Space-Time Cokriging Approach for Groundwater-Level Prediction with Multiattribute Multiresolution Satellite Data

The availability of complete, continuous, and reliable groundwater data is essential for preparing a regional-scale groundwater model. Poor sampling of groundwater level is often encountered. Satellite-derived data can provide possible solution to such a problem. Better coverage and timely sampled data can be easily obtained from satellite-derived data. However, getting direct and exact groundwater-level measurements from satellite data is not possible. Satellite data, hence, are used as latent variables (or secondary attributes) to predict groundwater level. A geostatistical approach (cokriging) was brought into use to predict groundwater level by using the available secondary attributes. The study was carried out in the Indo-Gangetic Basin by using Gravity Recovery and Climate Experiment (GRACE) anomaly data and Tropical Rainfall Measuring Mission (TRMM) precipitation data as secondary attributes. The study was carried out at the finest resolution available. Groundwater-level measurements measurements are available at seasonal scale at point support, and the analyses were performed for the premonsoon season of 2005 and 2008. Four performance indicators—mean absolute error (MAE), bias, root-mean-square error (RMSE), and coefficient of variation of root-mean-square variation [CV(RMSE)]—determined the best result obtained from the analyses. The cokriging analyses were carried out for different neighborhood search radii. Four different neighborhood search radii were chosen for the present study on the basis of range of variogram of the primary attribute: (1) less than the range of variogram; (2) range of variogram; (3) more than the range of variogram; and (4) infinite radius. Different neighborhood radii were chosen to visualize any changes in predictability of the different models prepared. The best prediction was obtained for the premonsoon of 2005 with very low error values at infinite neighborhood search radius [$\mathrm{bias}=1.61\mathrm{m}$; $\mathrm{MAE}=1.77\mathrm{m}$; $\mathrm{RMSE}=2.35\mathrm{m}$; and $\mathrm{CV}(\mathrm{RMSE})=0.32$]. With a neighborhood search radius of less than the range of semivariogram, missing values were obtained for the premonsoon of both 2005 and 2008. The overall prediction was found to be satisfactory, suggesting very good predictability of the geostatistical approach.