Training Data

For each class in our landcover taxonomy, we created between 500 and 5000 observations divided among the years 2008, 2011, 2013, 2016, 2019, and 2021 based on the consensus among the following landcover-related datasets:

Dataset	Citation	Data availability
Dynamic World v1	Brown et al. 2022	2015-present
LCMAP v1.3	Xian et al. 2022	1985-2021
US National Landcover Database	Homer et al. 2020
Rangeland Analysis Platform	Jones et al. 2018
JRC Global Surface Water	Pekel et al. 2016
USDA CroplandCROS	Boryan et al. 2011
Canada Annual Crop Inventory	Fizette et al. 2013
North American Land Change Monitoring System
ESA WorldCover

More observations were produced for more common landcover classes compared to rare classes. Not all datasets were available for every year across the full extent of the region. Also, the landcover datasets in the table above only partially conformed to our landcover taxonomy, and some landcover classes were widely included in these datasets while others were only available in a small subset. Given these limitations, there were a variable number of datasets that were available to assess model agreement depending on the class, location, and year. However, in all cases, at least 2 datasets agreed on the class of the observation in our training dataset. The total number of training data samples was 78,500.

For each observation of a consensus landcover class, our Google Earth Engine (GEE) based workflow extracted the value of all covariates matched to the year corresponding to the sampling year of the consensus landcover observation. The result was a table (a GEE FeatureCollection) with attributes for the consensus landcover class and all covariates.

All training data locations were assigned a 'fold' (value from 1 to 10) that was used to reduce autocorrelation during model validation. Folds were assigned based on a 10 km x 10 km grid of randomized folds across the region. During model training, one fold of data was always withheld, and because of the spatial assignment of folds, the withheld data was spatially separated from all other folds and therefore expected to be a more independent assessment compared to validation based on folds that were spatially intermixed.