Methods and data sources for spatial prediction of rainfall

This chapter reviews the data sources, both rain gauge (point) data sources and remote sensing imagery sources (visible, infrared, thermal and microwave) used for producing precipitation maps, and then shows "in action" a number of mechanical and stochastic spatial prediction methods that can be used to generate maps of rainfall intensity. Special focus was put on using geostatistical techniques implemented in the R environment for statistical computing (via stats, gstat and geoR packages). The spatial prediction methods are illustrated using a small case study (97 points in the area used for geostatistical analysis were obtained from the NCDC Global Summary of Day) covering the scanning radius of the Bilogora weather radar in Croatia (366 daily images), and the national rain gauge network in Italy (1901 stations). The results show that the rainfall estimated using ground based radar can be of variable accuracy. The radar images can carry many artifacts, especially at high distances from the ground radar, so that the correlation with ground measurements is often marginal. Daily rainfall intensity is commonly skewed toward small values with many zeros, hence rainfall data at shorter time intervals is suitable for modeling using Zero-inflated regression models. The chapter contains code snippets showing how to implement various prediction techniques from local trend surfaces to ordinary kriging, zero inflated regression models, and regression-kriging in R.