文摘
This study focuses on the spatial distribution of mean annual and monthly precipitation in a small island (1128 km2) named Martinique, located in the Lesser Antilles. Only 35 meteorological stations are available on the territory, which has a complex topography. With a digital elevation model (DEM), 17 covariates that are likely to explain precipitation were built. Several interpolation methods, such as regression-kriging (𝖬𝖫𝖱𝖪, 𝖯𝖢𝖱𝖪, and 𝖯𝖫𝖲𝖪) and external drift kriging (𝖤𝖣𝖪) were tested using a cross-validation procedure. For the regression methods, predictors were chosen by established techniques whereas a new approach is proposed to select external drifts in a kriging which is based on a stepwise model selection by the Akaike Information Criterion (AIC). The prediction accuracy was assessed at validation sites with three different skill scores. Results show that using methods with no predictors such as inverse distance weighting (𝖨𝖣𝖶) or universal kriging (𝖴𝖪) is inappropriate in such a territory. 𝖤𝖣𝖪 appears to outperform regression methods for any criteria, and selecting predictors by our approach improves the prediction of mean annual precipitation compared to kriging with only elevation as drift. Finally, the predicting performance was also studied by varying the size of the training set leading to less conclusive results for 𝖤𝖣𝖪 and its performance. Nevertheless, the proposed method seems to be a good way to improve the mapping of climatic variables in a small island.