Abstract: We modified a self-organizing map (SOM), a clustering artificial neural network, using variogram analyses to incorporate the spatial and temporal auto-correlation that exists in many surface and subsurface environmental datasets. The SOM reduces the dimensionality and clusters the data. The SOM is particularly effective with multiple data types (e.g. continuous, discrete and categorical variables). The standard SOM algorithm iteratively updates connection weights between the input parameters and the two-dimensional output mapping over a specified region of the estimation field. The method accounts for the anisotropy found in geologic and hydrologic datasets. The algorithm is tested on a unique dataset collected from a slab of Berea sandstone (1 m by 0.4 m). Air permeability, electrical resistivity and compressional wave velocity were measured on a 3 mm rectangular grid. Sparse testing data were drawn randomly from this exhaustive dataset for validating the new computational methods. We apply the method using biogeochemical data collected along a transect between the Vermont and New York shorelines of Lake Champlain, to demonstrate the ability to discriminate between different functional zones in the lake. This clustering method could be applied to a variety of terrestrial, aquatic, or subsurface biogeochemical or geophysical problems. Considering spatial auto-correlation in delineating regions or zones in environmental systems creates more accurate estimations.