Evaluation of L-Moments and Bulletin 17B Methods for Flood Frequency Analysis
Research Idea Scope
Standard practice in the U.S. within hydrology is to assume a Log-Pearson III probability distribution for predicting flood flows. This practice came about due to the computational resources that were available in the past. Now, standard desktop computers have the capacity to analysis gaging station data and watershed characteristics to cluster (regionalize) stations, select the best fit probability distribution for each cluster, and regress the watershed characteristics against the predicted peakflows. Existing research has shown that the Log Pearson III probability distribution is not the best fit for rural, unregulated streams in two states (Arkansas & Tennessee).
By selecting 6 states in different climatic and physiographic regions and comparing the results from the L-Moments method with the traditional method (Bulletin 17B), it would be possible to show which method is most appropriate for different conditions or if one method is superior to another.
To make this analysis the following data is required (all data is available on-line through USGS, NRCS, & EPA websites):
1) annual peak flow data (10 plus years) from gaging stations on rural, unregulated streams,
2) digital elevation maps (DEMs) for each watershed to be analyzed,
3) watershed characteristics (land use, vegetation, soil type) for each watershed,
4) climatic data for each watershed (precipitation, temperature).
The following analyses of the data will be performed:
1) The peakflow data for each watershed would be analyzed for outliers (discordant data points) and discrepancies corrected.
2) Watershed characteristics will be determined for each watershed (drainage area, basin slope, basin length, basin shape factor, mean basin elevation, mean annual precipitation, land use land cover (LULC), and hydrologic soil group.
Using traditional method (Bulletin 17B) the following analyses would be conducted:
1) Regionalization of watersheds by physiographic regions in the state and analysis of site and at-site characteristics to show statistical difference of the regions.
2) Determination of estimated peakflows using Log-Pearson III probability distribution.
2) Regression of site and at-site characteristics against the Log Pearson III estimated peakflows from gaging stations within each region to develop regression runoff equations for each region within the state.
Using the L-Moments method, the following analyses would be conducted:
1) A regionalization analysis of the at-site and site characteristics of each watershed will be performed to determine homogeneous regions (geographic or watershed characteristics).
2) The peak flows of each gaging station will be analyzed to determine the best fit probability distribution (log-Pearson III, log normal, generalized extreme value, generalized logistic, generalized Pareto, Kappa, or Wakeby) for each region.
3) Regression of site and at-site characteristics against the estimated peakflows within each homogeneous region (cluster) to develop regression runoff equations for each region within the state.
Analysis and comparison of the accuracy of the two methods will be determined using ordinary least squares and generalized least squares methods on the predicted versus actual peakflow data. The comparison will be conducted state by state, by site characteristics, and for all six test states.
Urgency and Payoff
To properly design bridges, culverts and channels, every 10 years each state conducts an analysis to develop regression runoff equations for flood prediction on rural, unregulated streams. The traditional method (Bulletin 17B) is out-of-date for the computational capacity of today’s computers and the digital watershed data available. Updating and improving the method to be used by all states could greatly improve flood predictions and could save state DOTS hundreds of millions of dollars in rebuilding under-designed bridges, culverts, and channels.
Findlay G. Edwards, PhD, PE, D.WRE, University of Arkansas