Assessment of different methods for estimation of missing data in precipitation studies

Rainfall is an important part of the hydrological cycle. One of the first steps in any hydrological and meteorological study is accessing reliable quality data. However, precipitation data is frequently incomplete. The incompleteness of precipitation data may be due to damaged measuring instruments, measurement errors and geographical paucity of data (data gaps) or changes to instrumentation over time, a change in the measurement site, a change in data collectors, the irregularity of measurement, or severe topical changes in the climate.

The accurate planning and management of water resources depends on the presence of consistent and exact precipitation data in meteorology stations. In cases where it has not been possible to accurately and consistently record precipitation data in a particular time section, it is necessary to estimate the missing precipitation data before applying it in hydrological models. The data in climate studies also face this issue, as well as measurements in the ocean share similar data problems in regard to missing precipitation data (Lyman & Johnson 2008; Abraham et al. 2013; Cheng et al. 2015a, 2015b).

The estimation of missing data in hydrological studies is necessary for timely implementation of projects such as dam or canal construction. This information is extremely valuable in areas that deal with heavy precipitation events and floods. The accurate estimation of the missing data makes a great contribution to accurate assessment of the capacity of flood control structures in rivers and also dam spillovers. It reduces the risk of floods in the downstream of these structures. Abraham et al. (2015) observed precipitation changes in the United States and stated that observations and projections of precipitation changes can be useful in designing and constructing infrastructure to be more resistant to both heavy precipitation and flooding.

Homogeneity and trend tests of data used in hydrological modeling or water resource analysis are essential. Numerous methods have been introduced for estimating and reconstructing missing data. They can be categorized as empirical methods, statistical methods, and function fitting approaches (Xia et al. 1999). Most of these methods derive the missing values using observations from neighboring stations. Selecting appropriate methods for estimating missing precipitation data may improve the accuracy of hydrological models. The literature points to rather arbitrary selection methods for estimating and reconstructing missing data (Hasanpur Kashani & Dinpashoh 2012). Some of the most significant studies involving estimation and reconstruction of missing rainfall data are discussed next.

Xia et al. (1999) estimated the missing data of daily maximum temperature, minimum temperature, mean air temperature, water vapor pressure, wind speed, and precipitation with six methods. They determined that the multiple regression analysis method was most effective in estimating missing data in the study area of Bavaria, Germany. Teegavarapu & Chandramouli (2005) applied a neural network, the Kriging method and the inverse distance weighting method (IDWM) for estimation of missing precipitation data. They demonstrated that a better definition of weighting parameters and a surrogate measure for distances could improve the accuracy of the IDWM. De Silva et al. (2007) used the aerial precipitation ratio method, the arithmetic mean method, the normal ratio (NR) method, and the inverse distance method to estimate missing rainfall data. The NR method was found to be most accurate. The arithmetic mean method and the aerial precipitation ratio method were most appropriate for the wet zone. You et al. (2008) compared methods for spatial estimation of temperatures. The spatial regression approach was found to be superior over the IDWM, especially in coastal and mountainous regions. Dastorani et al. (2009) predicted the missing data using the NR method, the correlation method, an artificial neural network (ANN), and an adaptive neuro-fuzzy inference system (ANFIS). The ANFIS approach performed best for the missing flow data. ANN was found to be more efficient in predicting missing data than traditional approaches. Teegavarapu (2009) estimated missing precipitation records by combining a surface interpolation technique and spatial and temporal association rules. The results suggested that this integrated approach improved the precipitation estimates. Teegavarapu et al. (2009) applied a genetic algorithm and a distance weighting method for estimating missing precipitation data. The genetic algorithm provided more accurate estimates over the distance weighting method. Kim & Pachepsky (2010) reconstructed missing daily precipitation data with a regression tree and an ANN. Better accuracy was accomplished with the combined regression tree and ANN rather than using them independently. Hosseini Baghanam & Nourani (2011) developed an ANN model to estimate missing rain-gauge data. The resulting feed-forward network was found to be accurate. Nkuna & Odiyo (2011) confirmed accuracy of the ANN in estimating the missing rainfall data. Hasanpur Kashani & Dinpashoh (2012) assessed accuracy of different methods of estimating missing climatological data. They concluded that although the ANN approach is more complex and time consuming, it outperformed the classical methods. Also, the multiple regression analysis method was found to be most suitable among the classical methods. Choge & Regulwar (2013) applied ANN to estimate the missing precipitation data. Che Ghani et al. (2014) estimated the missing rainfall data with the gene expression programming (GEP) method. The GEP approach was used to determine the most suitable replacement station for the principal rainfall station. Teegavarapu (2014) attempted to achieve statistical corrections for spatially interpolated missing precipitation data estimations.

The literature review indicates that there are no significant studies that evaluate various methods for estimating missing precipitation data in arid regions, such as southern parts of Iran and most of them have been performed in countries with almost mild or wet climates such as the studies of Xia et al. (1999), Teegavarapu & Chandramouli (2005), De Silva et al. (2007), You et al. (2008), Teegavarapu (2009), Teegavarapu et al. (2009), Kim & Pachepsky (2010), Che Ghani et al. (2014), and Teegavarapu (2014). Also most of the previous research is about the application of ANN and GEP methods in comparing classic methods, but there is not any remarkable study that evaluates the efficiency of the M5 model tree, which is one of the new and modern data mining methods.

The purpose of this study is to investigate the capability of 10 different traditional and data-driven methods to estimate missing precipitation data in arid areas of southern Iran and to identify the most appropriate method. The 10 examined methods include arithmetic averaging (AA), inverse distance interpolation, linear regression (LR), multiple imputations (MI), multiple linear regression analysis (MLR), non-linear iterative partial least squares (NIPALS) algorithm, NR, single best estimator (SIB), UK traditional (UK) and M5 model tree.

Normally, there are no particular issues regarding recording data at meteorology stations. However, the inconsistency of the data record may happen in certain time sections per se. Hence, in this study we have hypothesized that 10% of data might not be measured. It may need to be estimated.

In this study, the Bandar Lengeh and Bandar Abbas stations were considered the target stations. The Bandar Abbas station is likely to have a precipitation regime different from other stations because it is affected by the elevation of Hormozgan Province. Thus, this station was not taken to be a target one. On the other hand, Bandar Lengeh is located almost in the middle of the zone regarding its latitude and longitude.

After statistical analysis and quality control of the available data, including homogeneity and trend tests, an attempt has been made to evaluate the efficiency of different classic statistical methods and a decision-tree model to estimate missing data.

In the SIB method, the closest neighbor station is used as an estimate for a target station. The target station rainfall is estimated using the same data from the neighbor station that has the highest positive correlation with the target station (Hasanpur Kashani & Dinpashoh 2012).

LR is a method used for estimating climatological data at stations with similar conditions. In statistics, LR is an approach for modeling the relationship between scalar dependent variable y and one independent parameter denoted X. LR was the first type of regression analysis to be studied rigorously and to be used extensively in practical applications (Xin 2009). This is because models that depend linearly on their unknown parameters are easier to fit than models that are non-linearly related to their parameters because the statistical properties of the resulting estimators are easier to determine. In this study, the Kish Island station data was used to calculate the missing data of the target station (Bandar Lengeh) using the LR method.

A single imputation ignores the estimation of variability, which leads to an underestimation of standard errors and confidence intervals. To overcome the underestimation problem, multiple imputation methods are used, where each missing value is estimated with a distribution of imputation reflecting uncertainty about the missing data. MI lead to the best estimation of missing values. Since the rainfall data is skewed to the right, the data needs to be transformed by taking the natural logarithm of the observed data before the method is applied. In some cases, the data may not have a normal distribution with a logarithmic transformation. In these cases, other transformation methods such as the Box-Cox power transformations method (Box & Cox 1964) or the Johnson transformation method (Luh & Guo 2000) could be applied. Then, the average of imputed data is calculated to provide the missing data at the target station (Radi et al. 2015). In many studies, five imputed data sets are considered sufficient. For example, Schafer & Olsen (1998) suggested that in many applications, three to five imputations are sufficient. In this study, the statistical XLSTAT software was used to generate multiple imputations.

The NIPALS algorithm was first presented by Wold (1966) under the name NILES. It iteratively applies the principal component analysis to the data set with missing values. The main idea is to calculate the slope of the least squares line that crosses the origin of the points of the observed data. Here eigenvalues are determined by the variance of the NIPALS components. The same algorithm can estimate the missing data. The rate of convergence of the algorithm depends on the percentage of the missing data (Tenenhaus 1998). In this study, the statistical XLSTAT software is used to generate the NIPALS algorithm.

This method traditionally used by the UK Meteorological Office to estimate missing temperature and sunshine data was based on comparison with a single neighboring station (Hasanpur Kashani & Dinpashoh 2012). In this study, the ratio between the average rainfall at the target station (Bandar Lengeh) and the average rainfall at the station with the highest correlation (Kish Island) was calculated. Then, that ratio was multiplied by the rainfall at the station with the highest correlation to the target station.

Considering the importance of data accuracy in climate studies, the standard normal homogeneity test (SNHT) and the Mann-Kendall (MK) trend test were applied to the data sets using XLSTAT software (Table 2). The SNHT test was developed by Alexanderson (1986) to detect a change in a series of rainfall data. The purpose of the MK test (Mann 1945; Kendall 1975; Gilbert 1987) is to statistically assess if there is a monotonic upward or downward trend of the variable of interest over time.

In the SNHT, the null hypothesis (H₀) was homogeneity of the data and the alternative hypothesis (H₁) was heterogeneity of the data. In the MK trend test, the null hypothesis was randomness and absence of any trends in data, and the alternative hypothesis was non-randomness and presence of trends in the data. If the p-value is more than significance level (α), the null hypothesis is confirmed; otherwise, the alternative hypothesis is acceptable. The results in Table 2 show that the data related to monthly precipitation is homogeneous and random at all stations and can be used with confidence. The correlation of monthly precipitation at different stations is important and applicable in modeling. Hence, the correlation between the monthly precipitation at different stations was investigated (Table 3). The synoptic station of Bandar Lengeh was used as the target station.

As shown in Table 3, the precipitation at the Bandar Lengeh station is most correlated with the Kish Island, Bandar Abbas, and Abomoosa Island stations, respectively. Latitude is a key factor behind varying precipitation levels across different regions. Precipitation correlation values in different stations are, therefore, positively correlated with their respective latitude. As Table 3 indicates, precipitation correlation values are greater between Bandar Lengeh and Jask stations than between the Jask and Bandar Abbas stations. This could be attributed to latitudinal proximity of Bandar Abbas to Jask as well as to the evident comparability of the two cities in terms of condition, which also applies to other stations. Out of the total precipitation data at each station, 10% was randomly assumed to be missing. The missing data was used as a test section and the residual one for training. The number of neighboring stations employed in different methods was dependent on the method. For example, in the methods of LR, UK and SIB, only one station's data highly correlated with target station's data was employed, but in the AA, IDWM, MLR, NR and NIPALS methods, all of neighboring stations were used. In the M5 and MI methods, different combinations of input parameters varying from one station to five stations were used to see which one had better performance.

In the multiple imputation method, the best results were obtained for data at five stations. In estimating the missing values of precipitation at Bandar Lengeh, the M5 decision-tree model was selected. The best results were obtained when the data related to monthly precipitation at the stations of Bandar Abbas, Jask, Abomoosa and Kish Islands was used. The M5 model in the form of three decision rules (involving linear Equations (12)–(14)) estimates the monthly precipitation at the Bandar Lengeh station with relatively acceptable accuracy. These rules are provided in Table 4.

Table 4

The rules produced by the M5 model tree for monthly precipitation estimation

Rule no. .	If .	Then .	Equation no. .
1	and		(12)
2			(13)
3	otherwise		(14)

Rule no. .	If .	Then .	Equation no. .
1	and		(12)
2			(13)
3	otherwise		(14)

Note: *B represents Bandar in all equations.

View Large

Decision rule (1) above states that if the amount of monthly precipitation at Bandar Abbas is equal or less than 3.55 mm and the monthly precipitation at Kish Island is equal or less than 0.15 mm, then monthly precipitation in Bandar Lengeh is calculated from Equation (12). Rule 2 states that if the monthly precipitation at Kish Island is equal to or less than 24.7, then the monthly precipitation at Bandar Lengeh is calculated using Equation (13). According to rule 3, in other situations, the amount of monthly precipitation at Bandar Lengeh is computed using Equation (14). The results obtained from various classic statistical methods and the M5 decision tree model are presented in Table 5.

In the study reported in this paper, the monthly precipitation data at six stations located in arid areas was considered. The data collected was homogeneous, and no trends were found. However, numerous values were missing. Different methods were applied to fill in the missing data. The computational results demonstrated that among classical statistical methods, AA, MLR, and the NIPALS algorithm performed best. The high performance of AA might be related to the location of research stations at a similar elevation (between 5 to 30 meters above sea level). Therefore, using the AA method in arid areas with similar elevation is suggested. The results indicated that the MLR method was found to be suitable for estimating missing precipitation data. This result supports the findings of Eischeid et al. (1995), Xia et al. (1999), and Hasanpur Kashani & Dinpashoh (2012). Furthermore, Shih & Cheng (1989) stated that the regression technique and the regional average can be applied to generate missing monthly solar radiation data. They found the regression technique and AA satisfactory in interpolating missing values. The multiple imputation method performed best when precipitation data from five dependent stations was used. This finding was supported by the results reported in Radi et al. (2015). The research reported in this paper has demonstrated that the results if-then rules produced by the decision-tree algorithm provided high accuracy results with the correlation coefficient of 0.95, Nash-Sutcliffe coefficient of 0.89, root mean square error of 5.07 mm, and the mean absolute error of 2.48 mm. Due to its simplicity and high accuracy, the decision-tree model was suggested for estimating the missing values of precipitation in non-arid climates. Although the results reported in this paper were derived from regions in a single country, the results would be applicable to arid and semi-arid regions in other countries. This is due to the fact that all arid and semi-arid regions share the same or similar climate conditions.