COS 290-3 - A Comparison Shopper’s Guide to forest cover datasets

Abstract: Where is forest cover increasing and declining in the United States? This is an important question for climate change mitigation, biodiversity conservation, and many other goals, yet it is surprisingly difficult to answer. Although scientists, consultants, policy-makers, and the general public have access to an ever-growing number of spatial datasets including maps produced from Earth observation imagery, these resources vary in their definitions of forest cover as well as the methods used to map them.

In this study, we compared forest cover estimates from eight different land cover datasets and two forest inventory datasets for the conterminous United States. For spatial datasets, we used Google Earth Engine to produce forest/non-forest maps by aggregating map classes representing forests, trees, and other descriptions of woody biomass or applying a series of thresholds to class probabilities and continuous estimates of canopy cover. We then summarized changes in pixel counts by state for each year in each dataset. We also include state-level summaries from the USFS Forest Inventory and Analysis census and EPA GHG Inventory in our comparisons. Time series of forest cover estimates for each dataset were compared in terms of differences in spatial resolution, temporal coverage and update frequency, definition of forest classes, differences in magnitude of forest cover estimates, and differences in forest cover trends over time.

After accounting for different label aggregations and threshold options, we compared over twenty different estimates of forest cover. Spatial resolutions of the map datasets considered ranged from 500 meters (MODIS C6 Land Cover) to 10 meters (ESA World Cover, Google Dynamic World, and ESRI Global Land Cover). Temporal coverage was also determined by choice of input imagery, with Landsat-based datasets (USGS LCMAP and USFS LCMS) and the EPA GHG Inventory providing the longest histories (1985-2022 and 1990-2019, respectively). We find the direction, magnitude, significance, and agreement of trends varies across datasets, with MODIS C6 Land Cover legends and different thresholds on NLCD Tree Canopy Cover typically showing the greatest ranges of variability in terms of magnitude while LCMS and Dynamic World derivatives show the greatest year-to-year instability.

We show that forest cover estimates range up to 25% of the total land area of CONUS depending on source and definition with non-trivial implications for downstream use cases. Thus, this work provides important context for ecologists across sectors who rely on map- and inventory-based forest cover estimates.