Abstract
Rainfall disaggregation in time can be useful for the simulation of hydrologic systems and the prediction of floods and flash floods. Disaggregation of rainfall to timescales less than 1 h can be especially useful for small urbanized watershed study, and for continuous hydrologic simulations and when Hortonian or saturation-excess runoff dominates. However, the majority of rain gauges in any region record rainfall in daily time steps or, very often, hourly records have extensive missing data. Also, the convective nature of the rainfall can result in significant differences in the measured rainfall at nearby gauges. This study evaluates several statistical approaches for rainfall disaggregation which may be applicable using data from West-Central Florida, specifically from 1 h observations to 15 min records, and proposes new methodologies that have the potential to outperform existing approaches. Four approaches are examined. The first approach is an existing direct scaling method that utilizes observed 15 min rainfall at secondary rain gauges, to disaggregate observed 1 h rainfall at more numerous primary rain gauges. The second approach is an extension of an existing method for continuous rainfall disaggregation through statistical distributional assumptions. The third approach relies on artificial neural networks for the disaggregation process without sorting and the fourth approach extends the neural network methods through statistical preprocessing via new sorting and desorting schemes. The applicability and performance of these methods were evaluated using information from a fairly dense rain gauge network in West-Central Florida. Of the four methods compared, the sorted neural networks and the direct scaling method predicted peak rainfall magnitudes significantly better than the remaining techniques. The study also suggests that desorting algorithms would also be useful to randomly replace the artificial hyetograph within a rainfall period.