Abstract
Stomata are an attractive model for studying the physiological responses of plants to various environmental stimuli1–3. Of the morphological parameters that represent the degree of stomatal opening, the length of the minor axis of the stomatal pore (the stomatal aperture) has been most commonly used to dissect the molecular basis of its regulation. Measuring stomatal apertures is time consuming and labour intensive, preventing their use in large-scale studies. Here, we completely automated this process by developing a program called DeepStomata, which combines stomatal region detection and pore isolation by image segmentation. The former, which comprises histograms of oriented gradients (HOG)-based stomatal detection and the convolutional neural network (CNN)-based classification of open/closed-state stomata, acts as an efficient conditional branch in the workflow to selectively quantify the pores of open stomata. An analysis of batches of images showed that the accuracy of our automated aperture measurements was equivalent to that of manual measurements, however had higher sensitivity (i,e., lower false negative rate) and the process speed was at least 80 times faster. The outstanding performance of our proposed method for automating a laborious and repetitive task will allow researchers to focus on deciphering complex phenomena.