Information about spatial variations of the water table that occur
throughout catchments is useful to infer large scale flow patterns, but
conventional mapping using piezometric data is invasive, slow, and
expensive. Water flow in the subsurface generates an electrical current
called the streaming current. The resulting self-potential (SP)
(electrostatic) signals can be measured non-intrusively, quickly, and
inexpensively-at the ground surface. We considered two conceptual models
to relate SP signals to the water table. The ``infiltration model''
relates SP signals to the thickness of the vadose zone, while the
``water table model'' relates SP signals to the distribution of the
water table in unconfined aquifers. These models are first calibrated
against field data before a Bayesian method is applied to update a
kriged map of the water table obtained from piezometric observations
using a kriged SP map. The estimated water tables based on the two
conceptual models were combined into one final model using concepts from
Bayesian Model Averaging. The method was applied to a small agricultural
catchment (similar to 1 km(2)) in southern France. The posterior water
table estimates were improved mainly on the slopes surrounding the
basin. Understanding the variations in the water table on these slopes
is important because infiltration in these areas feed the basin with
groundwater flow. The Bayesian framework is useful to avoid
overconfident predictions when using SP data in hydrogeological
estimation because it provides realistic uncertainty estimates. (c) 2006
Elsevier B.V. All rights reserved.