Remotely-Sensed Predictor Data
Predictors included gridded environmental data derived from remote sensing. These rasters encompassed topographic, edaphic, hydrological, climatic, and vegetation gradients. We emphasized predictors shaping the distribution of each ecosystem property represented in our independent pool of response data; selections were informed from published models built to predict spatial variations in biodiversity andgeodiversity (Hjort and Luoto 2012, Mod et al. 2016, Simensen et al. 2020). Where possible, we also sought predictors with a direct influence on the distribution of ecosystem properties, instead of employing proxies (Soley-Guardia et al. 2024).
Following these conditions, we included 22 rasters in our final predictor dataset (see S1). We conducted all raster processing using the terra package (v1.7.29) (Hijmans 2023) in R (v4.3.2) (R Core Team 2023). In addition, we cropped and masked rasters to the extent of Nova Scotia; rescaled them to 10 m using bilinear interpolation, where necessary; and projected them to NAD1983 CSRS v6 UTM Zone 20N. To minimize false accuracies imparted by down scaling (Sillero and Barbosa 2021), we only selected predictor variables with a native spatial grain of ≤10 m (details in S1) to match the finest grain in nested provincial ecosystem surveys. Leaf area index was an exception; data were downscaled from 20 m (see S1).
We extracted predictor values at plot locations using the simple method (Hijmans 2023). Afterward, we used correlation (r = 0.7) (Guisan et al. 2017) and variance inflation factor (VIF = 3) analyses (Zuur et al. 2009), to identify correlated environmental predictors, employing the corrplot (v0.92) (Wei and Simko 2021) and usdm (v2.1-6) (Naimi et al. 2014) R packages respectively. To choose between highly correlated predictor pairs, we ran individual models with each predictor, from those pairs, and selected the predictor returning better model fit (see S1). These procedures reduced the number of predictors from 22 to 18. Coupled with geography (i.e., the Euclidian distance between each plot pair), those 18 environmental variables were employed as baseline predictors for modelling (see S1).