Tue, Aug 16, 2022: 5:00 PM-6:30 PM
ESA Exhibit Hall
Background/Question/MethodsCytotype, or number of chromosome copies, is a key component of genetic variation in many species. In quaking aspen, cytotype influences ecophysiological traits including water use efficiency and stomatal conductance, and thus may be a driver of plant-environment interactions. Once impractical, remote sensing and machine learning technologies present the opportunity to produce continuous, extensive maps of intraspecific cytotype variation from canopy spectra coupled to limited field sampling. This approach has been piloted in quaking aspen over small extents using high-resolution imagery collected via drones and aircraft but has not yet been scaled to regional levels.Here we scale-up cytotype mapping in quaking aspen from local to regional areas using multispectral satellite imagery and deep learning algorithms. We use genomic field samples and hyperspectral NEON imagery as a ‘data bridge’ to train an ensemble of fully convolutional neural networks to predict aspen coverage and cytotype from Sentinel-2 multispectral satellite imagery over larger areas. In doing so we ask 1-whether moderate resolution, multispectral satellite imagery is sufficiently sensitive to classify cytotype in quaking aspen at regional extents, and 2-how spectral signals of cytotype in the species vary across environments and years. Predictive skill is assessed in spatially explicit hold-out test sets.
Results/ConclusionsA preliminary version of our model classified diploid, triploid, and ‘other’ pixels with class-weighted average precision and recall scores of 92% and 93%, respectively. These promising early results suggest that it is possible to predict intraspecific cytotype variation in quaking aspen from multispectral satellite imagery using deep learning algorithms and a multi-scale data fusion approach.High resolution, continuous maps of quaking aspen cytotype over large areas will allow us to next test the hypothesis that models of mortality and range contractions in the species are more skillful when they incorporate intraspecific genetic information, relative to models based on environmental predictors alone. The difficulty of producing continuous datasets describing intraspecific genetic variation over large areas using costly molecular methods has traditionally prevented researchers from including such information in spatially explicit models. Maps of intraspecific genetic variation may be key to improving the predictive ability of species distribution and ecological forecasting models, as intraspecific genetic variation plays a critical role in determining how organisms respond to environmental change.
Results/ConclusionsA preliminary version of our model classified diploid, triploid, and ‘other’ pixels with class-weighted average precision and recall scores of 92% and 93%, respectively. These promising early results suggest that it is possible to predict intraspecific cytotype variation in quaking aspen from multispectral satellite imagery using deep learning algorithms and a multi-scale data fusion approach.High resolution, continuous maps of quaking aspen cytotype over large areas will allow us to next test the hypothesis that models of mortality and range contractions in the species are more skillful when they incorporate intraspecific genetic information, relative to models based on environmental predictors alone. The difficulty of producing continuous datasets describing intraspecific genetic variation over large areas using costly molecular methods has traditionally prevented researchers from including such information in spatially explicit models. Maps of intraspecific genetic variation may be key to improving the predictive ability of species distribution and ecological forecasting models, as intraspecific genetic variation plays a critical role in determining how organisms respond to environmental change.