Prediction of drought indices based on hydroclimatological data

Drought is a cause that adversely affects the land and water resources as a consequence of the fact that the normal levels of precipitation are considerably below that level, thereby causing deterioration of the hydrological balance. Early indications of droughts can provide valuable information to help mitigate some of its consequences of drought; for instance, decision makers may plan and manage the water resource systems accordingly. The success of drought preparedness and mitigation depends upon timely information on the drought onset and development in time, which can be obtained through continuous drought monitoring using common indices. Over a region, many drought indices have been developed for monitoring and examining meteorological droughts. The intensity and spatial extent of droughts are generally quantied by drought indices. In this study, the estimation of the drought indices with the influence of atmospheric oscillations was aimed. For this purpose, in the first part of the thesis, precipitation stations across Turkey with nearly uniform distributed and continious data have been identified. Thus, the data set for 160 precipitation stations and the homogeneity of monthly precipitation data between 1974 and 2014 were analyzed by Standard Normal Homogeneity, Pettitt, Buishand and Von Neumann Ratio tests. Our decision criterion for a station to be homogeneous was rigidly set as a full confirmation through all of the four tests used in this study. Previous similar studies in Turkey did not adopt such a rigid criteria for the confirmation of homogeneity of precipitation series as applied in this study. Keeping this criterion in mind, all stations were subjected to the four homogeneity tests. As a result of overall evaluation, we determined 44 out of 160 stations to be inhomogeneous since either one or more tests did not confirm. Test-wisely speaking, the results of the SNHT and the Pettitt test showed that a set of 25 out of 160 stations were impacted from inhomogeneity. According to the Buishand test, 13 out of 160 stations were found to be inhomogeneous while the Von Neumann test revealed 22 inhomogeneous stations. Moreover five out of 160 stations possess inhomogeneity characteristic with the respect to the results of all four tests. Since we set the level of significance at 5%; the critical values are 167, 8.10, 8.07, and 1.49 for the Pettitt, SNHT, Buishand, and Von Neumann ratio tests, respectively. It is important to visually inspect the distribution of inhomogeneous stations in order to catch an underlying unexpected specific reason or to justify the required homogeneity correction. For this purpose, we mapped the distribution of inhomogeneous stations depending on applied test method. The Pettitt test revealed 25 inhomogeneous stations as twice as that of the Buishand test (implying that the former exhibits more conservative outcomes than the latter) as both tests are capable of detecting a break point in the middle of a series. It is important to note that all 13 inhomogeneous stations detected by the Buishand test were also comprised in the set of inhomogeneous stations of the Pettitt test. The SNHT, known as a good tool for detecting a break point at the head or end part of a time series, brought out a fact that the reason behind inhomogeneity of 25 stations was a break point appearing mostly in the second half of time series (in particular 1996 or later). It is also noticeable that the year of 2008 was the most frequent break point as happened at six stations. The Von Neumann ratio test, which focusing on all parts of the time series, pointed out to seven inhomogeneous stations that were not caught by the other three tests; namely, Cankiri, Tatvan, Hakkari, Keskin, Ilgin, Baskale, and Milas stations. Among the 44 inhomogeneous stations, five stations (Nallihan, Sinop, Ardahan, Kars, and Mardin) could not pass the four tests. Two stations (Nallihan and Odemis) still remain inhomogeneous after applying Double Mass Curve Analysis (DMCA) to the set of 44 non-homogeneous meteorological stations. Subsequently, we adopted the DMCA to each of 44 inhomogeneous stations to see if any station could possibly turn out to be homogeneous. We drew a double mass curve by setting the annual precipitation total of an inhomogeneous station under consideration at ordinate axis and corresponding an average of precipitation total values of neighboring homogeneous stations at horizontal axis. We used minimum four neighboring stations in this analysis. Our results showed that 42 out of inhomogeneous 44 stations passed all the four analysis tests after being applied to correction based on the DMCA. However, it was not possible for only two stations, namely Odemis and Nallihan, to make them homogeneous. In the last of this section, interpretations were made using meta-data obtained from the Turkish State Meteorological Service to understand the reasons behind the inhomogeneity of a non-homogeneous station. In the second part of the thesis, drought analyzes were carried out using the Linear Genetic Programming-Discipulus software program for the Konya Closed Basin, which was designated as pilot region. In genetic programming which is an evolutionary computational technique, processes are performed by generating a structured representation of the examined system using previous possible solutions and genetic operators. Palmer Drought Severity Indices (PDSI) of 8 precipitation stations and above-mentioned atmospheric oscillation indices were determined as inputs in the 1970-2016 time period of Konya Closed Basin and PDSI estimations were made at 3, 6, 9 and 12 months lead times and the obtained results were interpreted. In the program, monthly observation data were divided into training and test data. 70% of the entire data set was used in the training period, and the remaining 30% was used to test the validity of the model. When the results of the general analysis for the 8 stations are evaluated, it is observed that the scenarios used for all three oscillations in Aksaray and Nigde stations, particularly 3 months lead time drought data are reasonably estimated. In the third part of the study, prediction intervals were obtained by Bootstrap method for future values of PDSI drought indices for 4 stations in Konya Closed Basin. In addition, the effects of the North Atlantic and Arctic Oscillation indices on these intervals have been examined. To obtain out-of-sample prediction intervals, we divide each time series into the following two parts: (i) the first part consisting of a length spanning from January, 1970 to December, 2013 (a total of 528 observations) on which the model construction will be based to calculate 24 steps ahead (monthly), 8 steps ahead (mid-term) and 4 steps ahead (short-term) predictions (ii) and the second part containing the period January, 2014-December, 2015 at which comparison with the observed values will be made. For each series, bootstrap simulations were performed, and we set the significance level to obtain 95% bootstrap prediction intervals for future PDSI values. The calculations were carried out using R 3.3.3. While the effects of these atmospheric oscillation indices were observed in the short- and mid-term drought forecasts, it was found that the monthly drought forecasts had no effect on the prediction intervals of the study area. Our findings showed that the bootstrap method with AR(1) and ARX(1) models resulted in reasonable prediction intervals for future PDSI values so that almost all such values are well covered by the constructed prediction intervals.We might conclude that the bootstrap method produces narrower prediction intervals for the short-term forecasts compared to those constructed for long-term forecasts. This is because the latter is expected to contain more uncertain. The effects of AO are statistically significant only for Aksaray station, similarly the effects of NAO are also significant for Karaman and Seydisehir stations. The overall result suggests that the AO and NAO indexes do not play a crucial role for the monthly PDSI values in Konya basin since the prediction intervals constructed by oscillation indices are not significantly different from the intervals obtained by using only AR(1) model. On the other hand, the significant effects of the oscillation indexes can be readily seen in the short-term and mid-term forecast in Karaman station. We may speculate that the future short-term and mid-term drought values fluctuate in a increasing manner with time and even becomes more evident like a form of linear-trend appearance as we move to larger scale drought duration. It is worthwhile to note that the bootstrap prediction intervals seem to be self-adjusted to exhibit similar behavior having imbedded effects of the oscillation indexes.