INTELLIGENT EXPLORATION FOR SHALLOW GROUNDWATER IN FRACTURED ROCK SYSTEMS

It is not possible to predict well yield in the semi-arid climate and fractured crystalline rocks of northeastern Brazil by traditional groundwater modeling. For this reason, an alternative paradigm is used in which the relations associated with a sparsely populated set of hydrogeologic data (electrical conductivity, geology, temperature, and well yield) and airborne geophysical measurements (electromagnetic, magnetic, and radiometric) are found using the self-organizing map technique. Selected variables exhibiting a statistically significant relation to well yield are then used with symbolic regression to discover predictive models based on evolutionary heuristics. An objective function that simultaneously maximizes fitness and minimizes root-mean-squared error identifies the best well yield models evolved from unprocessed, processed, and mixed sets of the airborne measurements. All models exhibit unbiased predictions that are within a few percent of the known well yield observations. Estimates of nonlinear uncertainty limits reveal that models evolved from processed measurements may result in a biased prediction at low well yields (< 1 m³ hr^-1). For a particular combination of model and measurement type, the computed range of prediction uncertainty generally is reduced when increasing the number of measurement variables. The best well yield predictor is a function of three unprocessed airborne electromagnetic measurement variables. These findings suggest that the combination of data mining, knowledge discovery, airborne electromagnetic measurements, and predictive analysis may provide a low-cost alternative to traditional modeling in challenging groundwater environments, such as semi-arid and fractured rock aquifers in northeastern Brazil.