This study measured pollen germination percentage and pollen tube elongation in Douglas-fir (Pseudotsuga menziesii) across an incubation temperature series from 5°C to 40°C. Microscopic images of pollen samples incubated at these temperatures provided the raw data for phenotyping. The authors applied an image-based workflow to enable higher-throughput quantification compared with manual scoring, with a focus on deriving temperature response curves for germination and elongation across multiple populations collected along an elevational gradient.
A convolutional neural network (CNN) was developed to segment pollen grains in microscopic images and to classify germination status. For segmentation (separating pollen grains from background), the CNN achieved an intersection-over-union (IoU) of 0.846, indicating high accuracy in identifying pollen objects in the images. The network successfully distinguished pollen from background artifacts.
Classification performance for distinguishing germinated versus ungerminated pollen was more limited: the model showed only moderate accuracy when pollen had initiated early elongation. The authors note this reduced performance specifically during early tube emergence, when visual differences are subtle and harder to classify automatically. The CNN nonetheless supported automated extraction of germination percentages and pollen length measurements across the temperature treatments.
From the segmented images and classified objects, the study quantified two primary response variables across temperatures: the percentage of pollen grains that germinated and the length of pollen tubes (pollen elongation). Measurements spanned incubation temperatures between 5 and 40°C, allowing construction of full temperature response curves for both traits. Both traits displayed bell-shaped responses to temperature in the measured range.
To summarize temperature responses, the authors fitted gamma functions to the germination and elongation data and used these fits to estimate thermal optima for each response and population. Gamma-function fits produced characteristic bell-shaped curves and allowed estimation of the temperature at which each trait reached its peak performance.
Across the Douglas-fir samples, estimated optimal temperatures for reproductive performance fell within a relatively narrow interval, approximately 19 to 23°C. Notably, pollen elongation consistently reached its optimal temperature at a higher point on the thermal axis than pollen germination, indicating different temperature sensitivities for initiation versus subsequent tube growth.
The sampled Douglas-fir populations were collected across an elevational gradient to test whether thermal optima varied predictably with elevation. The analysis did not detect a statistically significant relationship between elevation and estimated thermal optima. The authors did observe that some higher-elevation populations exhibited lower optimal temperatures, but this pattern was not strong enough to constitute a significant correlation in the reported analyses.
The study compared its Douglas-fir results with prior analyses of three western North American conifers. These comparisons indicated that each species occupied a distinct reproductive thermal niche—that is, the thermal range for peak reproductive performance corresponded to typical spring temperatures in each species' native habitat. This provides context suggesting species-specific reproductive thermal niches among regional conifers.
The findings imply that Douglas-fir pollen performance has a narrow thermal window for germination and elongation. Because estimated optima fall near 19–23°C and elongation peaks at higher temperatures than germination, reproductive success may be sensitive to temperature fluctuations during the reproductive period. The authors highlight that warming climates could therefore alter reproductive success and potentially affect forest regeneration dynamics. The study also demonstrates that deep learning can provide an efficient, scalable framework for high-throughput quantification of pollen viability and growth from microscopic images, facilitating broader assessments of thermal sensitivity across populations and species.
This article is a preprint and has not been peer-reviewed. The CNN exhibited only moderate accuracy for classifying germination during early pollen tube emergence, a limitation acknowledged by the authors. Specific methodological details such as sample sizes, exact CNN architecture and hyperparameters, statistical test results, and the numerical outputs of gamma-function fits (beyond the reported optimal temperature range) were reported in the source but are not restated here in full. The authors declared no competing interests.
Note: details beyond what is presented in the source (for example, exact sample counts, training-validation splits, or p-values) were not provided in the summarized text and therefore are not reported here.