Variable Time Scale Rainfall Disaggregation Using Artificial Neural Networks

Variability of precipitation is the primary factor driving land surface hydrological processes. In order to accurately understand many land surface hydrological processes such as infiltration and Hortonian runoff, an accurate and fine time scale description of precipitation is necessary. In many cases, this is neither time nor cost effective. An alternative to maintaining fine time scale networks of rain gauges is to disaggregate records from gauges with coarser time steps. This research developed disaggregation methods for converting daily rainfall records into hourly records and hourly records into fifteen-minute records using artificial neural networks (ANNs). Artificial neural networks have been successfully utilized in the past in many different areas of natural science including climate change, seismology, and groundwater remediation. More recently, ANNs have been developed to disaggregate hourly rainfall records into fifteen-minute records. This study extends prior research in the use of ANNs by examining their performance when disaggregating into a range of time scales. The performances of the ANNs developed to account for seasonal variability of rainfall in west-central Florida were compared to those of the ANNs that did not account for seasonal variability of rainfall. Network architecture for each feed-forward, backpropagation ANN was developed and optimized. The performances of the final ANNs were compared to determine if the fractal structure of rainfall was conserved across the two time scales studied.