Evaluation of ordinary cokriging and artificial neural networks for optimizing rainfall estimate using stage III NEXRAD precipitation surfaces and rain gauge measurements

The deployment of the National Weather Service Weather Surveillance Radar-1988 Doppler (WSR-88D) has provided an improved tool for monitoring real-time areal mean precipitation spatial distribution (4-km resolution) for hydrometeorological modeling. Unfortunately, a number of factors introduce discrepancies between radar precipitation estimates and actual precipitation at the Earth's surface. In this project, a pilot study was performed by making two types of statistical analyses to describe the correlation between SEMCOG rain gage values and stage III NEXRAD: (1) the agreement of occurrence of precipitation between the two sources and the magnitude of error in precipitation when there is disagreement of occurrence, and (2) the error in magnitude between rain gage measurement and NEXRAD estimate when they both register a precipitation amount. These analyses provided a basis of justification for using models to improve the correlation between the two sources of precipitation measurements. Twenty-two daily precipitation events (partial and full rainfall coverage) during the months of May through September in 1999 and 2000 were selected to estimate the precipitation using Stage III NEXRAD data and SEMCOG rain gage measurements. Artificial neural network and ordinary cokriging models were evaluated by the performances of improved precipitation estimates. The best performing model, the ANN model, significantly improved the accuracy of the radar-derived precipitation surfaces (Average correlation coefficient was improved from 0.61 to 0.76).; The ANN model was applied to improve the precipitation estimate of the entire state of Michigan. The 16-km NEXRAD grid size was used for improving the Stage III NEXRAD data for the entire state of Michigan. Six daily precipitation events (partial and full rainfall coverage) were selected to optimally estimate the precipitation by combining Stage III NEXRAD data with forty-two National Weather Service Fisher & Porter rain gages distributed around the Michigan. The results showed that the Stage III NEXRAD precipitation surfaces were fairly improved (Average correlation coefficient was improved from 0.72 to 0.87).