- Title
- Nonhomogeneity in Eastern Australian flood frequency data: identification and regionalisation
- Creator
- Micevski, Tom
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2007
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Flood frequency data from the eastern Australian states of New South Wales (NSW) and Queensland (Qld) were investigated to determine the magnitude and extent of multidecadal variability (nonhomogeneity) in flood risk. Some flood data from NSW were found to be systematically in error because daily-read discharges were used instead of instantaneous peak discharges. A new approach, based on the method of maximum likelihood, was developed to overcome the potential artefacts introduced by the use of daily-read data in flood frequency analysis. However, it was shown for flood data typical of NSW, the treatment of daily-read data as instantaneous peaks did not introduce sufficiently large quantile bias and loss of mean-squared-error performance to warrant use of the new estimation method. The flood data were stratified by Interdecadal Pacific Oscillation (IPO) value and flood frequency analyses performed on the IPO-stratified flood data --- the IPO is a climate index of Pacific Ocean sea surface temperature anomalies, which displays variability on a long-term (multidecadal) time scale. The IPO was found to modulate the flood risk in NSW and southern Qld, with flood quantiles being increased, on average, by approximately 1.7 times during IPO-negative epochs, whereas little effect was detected for sites in northeast Qld located approximately north of the Tropic of Capricorn. The IPO modulation (nonhomogeneity) of flood risk has great practical significance --- the use of at-site flood data with inadequate coverage of both IPO epochs may result in biased estimates of long-run flood risk. A Bayesian regional flood model framework, based on hierarchical modelling concepts, was developed to overcome the possible bias in long-run flood risk associated with a nonhomogeneous flood record. Importantly, the model allows for the consideration of intersite correlation. Bayesian methods were used to enable a rigorous treatment of uncertainty in the flood regionalisation problem. The Gibbs sampler was used to infer uncertainty in the regional model parameters, while importance-sampling-based procedures were developed to compute the Bayesian predictive distribution and the posterior distribution of quantiles at a new site, which may be ungauged or gauged. This represents the first truly-general Bayesian solution for combining regional and gauged information in flood frequency analysis. The flood data from eastern Australia were partitioned into four regions and analysed using the Bayesian hierarchical regional model. The (correlated-site) regional model found significant differences (at the 10% level) in the regional means for the two regions in NSW. The regional standard deviations showed significant differences for the two regions in Qld, but these differences were opposite in sign, with IPO-positive standard deviations being greater than IPO-negative values. The equivalent gauged length provided by the regional model had a maximum of 4 to 8 years. There is a large overlap in the probability limits between the IPO-positive and IPO-negative regional distributions for a flood quantile. However, because the regional model errors for the IPO phases are highly correlated, the difference in IPO-positive and IPO-negative quantiles is likely to be significant. For larger return periods, the opposite-in-sign differences in the regional mean and standard deviation may reduce the resultant IPO-related differences (in discharge). The value of the regional model was demonstrated by pooling the regional information with the information in short gauged records at selected sites in each region. The pooled at-site flood frequency distribution provided substantial improvements over the gauged record alone (in terms of prediction limits and bias). Indeed, this was especially evident in a situation where a shortened (10-year) gauged record was found inconsistent with the true (long-run) gauged record --- the shortened gauged record consisted mainly of years from the IPO-positive epoch. These results suggest that the use of the regional model may protect against bias in long-run flood risk at sites with short records sampled largely in one IPO epoch.
- Subject
- regionalisation; hierarchical model; Bayesian analysis; flood frequency analysis; nonstationarity; ungauged site
- Identifier
- http://hdl.handle.net/1959.13/937245
- Identifier
- uon:12530
- Rights
- Copyright 2007 Tom Micevski
- Language
- eng
- Full Text
- Hits: 1081
- Visitors: 1408
- Downloads: 402
Thumbnail | File | Description | Size | Format | |||
---|---|---|---|---|---|---|---|
View Details Download | ATTACHMENT01 | Abstract | 132 KB | Adobe Acrobat PDF | View Details Download | ||
View Details Download | ATTACHMENT02 | Thesis | 1 MB | Adobe Acrobat PDF | View Details Download |