Table 1.
Summary of predictor variables.
Fig 1.
Illustration of training polygons (grey polygons in a and c) and test polygons (grey polygons in b and d) used to evaluate the goodness of fit with the downscaling (a—b) and gap-filling (c—d) methodology.
Table 2.
Training data and factors tested in the different evaluations.
Fig 2.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling, using the stratified regression modelling method, density as dependent variable, the full set of predictor variables, the 1 km resolution models, and varying the number of bootstraps. The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 3.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling, using the stratified regression modelling method, density as dependent variable, the full set of predictor variables, 10 bootstraps, and varying the spatial resolution of the modelling process (1 km: 0.0083333 decimal degrees resolution; 10 km: 0.083333 decimal degrees resolution). The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 4.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling, using density as dependent variable, the full set of predictor variables, 10 bootstraps, 1 km resolution modelling and varying the modelling method (SR: Stratified regression corresponding to GLW2; RF: Random Forest). The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 5.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling, using Random Forest as modelling method, the full set of predictor variables, 10 bootstraps, 1 km resolution modelling and varying the dependent variable (Dn: suitability-corrected density corresponding to GLW2; Pc: number of animals per capita). The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 6.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling, using Random Forest as modelling method, density as dependent variable, 10 bootstraps, 1 km resolution modelling and varying the set of predictor variables variable (Ap: all predictors corresponding to GLW2; Fp: reduced set of predictor variables). The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 7.
Correlation coefficients between log-transformed observed and predicted densities evaluated through downscaling (top) and gap-filling (bottom) for both the GLW2 methodology (stratified regression modelling, density as dependent variable, 25 bootstraps, 1 km resolution modelling, full set of predictor variable) and the proposed GLW3 methodology (random forest modelling, density as dependent variable, 10 bootstraps, 1 km resolution modelling, reducted set of predictor variables). The correlation coefficient is estimated by breaking down evaluation polygons by their size, and for both species and continents.
Fig 8.
Predicted distribution of cattle in Africa (top) and chickens in Asia (bottom) from the GLW2 methodology (a and d), the proposed GLW3 methodology (b and e) based on random forest, and of GLW3 methodology using animals per capita instead of absolute density (c and f).
The data used to produce these maps were all from public sources, and the country limit data are from the FAO Global Administrative Unit Layers (GAUL) database.