Automated starch granule classification using high-throughput microscopy and machine learning

Starch analysis is widely used in archaeology to investigate the processing of wild plants for food and medicine, as well as the domestication and spread of cultigens. Starch analyses are dependent on the development and use of identification keys. To date, all published methods to identify starch granules are either time-consuming to produce and apply, or difficult to statistically validate for their accuracy. The method outlined in this paper mitigates two major production bottlenecks experienced by current methods. First, the necessary task of collecting reference images of starch granules is accelerated using a multispectral imaging flow cytometry (MIFC), a high throughput microscope that collects thousands of images per second. Second, the traditional step of collecting measurements of individual granules by hand is eliminated. Processed image sets are used directly to train image recognition (machine learning) algorithms for species identification. This method produces identification accuracies that are comparable to, or better than, other published methods. When this method is applied using 15,000 images of starch granules from 15 plants occurring in North America, validation accuracies were observed to be as high as 100 %. This method promises to provide a feasible, cost-effective, and accurate means to identify starch granules recovered from archaeological materials. • Deep learning algorithms can be used to classify images of modern starch granules with a high degree of accuracy. • Multispectral imaging flow cytometry can be used to collect image datasets large enough to train deep learning algorithms. • This methodology promises a feasible alternative to other published statistical starch grain classification methods.