Information Entropy Suggests Stronger Nonlinear Associations between Hydro-Meteorological Variables and ENSO

Abstract

Understanding the teleconnections between hydro-meteorological data and the El Niño–Southern Oscillation cycle (ENSO) is an important step towards developing flood early warning systems. In this study, the concept of mutual information (MI) was applied using marginal and joint information entropy to quantify the linear and non-linear relationship between annual streamflow, extreme precipitation indices over Mekong river basin, and ENSO. We primarily used Pearson correlation as a linear association metric for comparison with mutual information. The analysis was performed at four hydro-meteorological stations located on the mainstream Mekong river basin. It was observed that the nonlinear correlation information is comparatively higher between the large-scale climate index and local hydro-meteorology data in comparison to the traditional linear correlation information. The spatial analysis was carried out using all the grid points in the river basin, which suggests a spatial dependence structure between precipitation extremes and ENSO. Overall, this study suggests that mutual information approach can further detect more meaningful connections between large-scale climate indices and hydro-meteorological variables at different spatio-temporal scales. Application of nonlinear mutual information metric can be an efficient tool to better understand hydro-climatic variables dynamics resulting in improved climate-informed adaptation strategies.

1. Introduction

For many water resources planning and management studies, reliable preliminary estimates of dependence between two hydroclimatic variables are extremely important. For example, knowledge of dependence between large-scale climate patterns such as El Niño–Southern Oscillation (ENSO) [1], the Pacific Decadal Oscillation (PDO) [2], and the Atlantic Multi-decadal Oscillation (AMO) [3] with local precipitation, temperature, or streamflow has resulted in improved longer lead-time forecasting models [4,5,6]. Large-scale climate patterns can also predict ecological processes better than local weather [7]. In addition to that, several studies indicated the presence of a significant relationship between large-scale climate phenomena and hydrologic extremes, such as extreme precipitation events [8,9,10], droughts [11,12,13], and floods [14,15]. Therefore, the presence of any kind of significant dependence forms the preliminary metric to identify appropriate predictors for forecasting streamflow and other hydroclimatic variables in ungauged river basins [16,17,18]. As a result, the predictability of these large-scale climate patterns much in advance is extremely important to improve the design of early warning systems of extreme events [19].

A wide range of methods is available for detecting the presence of an association between bivariate data set. Among them, Pearson’s correlation coefficient is the most widely used metric to quantify the linear dependence between any two variables [20]. Pearson correlation coefficient is based on the assumption that the considered variables follow a Gaussian distribution. Therefore, using Pearson correlation in case of variables that follow non-Gaussian distributions may be suboptimal.

However, several insights were derived from these linear associations. For example, Zhang et al. [21] quantified the Pearson linear correlation between different sea surface temperature (SST) anomalies and seasonal precipitation for Huai River basin in China. The authors identified some positive/negative correlation with the coefficients ranges from absolute 0.2 to 0.3. Whereas, Wrzesinski [22] characterized and confirmed the strong influences of large-scale climate index NAO to seasonal river flow in Poland by comparing the difference in average runoff during the positive and negative NAO phases. On the same lines, Wrzesinski [23] also found the linkage between NAO and European river streamflow at the 140 gauging station by analyzing the high and low water flows according to the positive and negative NAO phases.

In addition, most of the hydro-climatic series do not follow a Gaussian distribution leading to a possible misinterpretation of the dependency between the variables when using linear measures [24,25,26]. As a result, the non-parametric rank-based correlation metrics of Kendall’s tau and Spearman’s rho are applied to quantify the relationship between two given variables [27,28,29]. However, even though the associations evaluated by these two metrics are independent of any probability distribution assumptions, they are more successful in detecting the monotonous relationships between any two variables. In addition, assuming a monotonous relationship among two hydroclimatic variables might be too restrictive to characterize the existing complex dependence structures between the hydroclimatic variables [30,31,32].

To overcome the limitations of linear dependence, several studies in hydro-climatology adopted the concepts based on nonlinear statistics to evaluate the strength of association among distinct hydroclimatic variables. For instance, Fleming et al. [33] observed a nonlinear association between northern hemisphere river basin streamflow and teleconnection patterns. In regard to climate extremes, Cannon [34] utilized generalized extreme value distributions to investigate the relationship between ENSO and winter extreme station precipitation in North America. Whereas Lin-Ye et al. [35] investigated the relationship between extreme events of wave storms and large-scale climate covariates using generalized additive and linear models. While Kusumastuti [36] studied the nonlinearity between threshold and rainfall-runoff transformation. More recently, Konapala et al. [37] have investigated the nonlinear relationship between low flows in Texas river basins with large scale climatic patterns.

Recently entropy-based approaches are gaining popularity for different applications, such as, climate variability [38], uncertainty analysis [39], hydrometric network analysis [40,41,42], selection of predictors [43,44], and drought regionalization [45]. Similarly, the entropy theory based mutual information (MI) is widely used for quantifying the non-linear association between multiple variables [40,41]. The MI concepts have been utilized to quantify nonlinear dependence structure between several climate variables across the globe [46,47]. Also, several MI metrics were developed to assist in predictor selection for hydroclimatic forecasting [48,49]. More importantly, MI was able to detect the presence of strong nonlinear associations of streamflow, rainfall, and global mean temperature with several large-scale climatic variables [31,50,51,52,53,54].

Unlike the signals in other research fields, hydroclimatic data being a subset of geophysical research field are non-repeatable, shorter in length and contaminated with significant noise levels [31]. In addition to that, studies have indicated a presence of chaotic component [55]. Therefore, it becomes evident that the selected measure should be relatively robust to noise and chaos and, more importantly, detects signals possibly shorter in length. Thus, the goal of this study is to investigate and compare the performance of nonlinear estimation method based on mutual information (MI) with the traditional linear association metric of Pearson correlation. For this purpose, the hydro-meteorological data are considered such as total annual precipitation, annual average streamflow, and extreme precipitation indices specific to Mekong River Basin in Southeast Asia and estimated its association with the large-scale climate index of ENSO.

2. Materials and Methods

2.1. Hydro-Meteorological Data and ENSO

The Mekong River Basin (MRB), originates from Tibetan Plateau in China and flows through the territory of five other countries in the Southeast Asia (Myanmar, Laos PDR, Thailand, Cambodia, and drains to the sea in Vietnam) (Figure 1a). This is currently home to more than 70 million people and it is expected to increase by 100 million in 2050 [56]. The detailed topography map of study region is displayed in Figure 1a with relative high terrain in the upstream of the basin (upper Mekong). In this study, the gridded precipitation data is provided from Asian Precipitation Highly Resolved Observational Data Integration Towards the Evaluation of Water Resources (APHRODITE) with a spatial resolution of 0.25° and daily temporal resolution with the sufficiently long record period from 1951–2007 covering the Asian monsoon domain. The dataset was created primarily with data obtained independently from rain gauge observation network across all the regions of Asia. The daily precipitation values from rain gauges were interpolated using the sphere map technique [57] and the first six components of the fast Fourier transforms were taken to obtain daily data for all land areas in the Asian monsoon domain [58]. This dataset is referred to as “APH” in this paper, for the purpose of brevity. APH has been previously utilized as a reliable gridded rain data set over this area for various hydrological applications [59,60,61,62,63]. The spatial distribution of total precipitation from APH data over MRB is displayed in Figure 1b with high rainfall volumes (around 2000 to 2500 mm) gathering around the eastern part of the river basin near to the Annamite mountain range.

The southwest monsoon wind originated from the Gulf of Thailand blocked by the Annamite mountain range is the main factor that causes the high seasonal rainfall in summer months (June, July August, September) [61] (Figure 2a). The APH daily gridded rainfall values were extracted to station locations using bilinear interpolation [59,63,64] to four meteorological stations located in MRB—Chiang Saen, Vientiane, Nakhon Phanom, and Pakse (Figure 1a)—for the time period of 1951 to 2007 over the Mekong River Basin. In addition to rainfall data, the corresponding annual streamflow data were also collected at these four hydrological stations from the MRC website for the same period. The monthly mean climatology over 57 years of hydro-meteorological stations is displayed in Figure 2a. Even though different patterns are found over four stations, the hydrographs are quite similar from the upstream station at Chiang Saen to downstream station Pakse, with flooding season from May to November, while the peak discharge months are around August and September. This study considers the seasonal cycle which includes all 12 months of the hydrologic year (starting from the month with the lowest rainfall—i.e., May—and ending in April of the next year; the streamflow month is considered to lag one month which starts from June and ends in May next year).

The study focuses on the total annual rainfall, annual streamflow, as well as extreme precipitation indices [65] for hydrologic years derived from daily precipitation datasets for both wet and dry spell statistics, which are (1) R5d: Max consecutive five-day rainfall, (2) P90p: 90th percentile of the daily precipitation time series for the year, (3) SDII: average daily precipitation on a wet day, and (4) dry spell length computed from daily precipitation for a year. Among four stations, Pakse has the highest streamflow data due to its furthest downstream location. Nakhon Phanom and Pakse have the highest rainfall as seen in Figure 1 and Figure 2.

The assumption has been that the long-term hydro-meteorological data variability can be captured by the fluctuation of the SST over seasonal scales. In this study, the ENSO index has been considered by using the SST over NINO 3.4 region [60,66], which is the area averaged monthly SST over the region bounded by the coordinates 5° N–5° S, 170° W–120° W. This time series could effectively indicate the occurrence of ENSO events [66]. The dataset can be downloaded from the Climate Data Guide website by NCAR [67]. Therefore in this study, the ENSO indices are computed at multiple three-month time windows in terms of quarters [31] consisting of the average SST anomalies over months of JFM, FMA, MAM, etc., for all the years (J: Jan, F: Feb, M: Mar). The SST over Nino 3.4 regions boxplots bound for 57 years for 10 different quarters are displayed in Figure 2b. The red crosses in Figure 2b denote outliers. There are four quarters just before the seasonal cycle, four quarters corresponding to the seasonal cycle, and two quarters after seasonal cycle, computed with the mean monthly SST over Nino 3.4 region. Among all 10 quarters, the fourth quarter (AMJ) has the highest median and whisker values compared to other quarters whilst the last two quarters (9 and 10) have the lowest median values.

2.2. Mutual Information Estimation

The mutual information (MI) has been utilized to capture the nonlinear dependence structure between two random variables. When analyzing experimental times series from the non-linear system, the MI is especially an important statistics [68]. According to [69], there are three theorems for MI between two random variables X, Y: (1) MI is non-negative and is zero if X and Y are strictly independent; (2) MI is infinity if there exists a function “g” such that X = g(Y); (3) MI is invariant to separate one to one transformations.

The MI can be computed using the relative entropy suggested by Joe [70]. Assuming that a pair of continuous random variables (X, Y) exist which have a joint probability density function (pdf) $p_{XY}$ and with its marginal pdf accordingly $p_X$ and $p_Y$. The mutual information or relative entropy can be defined as

$$MI(X,Y)={\displaystyle \int {\displaystyle \int p_{XY}\log \frac{p_{XY}}{p_X{p}_Y}dxdy}}$$

(1)

It is noted that the Equation (1) measures the distance between a joint distribution and the distribution when there is independence [69]. In case of continuous variables, there is no direct way to accurately determine continuous probability distributions. Therefore, several methods have been introduced to approximate the continuous probability distribution functions as discrete distributions. Among them, Khan et al. [71] compared four different methods to estimate the probability distribution function to calculate MI using: kernel density estimator (KDE) [68], K-nearest neighbors (KNN) [72], Edgeworth approximation of multivariate differential entropy [73], and adaptive partitioning of the XY plane [74]. The authors found that KDE and KNN outperform the other two methods in term of their ability to capture the dependence structure. Khan et al. [31] indicated that KDE is able to capture the underlying nonlinear dependence more consistently compared to KNN and Edgeworth when they are short and noisy assuming such dependence exists. Therefore, this article utilizes the KDE to estimate probability density function using the equation

$$p\left(x\right)=\frac{1}{n}{\displaystyle \sum _i^{n}K\left(u\right)}$$

(2)

in which, K(u) is a multivariate kernel function

$$K\left(u\right)=\frac{1}{h^d\sqrt{{\left(2\pi \right)}^d\det (S)}}\exp \left(-\frac{u}{2}\right)$$

(3)

$$u=\frac{{\left(x-x_i\right)}^T{S}^{-1}\left(x-x_i\right)}{h^2}$$

(4)

From (4), x_i a d-dimensional random vector for the multivariate data set (x₁, …, x_n). S is the covariance matrix on the x_i, det(S) is a determinant of S and h is the kernel bandwidth or smoothing variable. The optimal Gaussian bandwidth “h” can be computed using Equation (5) [75]

$$h={\left(\frac{4}{d+2}\right)}^{\frac{1}{d+4}}{n}^{-\frac{1}{d+4}}$$

(5)

The “d” is taken from [76] for the value of d = 2, and is similar to [68].

Substituting Equations (3) and (4) to Equation (2) to obtain the approximate probability density function as

$$p\left(x\right)=\frac{1}{n}{\displaystyle \sum _i^n\frac{1}{h^d\sqrt{{\left(2\pi \right)}^d\det (S)}}\exp \left[-\frac{{\left(x-x_i\right)}^T{S}^{-1}\left(x-x_i\right)}{2h^2}\right]}$$

(6)

The detailed procedure using KDE to estimate pdf can be found in [31,68,71]. Here, the discrete formulation of MI is shown in Equation (7)

$$MI(X,Y)=\frac{1}{n}{\displaystyle \sum _{i=1}^n\ln \frac{p_{XY}\left(x_i,y_i\right)}{p_X(x_i)p_Y(y_i)}}$$

(7)

In which $p_{XY}\left(x_i,y_i\right)$ is the joint pdf and $p_X\left(x_i\right),p_Y\left(y_i\right)$ are marginal pdfs at $\left(x_i,y_i\right)$. The MI values range between independent (MI = 0) to completely dependent (MI = $\infty$). In order to make the generalization of the correlation with a range from 0 (independent) to 1 (dependent), Joe [70] proposed a formula to transform the MI to nonlinear correlation coefficient (NLCC)

$$NLCC=\sqrt{\left[1-\exp \left(-2MI\right)\right]}$$

(8)

The NLCC range is similar to linear correlation coefficients (LCC) and has been used in most of the studies by [31,69,70].

Figure 3 modified [77] using their python code to demonstrate the LCC and NLCC using the scatter plots of two random variables obtained from a different sample of data. The advantage of NLCC is based on the fact that MI makes no assumption on the distribution of the variables or the nature of the relationship between them and is sensitive to nonlinear and non-monotonic effects [77]. It can be observed that (top row of Figure 3) both the LCC and NLCC can able to capture the linear relationship between the two random variables with a very close range of values. However, the advantage of NLCC compared to LCC is due to its ability to recognize the different distributions of the two random variables (bottom row of Figure 3).

Subsequently, the LCCs are computed between each of the precipitation indices and compared with NLCCs at four hydro-meteorology stations located in MRB (Figure 1a). The four stations are chosen based on their differences in terms of seasonal cycles, total annual rainfall amount and geographical height. Finally, the gridded linear/non-linear CC is constructed for the MRB to showcase the spatial variability of the correlation coefficient.

3. Results

3.1. Linear and Nonlinear Correlation between Annual Precipitation/Streamflow and ENSO Index

We first illustrate the linear dependence as a bivariate normal distribution and nonlinear dependence as a kernel bivariate distribution following the work of Khan et al. [71]. The bivariate normal and kernel density between the annual average streamflow at different hydrologic gauging stations and different quarters of ENSO indices are computed and plotted in Figure 4 for the highest and lowest linear and nonlinear CCs. For kernel density, a Gaussian kernel with an optimal Gaussian bandwidth computed by $h=N^{-1/6}$ with N is the total number of observed points (57 in this case).

Subsequently, the linear and nonlinear CCs between total annual precipitation and average streamflow were computed at different window lengths of SST Nino 3.4 (quarters) for four different hydro-meteorology stations and display in Figure 5. Based on the formulation Equation (8) NLCCs are positive, therefore the absolute value of LCCs are computed for comparison plotting in Figure 5 and all other figures henceforth. In order to compute the 90% confidence intervals for the absolute correlation coefficients, the bootstrapping approach [71] was applied using 100 simulations and plotted in Figure 5. It can be seen in Figure 5 that the NLCCs have higher values than the LCCs for all quarters. It indicates that the KDE captures the more extrabasinal connection between ENSO and precipitation as well as river flows compared to linear modeling [31].

In particular at station scale, Nakhon Phanom has the highest precipitation amount, its LCCs and NLCCs are also among the highest for the first four quarters Figure 5(c1,c2) which subsequently decrease. The NLCC values also have the same patterns as LCCs for this station as well as Pakse in Figure 5(d1). Although both stations have highest CCs for the first quarter, however, the lowest CCs values are not found at the same: Quarter 7 (Nakhon Phanom) and Quarter 4 (Pakse). Chiang Saen and Vientiane have variation trends for LCCs Figure 5(a1,b1) among all quarters, this is perhaps due to the less rainfall amount observed at these two rain gauges.

Among four hydrology gauging stations, Pakse also has highest LCCs/NLCCs with ENSO at the first four quarters Figure 5(d2). This is because Pakse is the furthest station downstream and it has the highest streamflow measured. Therefore, it shows more dependences by the fluctuations of ENSO indices. The second highest streamflow station is Nakhon Phanom (see Figure 5(c2)) that also indicates similar patterns of LCCs and NLCCs with higher dependences with the first four quarters of ENSO indices compared with the rest. When the LCC values are close to zero (as in the last four quarters of Figure 5(c2)), the differences between LCCs and NLCCs should be interpreted with caution because of the exponential scaling of MI in NLCC as shown in Equation (8). The two upstream gauging stations Chiang Saen and Vientiane (Figure 5(a2,b2)) also have similar patterns for LCCs and NLCCs at different quarters. Compared to total annual precipitation, the annual average streamflow has slightly higher LCCs and NLCCs, except for Nakhon Phanom station. Overlapping the 90% confidence intervals of LCCs and NLCCs for streamflow at Chiang Saen, Vientiane, and Pakse (Figure 5(a2,b2,d2)) for the first three quarters indicates that both KDE and linear regressions effectively capture the strong dependence structure.

3.2. Linear and Nonlinear Correlation between Precipitation Extreme Indices and ENSO

In addition to total annual precipitation and streamflow dependences, further investigations on LCCs and NLCCs are carried out between precipitation indices for extreme values indices (R5d, SDII, P90p, dry spell) with different quarters of ENSO events. This analysis aims to quantify the dependence structure between annual extreme precipitation events and ENSO indices. The wet indices (R5d, P90p, SDII), as well as dry indices (dry spell), are also taken into account for the analyses. The detail dependences based on LCCs and NLCCs can be found in Figure 6 for four indices and four stations over 10 quarters. Similar to Figure 5, the LCCs are displayed as absolute values comparable to NLCCs. The 90% confidence intervals for the absolute correlation are computed based on 100 simulations using bootstrapping approach. The detail analyses on Figure 6 reveals several significant dependencies between annual extreme precipitation events and ENSO indices. Chiang Saen exhibits high LCCs/NLCCs for a maximum five consecutive days of rainfall with the last three quarters of ENSO indices and low CCs values for the first three quarters Figure 6(a1). Similar patterns are found for Pakse station in Figure 6(d1) but with the highest CCs among Quarters 6, 7, 8 and lowest at the first three quarters. This variation is different from the total annual precipitation in Figure 5(a1,d1). In contrast to R5d, the other two indices: P90p and SDII have similar patterns of LCCs/NLCCs compared to the total annual precipitation in Figure 5. The last three-quarters of ENSO indices show that it has higher LCCs/NLCCs for Vientiane and Pakse Figure 6(b2,b3,d2,d3) compared to other quarters whilst the first three quarters of Nakhon Phanom has the highest values than the last three quarters. The dry spell indices of all four stations show nearly opposite patterns compared to R5d. The overall pattern illustrates that nonlinear CCs have higher values (more dependencies) than linear CCs for all stations/indices, similar to Figure 5 and [31].

The analyses with total annual precipitation, annual average streamflow, and extreme precipitation indices on wet and dry conditions reveal that there exists a nonlinear extrabasinal connection between ENSO and the hydro-meteorological data over Mekong river basin. The analyses based on LCCs/NLCCs exhibit the increasing trend in the variation of annual statistics on hydro-meteorology in connection with ENSO by computing nonlinear relationship as compared to linear measures. Therefore, it is expected to give additional support for early prediction (compared to traditional linear measurement) based on the ENSO forecast with the hydro-meteorology connection when MI-based approaches are utilized. This approach, somehow, would be helpful in water resources management for drought mitigation, flood control, as well as an irrigation system for agricultural areas.

3.3. Spatial Linear and Nonlinear Correlation Maps

Based on the correlation analysis at four rainfall stations, it was observed that the NLCCs have higher values in comparison to LCCs at all the selected stations located in MRB. Further analyses of the spatial pattern using the gridded data from APH based on the selected statistics of annual precipitation and extreme indices have been carried out. The linear and nonlinear CCs between ENSO Quarter 1 and annual rainfall are displayed in Figure 7 (using the same color bar), and the correlation values range from 0 to 0.6. Quarter 1 was selected arbitrarily as the first quarter, even though, it shows the highest CCs over Nakhon Phanom and Pakse for total annual precipitation but it also displays the lowest CCs for other analyses such as for Vientiane extreme indices (Figure 6b) or Pakse in Figure 6(d1,d3). Figure 7a exhibits the higher CC values between total annual precipitation and ENSO Quarter 1 (about 0.4) from Nakhon Phanom to Pakse station, along with Laos, eastern Thailand, and northern Cambodia. This can be clearly observed during the first quarter time frame Figure 5(c1,d1). On the other hand, the NLCCs have higher CC values (about 0.5 to 0.6) for the same locations compared to other grids. The use of spatial distribution map can be extended to generate the teleconnection patterns in ungauged regions. That, in turn, helps to better inform local stakeholders in building better tools for water resource management.

Similarly, the dependency between selected precipitation indices (R5d, P90p, SDII, and dry spell) and Quarter 1 of ENSO are computed and displayed in Figure 8. The linear CCs obtained from precipitation indices (Figure 8) are slightly lower than that of annual precipitation in the previous analysis (Figure 7) for Laos, eastern Thailand, and northern Cambodia. There is slightly higher LCCs/NLCCs magnitude observed at southern Cambodia during a dry spell Figure 8(a4,b4). The magnitudes of NLCCs (Figure 8) are also slightly lower than NLCCs based on the annual precipitation (Figure 7). However, the values obtained from NLCCs are still significantly higher than LCCs for all extreme indices. The lower Mekong basin (the southern part of river basin spread over Laos, Thailand, and Cambodia to Vietnam) has higher correlations to ENSO indices compared to the northern/upper basin, similar to [78].

4. Discussion

Although ENSO has a direct influence on rainfall anomalies over the tropical and subtropical regions, only a portion of the variation in the annual flow of rivers located in these regions is associated with ENSO events [31]. This study discusses in detail the possible dependences between different quarters of ENSO indices and hydro-meteorological dataset over Mekong river basin. Other existing studies on MRB highlighted that the linear correlations between ENSO and streamflow over the selected locations in Chiang Saen, Vientiane, and Pakse are around −0.4 to 0.2 [78] and maximum linear correlations between ENSO and precipitation over several stations in Thailand are around −0.18 to 0.22 [79]. The above correlation figures are in line with the study that the maximum LCCs obtained are around 0.35 for both annual precipitation and streamflow as displayed in Figure 5. The results from mutual information derived NLCCs therefore suggests an additional approach to look into the dependence structure for multivariate hydro-meteorological data with the large-scale climate patterns. ENSO is a periodic climatic phenomenon with 3–7 years of the cycle and can be predicted several seasons in advances [80]. For instance, Ludescher et al. [81] were able to predict the likelihood of El Niño conditions in 2014 almost a year in advance indicating an improvement in our prediction capacity. In addition, other researchers [82,83] also revealed the intensification of ENSO related precipitation and El Niño frequency in the future due to the warming associated with an increase in greenhouse gas emission. Moreover, Ward et al. [15] indicated that the global flood risk exists during El Niño or La Niña years, or both, in basins spanning 44% of the land surface of the Earth. Therefore, the predictability of ENSO has significant implications and the ability to predict in advance might lead to better local water resources management towards developing more efficient flood and drought early warning systems.

This study limits to using only the kernel density estimation approach to estimate the mutual information values. However, there are several methods that have been used to compute the mutual information, such as K-nearest neighbors (KNN) [72], Edgeworth approximation of multivariate differential entropy [73] and adaptive partitioning of the XY plane [74]. The detailed comparisons among these estimator approaches were carried out in [71] for a different type of data such as linear, quadratic, periodic function, chaotic system. Khan et al. [71] concluded that KDE is the best choice for very short data (50–100 data points). Therefore, KDE approach was utilized to estimate the bivariate mutual information in this study. In addition to mutual information method, there are several approaches used to build the nonlinear relationship between multivariate. Lin-Ye et al. [35] applied the VGAM/VGLM to quantify the nonlinear relationship between storm components and large-scale climate indices (NAO and others) using Global Climate Model data via the regression coefficients that were used to build the location and scale parameters of their statistical model. Zhang et al. [84] investigated the nonlinear relationship by employing the concept of mutual information to evaluate the dependency between the normalized difference vegetation index (NDVI) and meteorological variables for the middle reach of the Hei river basin. Higher dependency between NDVI and coupled precipitation/temperature was observed for the desert area whilst for oasis region, groundwater is an important factor for driving vegetation growth. The authors utilized the mutual information as a method to classify the study region into a smaller area (desert, oasis, artificial oasis).

The NLCCs and LCCs analysis between ENSO and hydro-meteorological data suggest that there exists a nonlinear correlation for the Mekong River Basin. This finding agrees well with a previous study [28] with a focus on river flow analysis over tropical and sub-tropical river basins. It can be observed that although the linear CCs can be 0, the smallest NLCCs can be around 0.3 to 0.4. This needs caution from an artifact of Equation (8) which scales nonlinear CCs exponentially with MI [31] when linear CCs are close to 0. It, however, does not affect the other values of CCs higher than 0.

5. Conclusions

This study analyses the possible influence of the large-scale climate index ENSO and hydro-meteorological data in the form of total annual precipitation, annual average streamflow, and extreme precipitation indices using linear and nonlinear correlation over the Mekong River Basin. The nonlinear correlation structure was computed based on mutual information approach using marginal and joint entropy. The kernel density estimation approach was selected among several other techniques to estimate the mutual information values as this approach works best for very short data length of 50 to 100 points. Nonlinear correlations were obtained by transforming the mutual information using Joe formula [70] and thus are comparable with absolute linear correlation coefficients. Bootstrapping approach based on 100 simulations was applied to find the confidence intervals for these absolute correlation coefficients. In-depth analysis was carried out at four hydro-meteorological stations located in the Mekong River Basin. The major conclusions that can be drawn from this study are listed below:

• Nonlinear correlation is able to reveal the additional dependence structures between hydro-meteorological data over Mekong River Basin and ENSO indices.
• Both linear and nonlinear correlation exhibits similar varying patterns among different ENSO quarters for most of the stations/indices.
• The results reveal that higher correlation coefficients can be found using the nonlinear correlation coefficients in comparison to the traditional linear correlation analysis.
• Spatial correlation structures for LCCs and NLCCs are also constructed based on extreme precipitation indices and ENSO. The use of spatial maps further complements our analyses based on a single station to other ungagged regions to better inform local stakeholders in building better tools for water resource management.
• Further analyses are required to reveal the non-linear association between other large-scale climate phenomena (SOI, PDO, NAO, etc.) with local meteorological variables. The mutual information between these indices and local meteorological variables can further help policymakers to improve climate-informed adaptation studies.