New dimension in leaf stomatal behavior analysis: a robust method with machine learning approach

Abstract

Stomata are specialized pores that play a vital role in gas exchange and photosynthesis. Microscopic images are often used to assess stomatal characteristics in plants; however, this can be a challenging task. By utilizing Matterport’s Mask R-CNN implementation as the foundational model, fine-tuning was conducted on a dataset of 810 microscopic images of Hedyotis corymbosa leaves’ surfaces for automated stomatal pores detection. The outcomes were promising, with the model achieving a convergence of 98% mean average precision (mAP) for both detection and segmentation. The training loss and validation loss values converged around 0.18 and 0.37, respectively. Regression analyses demonstrated the statistical significance (p values ≤ 0.05) of predictor parameters. Notably, the tightest cluster of data points was observed in stomata pore area measurements, followed by width and length. This highlights the precision of the stomatal pore area in characterizing stomatal traits. Despite challenges posed by the original dataset’s low-resolution images and artifacts like dust, bubbles, and blurriness, our innovative utilization of the Mask R-CNN algorithm yielded commendable outcomes. This research introduces a robust approach for stomatal phenotyping with broad applications in plant biology and environmental studies.