Machine learning and pollen Identification: CNN for identification of Ambrosia among late season pollen

Abstract text (incl. figure legends and references)

Introduction

In the past years Convolutional Neuronal Networks (CNN) have reached a state where this software has the potential to simplify and accelerate the processing of images and data. A promising field is aerobiology, where the current standard is to sample particles and make a slide for Light Microscopic inspection by eye. Microscope Image Analysis System (MIAS®) enables to digitize samples on a slide as a series of high quality images. A logical next step is implementing CNN for the identification of the particles. After successful tests on tree pollen the identification of Ambrosia among late season pollen was the next application for our CNN.Ambrosia artemisiifolia has been seen as an aggressive invasive species in Germany over the last decade and been object to ongoing surveys in Europe.

Materials and Methods

The MIAS® relies on Standard methods for the sampling and staining of samples. Operating our automatic light microscope aeroIScope® simulates the routine particle counting: It produces a series of adjacent images and the details of the particles are documented by images from adjoining focus levels. Samples for the training of the CNN, the inspection of these images is supported by a tool named Marker, which enables efficiently labelling and image-related listing of structures.The CNN of MIAS® was trained to discriminate Ambrosia pollen against pollen from relevant late season species at an accuracy over 80%.

Results: The CNN reached an identification accuracy in comparing pollen of Ambrosia (99%), Amaranth (96%), Brassica (95%), Plantago (84%), Rumex (100%) and Urtica (100%), using a training set of only a few hundred pollen images for each species.

Results and Discussion

Using CNN in MIAS® makes the analysis of pollen objective and effective: Pollen from a daily slide are recognized in few minutes. - Based on a series of adjacent images provided by the aeroScope® - photographic documentation including clear relation to position on the slide is provided - comfortable evaluation anywhere is possible - a quality suitable for digital analysis to detect and classify distinct pollen is reached.

The digitalization of aerobiological samples makes it possible to share, conserve and compare these samples.

Conclusions

With further adjustments and training the CNN should be capable of facilitating and accelerating the analysis of large quantities of imigary.

References

Campesato, Oswald; "Artificial Intelligence - Machine Learning and Deep Learning"; Mercury Learning and Information; Dulles, Virgina; 2020.

Ansari, Shamshad; "Building Computer Vision Applications Using Artificial Neural Networks"; Apress Medi LLC, Springer Science+Business Media New York; 2020.

Campesato, Oswald; "TensorFlow 2.0: Pocket Primer"; Mercury Learning and Information; Dulles, Virgina; 2020.

Gulli, Antonio; Kapoor, Amita; Pal, Sujit ;"Deep Learning with TensorFlow 2 and Keras: Regression, ConvNets, GANs, RNNs, NLP, and more with TensorFlow 2 and the Keras API, 2nd Edition"; Packt Publishing; 2019.

Starfinger, Uwe; "The Action Programme Ambrosia in Germany – State of the art and future prospects", Julius Kühn Institut, Bundesforschungsinstitut für Kulturpflanzen; Quedlinburg, Germany; 2012.

Beate Alberternst, Stefan Nawrath und Frank Klingenstein; " Biology, distribution and pathways of introduction of Ambrosia artemisiifolia in Germany and assessment from a nature conservation point of view"; Eugen Ulmer KG, Stuttgart, Germany; 2006.