Geostatistical model for correlating declining groundwater levels with changes in land cover detected from analyses of satellite images

A combination of satellite-data processing and geostatistical spatial modeling is examined to investigate long-term changes in groundwater levels within a catchment basin and identify the causes of the changes. The data processing was used to obtain an accurate image classification of land covers while the geostatistical modeling was used to interpolate residual components of groundwater levels by removing identified trends from the original data. Our case study focused on the Kumamoto Plain, situated in central Kyushu, southwest Japan, where all of the water supply is derived from groundwater. A Linear Spectral Mixture was adopted to the 10 sub-scene images of Landsat 5 Thematic Mapper (TM) and Landsat 7 Enhancement Thematic Mapper Plus (ETM+) sensors covering the study area for a 16-year period (1987–2003); this proved to be highly effective in detecting declining amounts of groundwater-recharging material. Descriptive statistical parameters from the groundwater-level data were integrated with the process of identifying, characterizing, and removing the trend components that are closely correlated with rainfall. Variogram modeling and ordinary kriging were used to construct a spatial model of the residual groundwater-level data. Finally, a multivariate model is proposed to estimate the degree of reduction in groundwater levels. The correlation coefficients between the observed and estimated residual levels by the multivariate regression model at 14 wells for the cross-validation range from 0.95 to 0.98. This demonstrates the applicability of the model to both the local areas around well sites and across the entire study area. To deepen the regression result, the Tukey honestly significantly different statistical test was applied. The test clarified that changes in land cover from rice field to urban areas have the greatest impact on reductions in groundwater within the recharge area.