GEOSTATISTICAL METHODS FOR PREDICTION OF SPATIAL VARIABILITY OF RAINFALL IN A MOUNTAINOUS REGION

摘要

Reliable estimation of rainfall distribution in mountainous regions poses a great challenge not only due to highly undulating surface terrain and complex relationships between land elevation and precipitation, but also due to non-availability of abundant rainfall measurement points. Prediction of rainfall variability over mountainous islands is a logical step towards meaningful land use planning and water resources zoning. In this context, geostatistical techniques were developed for mapping the rainfall variability over the island of St. Lucia in the Caribbean, using the elevation information extracted from a Digital Elevation Model (DEM) and long-term mean monthly rainfall (MMR) data of 40 raingauge stations spread over 616 km{sup}2. The ordinary co-hriging (OCK) and collocated co-kriging (CCK) methods of interpolation were applied for the standardized rainfall depths associated with elevation, as the primary variate, and the surface elevation values as the secondary variate. The best semivariogram model algorithm generated, using either of the above co-kriging (CK) methods, was used to predict standardized values for the elevation points extracted from the DEM for which the rainfall depths were not known. The predicted values were further destandardized to generate the rainfall depth at the unmeasured locations. Ordinary kriging (OK) was then performed for the destandardized and observed rainfall depths to generate the prediction map of MMR over the entire island. These sequential steps were repeated for the MMR data of all twelve months to generate rainfall prediction maps over the island. The spherical semivariogram model fit well (0.84 < R{sup}2 < 0.98) for both the OCK and OK methods. The cross-validation error statistics of OCK presented in terms of coefficient of determination (R{sup}2), kriged root mean square error (KRMSE), and kriged average error (KAE) were within the acceptable limits (KAE close to zero, R{sup}2 close to one, and KRMSE from 0.55 to 1.45 for 40 raingauge locations) for most of the months. The exploratory data analysis, variogram model fitting, and generation of MMR prediction map through kriging were accomplished through use of ArcGIS and GS+ software.