Application of a Self-learning Algorithm to Analyse Microscopic Images of Stomata

Abstract

The stomata on plant leaves are crucial for gas exchange and, when observed on (sub)fossil leaves, also provide insight into historic plant growth conditions such as atmospheric CO2 and humidity. Current methods to quantify stomatal traits rely on manual analysis of microscopic images, which is labour-intensive and requires expert knowledge. The use of computer-aided methods has the advantage to increase efficiency in data acquisition due to time-savings as the sample throughput is higher in a shorter time and thereby the potential of more reproducible results. This study applied a machine-learning approach developed by Jayakody, Liu, Whitty, & Petrie (2017). The aim was to test the effectiveness of the automated detection method for identifying stomata in microscopic images and its sensitivity towards stomata size and image quality. Two major experiments have been conducted using the originally introduced image dataset of grapevines from Jayakody et al. (2017), and an additional set of images from different plant types consisting of ferns and grasses, since specifically grasses feature a more complex stomata type. The outcome was compared to results by Jayakody et al. (2017) and manual counts of the fern and grass images. The images of the original dataset were manipulated by undergoing treatments of enlargement and reduction to mimic stomata size differences. For testing the influence of image quality, the original images were downsampled to achieve quality loss, while the fern and grass image dataset was classified into quality categories based on the perceived image quality by a human viewer. Additionally, comparisons between the detection accuracy of the different stomata types were carried out. It has been found that the algorithm reaches a limit in respect to stomata size leading to greater numbers of missed stomata indicating a lower tolerance towards variations in stomata size. In contrast, the method shows a high level of robustness in terms of image quality for the downsampled images generating a low number of incorrect detections. Also, the image quality of the fern and grass images does not seem to have a significant influence on the effectiveness of detecting stomata. In terms of stomata type, the algorithm handled both types well and no significant difference in accuracy has been found.