Analysis of the variation coefficient of the Normalized Vegetation Index (CV NDVI)
The variation coefficient of the Normalized Vegetation Index (CV NDVI) was calculated in each pixel - of 250 m resolution - from MODIS MOD13Q1 satellite images of the period 2010 to 2014, available at https://lpdaac.usgs.gov. The average values of the CV NDVI in each segment (1.8 x 2 km2 see in the main paper the section “Density surface model (DSM)”) and in each cell (4 km2) of the prediction grid (see in the main paper the section “Abundance and variance estimation”) were calculated. The map of the spatial variation of the CV NDVI was constructed (Fig. SI.1.1a) and the boundaries of the vegetation units of Península Valdés - defined by Bertiller et al. (2017; Fig. SI.1.1b) - were superimposed (Fig. SI.1.1a). Then, the mean NDVI CV was calculated in each vegetation unit (Table SI.1.1). The behavior of the variable in each stratum was visualized by the 'box-plot' chart (Fig. SI.1.2), while the significant differences were evaluated by means of Wilcoxon rank sum test (Table SI.1.1).
Concurvity measures between smooth terms
As we described in the article, we evaluated concurvity measures between smooth terms throughout the model fitting procedure. Here we presented the pairwise concurvity measures by three related indices (worst, observed and estimated) for the base model of the Tweedie response distribution (Tables SI.2.1, SI.2.2 and SI.2.3), and for the final model selected (Tables SI.2.4, SI.2.5 and SI.2.6).
Spatial autocorrelation in the residuals
Spatial autocorrelation in the residuals was evaluated using the ‘dsm.cor’ function of the‘dsm’ package. As described in the article, the correlogram show a small amount of spatial autocorrelation in the residuals (Fig. SI3.1). The confidence interval increased in width as the number of lags increased.