Landcover

TerrAdapt's landcover model employs an advanced classification system to identify and categorize the 19 types of landcover across Washington. Using satellite imagery, geospatial data, and machine learning algorithms, the model is capable of distinguishing between different landcover types such as forests, urban areas, water bodies, agricultural lands, native shrubsteppe habitats and more (see the taxonomy in the table below).

Level 1	Level 2	Level 3	Value
Non-vegetated	Water	Water	1
Non-vegetated	Snow and Ice	Snow and Ice	2
Non-vegetated	Barren	Barren	3
Natural vegetation	Wetland	Emergent Wetland	4
Natural vegetation	Wetland	Woody Wetland	5
Natural vegetation	Grass and Forb	Mesic Grass and Forb	6
Natural vegetation	Grass and Forb	Xeric Grass and Forb	7
Natural vegetation	Shrub	Mesic Shrub	8
Natural vegetation	Shrub	Xeric Shrub	9
Natural vegetation	Forest	Deciduous Forest	10
Natural vegetation	Forest	Coniferous Forest	11
Human modified	Irrigated Agriculture	Pasture	12
Human modified	Irrigated Agriculture	Irrigated Rowcrops	13
Human modified	Irrigated Agriculture	Orchard	14
Human modified	Non-irrigated Agriculture	Non-Irrigated Rowcrops	15
Human modified	Non-irrigated Agriculture	Fallow	16
Human modified	Developed	Developed - High Intensity	17
Human modified	Developed	Developed - Medium Intensity	18
Human modified	Developed	Major Road	19

Training the model

For each class in our landcover taxonomy, we created between 500 and 5,000 observations divided among the years 2008, 2011, 2013, 2016, 2019, and 2021 based on the consensus among the following existing 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	1992, 2001, 2006, 2011, 2016, 2019, and 2021
Rangeland Analysis Platform	Jones et al. 2018	1986-present
JRC Global Surface Water v1.4	Pekel et al. 2016	1984-present
USDA CroplandCROS	Boryan et al. 2011	2008-present
Canada Annual Crop Inventory	Fizette et al. 2013	2009-present
North American Land Change Monitoring System	Houseman et al. 2022	1985-present
ESA WorldCover v200	Zanaga et al. 2021	2021

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 two 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-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 Google Earth Engine 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.

TerrAdapt's landcover model uses a variety of covariates as environmental predictors of the landcover classes in our taxonomy. The data sources for these covariates are described in the DATA INPUTS section. All covariates were stored in Google Earth Engine (either in the public data catalog or TerrAdapt's private asset storage) and available for use in our dynamic workflow.

Developing the landcover data product

We used the random forest classifier (Breiman 2001) implemented in Google Earth Engine to predict each landcover class in our taxonomy from the covariates at all training data locations. We trained 10 replicate models per landcover class in a k-fold approach (Fushiki 2011) with a unique 'fold' of 1/10th of the data withheld for validation as described above in the Training the model section. The model predicted the probability that the observation belonged to the class of interest rather than any other class.

In addition to assessing the probability that a pixel belonged to each landcover class in our taxonomy, we also created a discrete landcover classification based on the landcover class with the highest probability in each pixel. We modified this discrete classification by assigning all pixels that coincided with the road classes of 'freeway' or 'primary road' a class value of 19 to indicate the presence of a major road.