Abstract
Gridded livestock distribution datasets have been produced for several years and are used in various fields, including epidemiology, livestock impact assessment, and territory management. Those datasets are based on census conducted at national/sub-national scale which are then downscaled using machine learning algorithm and relevant spatial-explicit environmental predictors. The most known dataset of livestock disaggregated observations is the Gridded Livestock of the World (GLW), which produces global maps of livestock density at 10 km spatial resolution for several livestock species. Though this spatial resolution can be appropriate to describe livestock distribution at the global scale, it inherently leads to a coarse representation of breeding species density for smaller territories such as the Caribbean Islands. In this study, we propose an adaptation of the GLW methodology that accounts for the spatial autocorrelation in observed cattle distribution, thereby better capturing the specific characteristics of geographically limited areas such as the Guadeloupean archipelago. Cattle census data were collected for the 32 municipalities of the archipelago and associated to environmental predictors derived from remote sensing and land cover datasets. Together with the Random Forest (RF) algorithm used in the standard GLW methodology, we tested the performance of a Geographical Random Forest (GRF), a novel methodology allowing for taking into account the spatial autocorrelation of the response variable. The GRF algorithm demonstrated significantly better performance compared to the RF algorithm, albeit with longer processing times, and allowed us producing cattle distribution maps for the entire Guadeloupe archipelago at a spatial resolution of 225 m using both algorithms. The approach developed holds potential for application to other small territories, including other islands in the Caribbean.