Modelling the distribution of domestic ducks in Monsoon Asia

Abstract

Domestic ducks are considered to be an important reservoir of highly pathogenic avian influenza (HPAI), as shown by a number of geospatial studies in which they have been identified as a significant risk factor associated with disease presence. Despite their importance in HPAI epidemiology, their large-scale distribution in monsoon Asia is poorly understood. In this study, we created a spatial database of domestic duck census data in Asia and used it to train statistical distribution models for domestic duck distributions at a spatial resolution of 1 km. The method was based on a modelling framework used by the Food and Agriculture Organisation to produce the Gridded Livestock of the World (GLW) database, and relies on stratified regression models between domestic duck densities and a set of agro-ecological explanatory variables. We evaluated different ways of stratifying the analysis and of combining the prediction to optimize the goodness of fit of the predictions. We found that domestic duck density could be predicted with reasonable accuracy (mean RMSE and correlation coefficient between log-transformed observed and predicted densities being 0.58 and 0.80, respectively), using a stratification based on livestock production systems. We tested the use of artificially degraded data on duck distributions in Thailand and Vietnam as training data, and compared the modelled outputs with the original high-resolution data. This showed, for these two countries at least, that these approaches could be used to accurately disaggregate provincial level (administrative level 1) statistical data to provide high resolution model distributions.

Fig. 1

Stepwise integration procedure of independent variables in regressions models: Each step of the model consists of integrating a pair of predictor variables (a variable and its quadratic term) that leads to the best AIC improvement of the model among the different predictors. This operation is successively repeated for each of the pairs of predictors that have not been selected previously. A new variable is thus added to the regression model after each iteration. The process stops under two conditions i) the improvement in the AIC criterion of two successive models is inferior to 1%, ii) the number of unique points within an Agro-Ecological Zone is insufficient given the number of unique points needed by variable (15 points, for linear and squared term integration).

Fig. 2

RMSE and correlation coefficients obtained with different stratifications (top left and right) and with the different methods for combining their predictions (bottom left and right). Different letters denote significantly different means at the p = 0.05 level (One-way ANOVA; RMSE: F_10,264 = 308.8, p < 0.001; COR: F_10,264 = 213.71, p < 0.001).

Fig. 3

Maps of observed duck density in data pooled from national statistics (top), predicted values using the Livestock Production Systems as stratification layer (middle), and variability of the predictions expressed as coefficient of variation (bottom)

Fig. 4

RMSE and coefficient of correlation of predictions vs. observed data grouped by country using the Livestock Production Systems as stratification layer.

Fig. 5

Predictions in the Mekong obtained with different levels of input data: administrative level 3, level 2, level 1, level 0, and no data.

Fig. 6

RMSE and Correlation coefficient in the Mekong, by administrative level of training data.