2010;Richardson et al 2019), light scattering (Crouzy et al 2016;Šaulienė et al 2019Šaulienė et al , holographic images (Sauvageat et al 2019, size and morphological characteristics (O'Connor et al 2013), real-time PCR (Longhi et al 2009), texture and infrared patterns of microscopic images of pollen (Marcos et al 2015;Gottardini et al 2007;Chen et al 2006), or a combination of several of these. Many studies applied machine learning algorithms to the problem (Punyasena et al 2012;Tcheng et al 2016;Crouzy et al 2016;Gonçalves et al 2016;Gallardo-Caballero et al 2019;Šaulienė et al 2019). These automated pollen identification methods have been applied not only to aerobiological research but also to palynological studies for the identification of fossilized pollen (France et al 2000;Kaya et al 2014;Li et al 2004;Zhang et al 2004;Rodríguez-Daminán et al 2006).…”

Quantitative methods in microscopy to assess pollen viability in different plant taxa

“…To reduce proportional bias, Tello et al ( 2018 ) also implemented an approach based on dimensional and shape characteristics of identified grains. Separating overlapping objects is a complicated task, but capabilities in computer vision are steadily improving and possible solutions are starting to appear in the literature (Gallardo-Caballero et al 2019 ; Cohen et al 2017 ; Molnar et al 2016 ).…”

“…The training set for the supervised classification process should be at least made of 40 objects for each class. Random forest usually offers a good balance between computing time and classification accuracy, nevertheless other models should be considered and tested for specific needs (Kuhn 2008 ; Gallardo-Caballero et al 2019 ).…”