Modeling Streambed Vertical Hydraulic Conductivity, Water Quality Pollutants, and Best Management Practices Using Machine Learning and the Soil and Water Assessment Tool

Spatio-temporal variability of natural and man-made watershed characteristics controls hydrological processes. With respect to quantity and quality, water resources management in a watershed involves understanding these variations in order to make better water policy changes and for the implementation of best management practices (BMPs). Two watersheds were studied at different spatial and temporal scales to model streambed vertical hydraulic conductivity (Kv), water quality pollutants, and BMPs. At Frenchman Creek, a new approach was used to develop pedo-transfer functions (PTFs) for simulating the effects of complex sediment routing on Kv variability across ten sites in multiple stream orders using watershed characteristics as predictors. Coupling Kv and drainage area as a response variable reduced the fuzziness in selecting the best PTFs. Four variograms were also used to determine the spatial distribution of Kv. Cross-validation results showed differences in the median absolute deviation of the cross-validation residuals. Geostatistical analysis showed that using the ten geometric means of the sites performed better than using either the Kv values from 93 permeameter tests (at least 9 tests per site) or ten Kv values from the middle transects and center permeameters. BMPs were simulated at hotspots within the Big Sandy Creek to determine reductions in pollutant loads, and to determine if water-quality standards would be met at the watershed outlet. With scaled-down acreage based on the proposed implementation plan, a combination of filter strips, grassed waterways and atrazine application rate reduction would most likely yield measureable improvement both in the hotspots and at the outlet. To predict daily Escherichia coli concentrations in two cascading dams at US Meat Animal Research Center (a subwatershed of Big Sandy Creek), a novel approach that integrated hydro-climatic variables with animal density and grazing pattern was used to increase prediction accuracy. Models were developed using regression, artificial neural network, and adaptive neuro-fuzzy inference system (ANFIS). Cross-validation and model performance results indicated that ANFIS models resulted in fewer errors compared to other models. ANFIS models have the potential to be used for predicting E. coli concentration for monitoring programs and for implementing BMPs for grazing and irrigation during the growing season