AAAR 32nd Annual Conference
September 30 - October 4, 2013
Oregon Convention Center
Portland, Oregon, USA
Abstract View
Automated Pollen Identification and Counting System (APICS)
JAMES HOUSE, Gregory Griffin, Richard Flagan, Caltech
Abstract Number: 387 Working Group: Bioaerosols: Characterization and Environmental Impact
Abstract The process of counting pollen is laborious, slow, expensive, and of questionable accuracy. As a result, measurements are routinely made at fewer than 100 sites within the U.S. The available data are far too sparse to enable robust estimations of exposure. Moreover, data reporting from many of the present pollen counters is often restricted to coarse classes of pollen (e.g., grasses, trees, and weeds. Individual species counts are needed to assess health risks of allergic individuals. While other nations have national pollen counting networks that are operated by scientists, allergy clinics provide most of the documented pollen measurements in the U.S. Efforts to conduct epidemiological studies of the health impacts of pollen, and to elucidate the effects of climate change on pollen and allergic disease are confounded by the lack of historical trends on pollen, and by the lack of a robust pollen database that includes measurements from all regions with sufficient spatial resolution to assess geographic and local variations in concentrations, and that is accessible for scientific research. To enable such a network, the measurements must be made much more cost-effective and reliable.
We report on the development of an automated pollen identification and counting system (APICS) as a step in this direction. We have applied this system to the identification of pollen in samples that were collected using a conventional Burkard pollen sampler. The samples are affixed to microscope slides, each representing one day. A computer-controlled microscope, driven by software written in Java and using MicroManager, an open-source microscopy program, captures images from each slide. The images are processed to produce focused images of candidate objects within each field of view. The resulting candidate objects are then identified using computer-vision software that has been trained using a large library of labeled images. Using a desktop PC for the analysis, the system processes samples representing one week in less than 24 hours. We continue to work toward shorter analysis times.
The performance of APICS has been characterized in terms of its accuracy by comparison with human pollen counters. Initial measurements duplicated the scan procedures that are typically employed by human counters, i.e., a linear scan through the length of each slide. The automated system enables analysis of many more fields of view than can a human counter, including whole slide scans. Using these comprehensive scans, we have examined the biases associated with conventional methods, and explored alternate scan patterns that to reduce those biases. APICS has been used to provide daily pollen measurements for Pasadena, CA to reveal detailed patterns and trends in the pollen data that have not been possible with previous manual counts.