Using Space-Based Imagery for the Identification of Hemlock Forests to Model Future Invasive Species Expansion due to Warming Winter Temperatures

The hemlock woolly adelgid (HWA, Adelges tsugae), is a non-native, invasive species that currently threatens eastern hemlock trees (Tsuga canadensis) in US forests. The HWA has continued to expand north along the US east coast due to warming winters as temperatures reach above -20°C. Eastern hemlock is a keystone species in these regions and has a greater capacity to store carbon, regulate stream temperatures, and provide habitat for wildlife compared to sympatric tree species. Currently, detailed models generating distribution maps of hemlock and HWA lack high spatial resolution over areas of large spatial domain for land managers to use as tools to lead efforts mitigating HWA expansion. This study created two hemlock distribution maps over New York state using a random forest classifier in Google Earth Engine, each using data from Landsat 8 Operational Land Imager, Sentinel-2 Multispectral Instrument, as well as habitat parameters (e.g., elevation, slope, stream distance, soil type, temperature). Ground-surveyed hemlock presence points provided by land managers allowed the satellite model to be trained and compared to the accuracy of a previous AVIRIS hyperspectral model. A preliminary accuracy of the models found that the Landsat 8 model was 67% accurate compared to the 53% accuracy of the AVIRIS model over a sample region in New York state. The study continued to build off the satellite-based modeled distribution maps to predict hemlock mortality in New York through 2049 using two different forecasting techniques. The first technique used a linear rate of expansion of HWA without consideration for other environmental factors in creating a buffer that displayed the spread of HWA in New York through 2049 and masked areas below -20°C; a temperature HWA dies in. The output from the satellite-based hemlock distribution map was used to mask out regions that could not support HWA populations. The second technique used the present-day distribution model of hemlock presence and added present-day temperature data to train a separate random forest classification algorithm. The random forest classification algorithm input consisted of the original satellite-based hemlock distribution map, past and present temperature simulations using the NASA NEX-DCP 30 downscaled climate data and HWA presence points. This input data trained the algorithm under present-day conditions. The output from the random forest classification model was then paired with future projections from the NEX-DCP 30 temperature data through the year 2049. Both forecasting models yielded similar results in that hemlock will continue to decline due to HWA expansion in 2049 throughout much of New York due to warming winter temperatures. In contrast, the Adirondack Mountains region may not see declines in hemlock, due to colder temperatures in high elevations. The study concluded that satellite-based imagery paired with climate projections and robust ground-truth data could provide for an accurate and cost-effective model that is available to the public in creating habitat distribution maps and ecological forecasting models.