Estimating Uncertainty Of Forest Species Distributions To Better Forecast The Spread Of Forest Pests And Pathogens

The Pest or Pathogen Spread (PoPS) Forecasting System is a web-based interface that provides a flexible, generalized framework for modeling the spread of pests or pathogens across landscapes with fully specified uncertainties. This model provides land managers an interactive way to forecast biotic invasions in forest and agricultural areas and test adaptive management strategies. Underlying these models are species distribution data on which the model relies to predict biotic spread. However, species distribution data are often lacking in spatial resolution, accuracy, and temporal relevance. We have therefore been developing a species mapping module as part of the web-based framework to provide users a way to generate species maps on demand. We combine the use of machine learning probability estimations and a Bayesian framework applied to time-series remote sensing imagery to provide estimates of species distributions with fully quantified uncertainties across broad geographic areas. Google Earth Engine is used to create annual, summer composite spectral indices (NDVI, NDWI, EVI2) and Tasseled Cap Transformation from Landsat imagery. We then apply a calibrated Random Forest model to estimate the probability of species presence at each time step, trained on U.S. Forest Service Forest Inventory And Analysis (FIA) data. These annual probability surfaces are combined in a time-series Bayesian update framework that incorporates the strength of the model from each year and prior probabilities to generate posterior estimates of species presence. These refined estimates of species distribution with uncertainty are then used in the PoPs model to improve accuracy and understanding of host mapping uncertainty as a driver of forecast uncertainty in pest and pathogen spread. Results show that uncertainty estimates from this approach are improved over single year estimates alone by capturing land use change over time, and improve species discrimination when using broadband spectral data.