Assessment of hydrological drought in Northern Ontario using standardized streamflow index

Abstract

Momentum is gaining for the use of a Standardized Streamflow Index (SSI) for the assessment of hydrological drought, which is an extension of the well-established and popularly used meteorological drought index, Standardized Precipitation Index (McKee et al, (1993). Drought characterization using SSI consists of various data treatments to transform streamflow data to standard normal Z-score values, where the mean is zero and the standard deviation is one. The underlying assumption when using Z-score in early warning systems and drought monitoring programs is that the scores associated probabilities may be represented by the normal (Gaussian) distribution. Applying data treatments that fail to achieve the conditions of normal distribution results inaccurate SSI drought assessments. There is an opportunity in Northern Ontario to conduct SSI drought assessments in early warning systems and drought monitoring programs using the abundance of long and continuous streamflow records; however, widespread applications would only be achieved if the resulting assessments are accurate, equitably and reliable. The data treatments investigated in the thesis include: Untreated and treated (i.e., Log Normal Transformation, and the Fitted Gamma Distribution). Monthly assessments for a total of 40 rivers from across Northern Ontario with an average record length of ≈ 45 years were utilized in the analysis. Historical droughts sample periods 1976/77 and 1998 were used in evaluations as they were determined to pertain to the well-defined regional drought events with confirmation of historically significant impacts in addition to containing seasonal influences. The Log Normal Transformation followed by the Fitted Gamma Distribution exhibited the most consistency in not rejecting normality assumption of the data set when using the Shapiro Wilks and the Anderson Darling tests. These tests also show the best performance of the Fitted Gamma Distribution for the winter season (November to May) and of the Log Normal Transformation for the summer season (June to October). Using these two data treatments for the respective seasons permitted the assumption of normality to be rejected 15.7% of the time. The use of best performing data treatments and removal of data for the month of March further reduced the assumption of normality being rejected only 12.0% of the time. In evaluating the frequency of occurrence for the severe and extreme Z-score values (- 1.64 and -2.33), the Fitted Gamma Distribution followed by the Log Normal Transformation demonstrated the most acceptable scoring distribution. Untreated data performed less efficiently. All data treatments underperformed for the expected frequency at the extreme Z-score of -2.33 (i.e., the extreme drought). The threshold level Q80 was determined to be the most appropriate trigger of monthly assessments for Northern Ontario because it was found to equitably delineate severe regional drought perceived by the respectively impacted sectors. The threshold level Q80 represents these sectors demands by identifying the point where streamflows are exceeded 80% of time. Since it is representative of the impacted sectors; it is used to assess the performance of SSI. Significant differences were recorded when comparing Q80 to the typical moderate drought classification with a Z-score equal to - 1. Untreated data tacitly refers to the assumption of the normal distribution of the monthly flows and correspondingly of SSI values. Thus, the analysis based on the assumption of the normality of the monthly SSI data turned out to be less reliable. All the data treatments underperformed for the expected frequency at extreme Z-score of -2.33. However, when applying Q80 to the theoretical equivalent Z-Score of - 0.84; identical results were achieved 76% and 66% of the time, respectively for the Log Normal Transformation and the Fitted Gamma Distribution. For the case of Q80, the average difference for the drought initiation, termination, and duration were found to range approximately 1.3 months with the Log Normal Distribution to 3.0 months for the lesser performing data treatments, such as the Untreated data.