Hidden Markov model-based probabilistic assessment of droughts
Abstract
Droughts are one of the most expensive and least understood natural disasters. Existing methods of drought assessment use drought indices that rely on subjective thresholds, and hence cannot be universally applied across different climatic regions. Many of the popular methods of drought characterization ignore temporal dependence in the data. In addition, current drought indices are not amenable to probabilistic treatment which is essential for quantifying model uncertainties in drought classification. This study explores the use of graphical models, specifically hidden Markov models (HMMs), to develop a new drought index that overcomes some of the limitations of current drought indices. This HMM-based drought index (HMM-DI) does not require specification of subjective thresholds. The parameters of the HMM model are determined from historical data. The HMM-DI reveals new insights into the frequency and severity of droughts and their spatio-temporal variations. The HMM-DI is applied to monthly precipitation and stream flow data over Indiana, and to gridded precipitation data over India. The results suggest that HMM-DI can be a promising alternative to conventional drought indices. However, available data records do not support a full HMM model with all the classifications adopted by the community as reflected in the U.S. Drought Monitor. This study suggests that simpler models may be more appropriate if memory is to be preserved.
Degree
M.S.C.E.
Advisors
Govindaraju, Purdue University.
Subject Area
Hydrologic sciences|Climate Change|Water Resource Management
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.