Deriving Spatially Distributed Precipitation Data Using the Artificial Neural Network and Multilinear Regression Models

Abstract

Precipitation is the primary driver for hydrologic modeling. Because hydrologic models often require long-term, spatially dis- tributed precipitation data sets for calibration and validation, a novel approach was developed to generate spatially distributed precipitation data using an artificial neural network (ANN) for the periods when Next-Generation Weather Radar (NEXRAD) data are either unavailable or the quality of the NEXRAD data is not good. The multilinear regression (MLR) technique was also evaluated for completeness. The study's focus was the Saugahatchee Creek watershed in southeast Alabama. In the study area, thewet seasons are dominated by frontal precipitations, whereas the dry seasons primarily contain patchy, convective thunderstorms. The basic approach was to train and validate the ANN and MLR models using recent NEXRAD and rain gauge precipitations, and then use the trained model with the rain gauge precipitation data to generate past, spatially distributed precipitation estimates at the NEXRAD grid locations. For the testing period, the ANN-simulated wet season precipitations in all the NEXRAD grids had a Nash-Sutcliffe efficiency greater than or equal to 0.72 and a mass balance error less than or equal to 14%. The same model performance parameters were 0.65 and 17%, respectively, for the dry season. The MLR model did not perform as well as the ANN model. For the MLR model, the wet season mass balance error ranged from 13-48%, whereas the dry season mass balance error ranged from 0.1-36% on the testing data set. An uncalibrated soil and water assessment tool model was used to assess the improvements in stream flow simulations with the ANN-simulated spatially distributed precipitation data. The stream flow simulations using ANN-generated, spatially distributed precipitations were closer to the observed stream flows relative to stream flows generated using the rain gauge precipitations. Overall, the results suggest that the method developed in this study can be used to generate past, spatially distributed precipitations at NEXRAD grid locations. DOI: 10.1061/(ASCE)HE.1943-5584.0000617. © 2013 American Society of Civil Engineers. CE Database subject headings: Precipitation; Neural networks; Hydraulic models; Hydrologic data. Author keywords: Precipitation; Next-Generation Weather Radar (NEXRAD); Artificial neural network; Multilinear regression; Hydrologic modeling; Soil and watershed assessment tool (SWAT).