Machine-learning-based morphological analyses of leaf epidermal cells in modern and fossil ginkgo and their implications for palaeoclimate studies

Leaf stomata form an essential conduit between plant tissue and the atmosphere, thus presenting a link between plants and their environments. Changes in their properties in fossil leaves have been studied widely to infer palaeo-atmospheric-CO₂ in deep time, ranging from the Palaeozoic to the Cenozoic. Epidermal cells of leaves, however, have often been neglected for their usefulness in reconstructing past-environments, as their irregular shape makes the manual analyses of epidermal cells a challenging and error-prone task. Here, we used machine-learning (using the U-Net architecture, which evolved from a fully convolutional network) to segment epidermal cells automatically, to efficiently reduce artificial errors. We furthermore applied minimum bounding rectangles to extract length-to-width ratios (R_L/W) from the irregularly shaped cells. We applied this to a dataset including over 21 000 stomata and 170 000 epidermal cells in 114 Ginkgo leaves from 16 locations spanning three climate zones in China. Our results show negative correlations between the R_L/W and specific climatic parameters, suggesting that local temperature and precipitation conditions may have affected the R_L/W of epidermal cells. We subsequently tested this methodology and the observations from the modern dataset on 15 fossil ginkgoaleans from the Lower to the Middle Jurassic (China). It suggested that the R_L/W values of fossil ginkgo generally had a similar negative response to warmer climatic backgrounds as modern G. biloba. The automated analyses of large palaeo-floral datasets provide a new direction for palaeoclimate reconstructions and emphasize the importance of hidden morphological characters of epidermal cells in ginkgoaleans.