Climate change or regional human impacts? Remote sensing tools, artificial neural networks, and wavelet approaches aim to solve the problem

Lake Urmia, as the largest lake in Iran, has suffered from water-level decline and this problem needs to be investigated accurately. The major reason for the decline is controversial. The current paper aimed to study the hydro-environmental variables over the Lake Urmia basin using remote sensing tools, artificial neural networks, wavelet transforms, and Mann–Kendall trend tests from 1995 to 2019 in order to determine the primary reason of the decline and to find the most important hydrologic periodicities over the basin. The results indicated that for the monthly-, seasonally-, and annually-based time series, the components with 4-month and 16-month, 24and 48-month, and 2and 4-year, respectively, are the most dominant periodicities over the basin. The agricultural increase according to the vegetation index and evapotranspiration and their close relationship with the water-level change indicated that human land-use is the main reason for the decline. The increasing agriculture, in the situations that the precipitation has not increased, caused the inflow runoff to the lake to decline and the remaining smaller discharge is not sufficient to stabilize the water level. Temperature time series, also, has experienced a significant positive trend which intensified the water-level change.


GRAPHICAL ABSTRACT INTRODUCTION
The impacts of climate change and its effects have put stress on natural processes such as hydrological cycles, environmental variables, and ecological balance of the eco-systems (Erler et al. ), and global climate change is resulting in a thermodynamic intensification of the hydrologic cycles. While climate change is one of the reasons for hydro-environmental changes, human land-and water-use activities have affected water and land resources and have made a great many changes in the natural processes (Pokhrel et al. ). The over-use of water resources by humans causes socio-environmental problems, especially in semiarid and arid regions such as the Middle East (Madani ). The main reasons for water resources problems are controversial among researchers due to the different approaches that they use to answer the question.
Iran, as a country in the Middle East, and due to the low annual precipitation, is situated among the semi-arid and arid regions. Located in the northwest part of Iran, Lake Urmia is the largest inland lake and has experienced rapid water-level decrement in the past years (Delju et al. ).
The decline of the water-level equaled 8 m from 1995 to 2010, and up to 2012, the lake lost about 60% of its area and more than 90% of its volume, which makes the shrinking problem of Lake Urmia an outstanding tragedy in the Middle East (Hassanzadeh et al. ). The reason for the water-level decline is controversial and a great number of researchers have tried to answer the question. Several studies argue that the reduction of the surface water inflow due to excessive agricultural extraction caused the problem (e.g., see Alborzi et al. ; Nourani et al. ). Some other researchers proposed that building a large number of dams over the rivers which inflow to the lake has created the problem. This idea was triggered due to the extensive construction of reservoirs, although it was rejected in a paper written by Fathian et al. (), showing that decreasing trends were observed in the headwater catchment areas.
Several other studies found that the climate change impacts over the region are the main cause of the decline in the water volume of Lake Urmia (e.g., see Delju et al. ; Tahroudi et al. ; Schulz et al. ).
Data gathering is one of the main processes in performing an accurate study. Various methods and indices are utilized to monitor and study hydrologic, environmental, and climate factors. These methods are divided into two main categories, which are remote sensing-based (RS-based) and site-based indices (Liu et al. ). Although some researchers use site-based tools to study trends in hydroclimatologic variables and climate change over regions, they are not available for every time and region due to the scarcity of ground-based data and they have low accuracy because of some disadvantages such as human mistakes and ungauged stations. One of the most important reasons for the disagreement among researchers about the reason for water resources problems is data. Using ground-based data has several drawbacks. Moreover, using different ground-based stations, some of which are far away from each other, is confusing to study the same basin. Also, even in more recent studies, researchers used old datasets and they could not use near to real-time data to study trends in hydro-environmental variables and the water level, which means what is happening with the lake in recent years is unknown and there is no certainty regarding their conclusions about the reasons for the lake problem.
Therefore, RS-based approaches are more efficient, reliable, and accurate to study climate change and factors affiliated to nature at a large scale with lower cost and less time consumption ( Ji & Peters ). In this regard, various RS sensors are employed to collect data, among which the Mod- Land Surface Temperature (LST) is a remotely sensed temperature index and a climate factor that is widely utilized to study hydro-environmental variables (e.g., see Varouchakis Wavelet transform (WT) is widely used to decompose the signals with either high or low frequency (e.g., see Roushangar et al. ; Nalley et al. ; Nourani et al. ). The classical signal analysis methods, including Fourier transform (FT), use a single-window analysis which leads to losing the frequency localizations at low frequency and the time localization at high frequencies. The WT decomposes a one-dimensional signal into two-dimensional time-frequency domains. Also, the irregular shapes of the mother wavelets make the method useful to study the time series with noises and discontinuities (Drago & Boxall ). The signals produced by nature, due to its complicated behaviors, are always difficult to study without using machine learning tools. WT can reveal various aspects of datasets, including discontinuities, trends, short-and longterm periodic intervals, and breakdown points, especially in long-term, complex signals.
The current study aimed to determine the main reason for the water-level decline problem of Lake Urmia using RS tools, ANNs, wavelet approaches, and trend test techniques as a high research priority for water resources engineering and management. To this end, RS tools were utilized to collect the hydro-environmental datasets from 1995 to 2019, and ANNs were employed to estimate the missing data. Thereafter, wavelet approaches were used to determine the most dominant periodicity over the basin for the water management plans. The Mann-Kendall trend tests were, also, utilized to study potential trends in the time series. Conclusions based on the fundamental hydrology theories are important in solving the controversial problems of water resources.
Water-level fluctuation is among the common hydrologic problems for big lakes across the world. These changes are mostly associated with climate change and anthropogenic activities. It is very important to determine the main reason for the fluctuations in order to make effective decisions in water resources management. The proposed methodology can be applied to other big lakes across the world in order to investigate potential impacts of climate change and regional human activities on them and, also, to determine the most dominant periodicities of the hydroenvironmental variables over the basins.

Case study
Located in the northwestern part of Iran (Figure 1), between 37 4 0 to 38 17 0 latitude and 45 13 0 to 46 longitude, Lake Urmia is a saline and shallow lake. The lake was among the largest lakes and the second hypersaline lake in the

Data collection
The present paper employed RS tools, RS-based datasets, and RS-based models to collect data for hydro-environmental indices and factors, and the water-level data to study precipitation, temperature, vegetation cover, and evapotranspiration of Lake Urmia from 1995 to 2019. All of the RS-based datasets were validated using ground-based measurements.
NDVI is one of the important indexes that researchers use to evaluate vegetation. The structural basis of this index is the presence of chlorophyll in different plants, which absorbs red light, also, the mesospheric layer of the leaf reflects near-infrared (NIR) light. This index well exhibits the response to photosynthetic effects, with higher values indicating denser and fresher vegetation, that greatly influence environmental parameters. Plants and their roots affect the physical properties of soil, such as moisture content, infiltration rate, and shear strength, which play a significant role in environmental conditions. NDVI is a normalized parameter and its value ranges between À1 and 1.
The general equation of the NDVI is (Pettorelli et al. ): where N and R denote the near-infrared band and the red band, respectively.
The NDVI dataset from satellite imagery was used and To collect the precipitation dataset, TMPA3B43 product was utilized, which creates a monthly precipitation average,

Proposed methodology
The current paper aimed to investigate potential trends in the long-term hydro-environmental time series and to analyze the interactions between them and the water-level fluctuations of Lake Urmia. In this way, RS instruments were used to collect data for NDVI, LST, precipitation, and ET over the Lake Urmia basin, and the water-level fluctuations of the lake during 1995-2019. WT, then, was used to decompose the hydro-environmental signals into sub-signals with different frequencies in order to find the most dominant periodicity of the variables over the basin.
To assess the performance of the model, determination coefficient (DC) and root mean square error (RMSE) efficiency criteria were used as Equations (3) and (4) (Nourani ).
where n, NDVI eval i , NDVI eval , and NDVI est i are, respectively, the number of years, averaged value of the evaluated NDVI (via the instrument), and estimated NDVI.

Autocorrelation analysis
Autocorrelation analysis should be applied to time series in order to study the seasonality patterns and correlations. The where R is the autocorrelation coefficient of the sample, x t , x t is the mean value of the sample, and n is the number of samples. If R satisfies Equation (6), there is not a significant autocorrelation in the dataset. In the case of the appearance of a significant autocorrelation, Nourani et al. () suggested using the pre-whitening MK test instead of the original MK.

Wavelet transform
The WT decomposes a non-stationary signal into multiple levels by shifting and scaling the mother wavelet and using high-and low-pass filters. WTs can analyze data on a local scale and reveal various aspects of the dataset at different frequencies.
WTs are widely used in hydro-environmental studies due to their robust properties (Nalley et al. ).
Generally, the WT is sorted into two main classes: discrete wavelet transform (DWT) and continuous wavelet transform (CWT). DWT is often used to decompose hydrologic time series because of the discrete instincts of hydrologic variables. DWT decomposes the signals into the approximation and detail components using high-pass and low-pass filters.
The high-pass filter is the wavelet function that produces the detailed sub-signals which are low-scale or high-frequency components of the original signal. The low-pass filter is the scaling function, which produces the approximate coefficient which is the low-frequency or the large-scale of the original signal (Nalley et al. ). DWT is defined as: where Ψ(t) is called the mother wavelet, and a and b are the numbers that control the wavelet dilation and translation, respectively. a 0 and b 0 are a dilation step greater than 1, and the location parameter greater than 0.
The mother wavelet function for DWT at timescale is generally defined as (Nourani et al. ): where t represents time, and the parameters a ¼ 2 m and b ¼ a × n are the scaling parameter and the location parameter, respectively.
DWT is performed at dyadic scales in hydro-environmental studies and the wavelet function for this method is (Nourani et al. ): Therefore, the DWT for discrete signals (f i (t)) can be defined as (Nourani et al. ): where T i (m,n) is the WT coefficient at level m for the sample n, and N is an integer in the power of two.
sgn(x j À where n is the number of samples and x j stands for the data point at the time j.
The null-hypothesis of the MK test is no-trend, therefore, S is normally distributed with mean ¼ 0 and the variance (σ) is calculated as (Rashid et al. ): where summation is over the ties and d is the extent of any tie. When the observations are not repeated, d equals to 0.
Thereafter, when the continuity correction is applied, the S-statistic becomes S 0 ¼ S À sgn(S), with a normal distribution. For testing the no-trend hypothesis, the Z-value associated with S-statistics of the test is defined as: High positive values of Z denote a positive trend and low negative values indicate a negative trend. The magnitude of the Z-value, also, represents the strength of the trend in the dataset. The probability value (p-value) obtained from the Z-value is used to determine the significance of a trend. In the conditions that p-value is less than pre-determined significant level (here, α ¼ 5%) or greater than the confidence level (here, ¼ 95%), the null hypothesis of no trend is not acceptable (Rashid et al. ).
The pre-whitening Mann-Kendall test (MK2) The pre-whitening Mann-Kendall test (MK2) proposes to first remove the autocorrelation such as lag-one or higher processes from the dataset and, then, apply the test. This method is called pre-whitening and is beneficial in terms of high autocorrelation (Burn & Hag Elnur ). According to Yue et al. (), the method contains four major steps: 1. Calculating the slope of the sample data (β) as: where x i and x j are the ith and jth observation of the dataset.
Then, remove the trend from data as: 2. Compute the autocorrelation of the de-trended data as in section 'Autocorrelation analysis'.
3. Remove the autoregressive component from the new dataset to get a residual time series as: 4. Add the trend back to the residual series as:

Wavelet transform-Mann-Kendall method
The WTMK technique contains two main steps. As the first step, DWT is applied to each signal in order to decompose it into its components. Choosing the best mother wavelet and the level of decomposition are important parts of this step.
The appropriate mother wavelet can be selected according to the similarity between the shape of the time series and that of the mother wavelet. Recent literature recommended Daubechies function as an appropriate mother wavelet to decompose hydrological signals due to the shape of the Dau- where L stands for the maximum number of decomposition levels, v is the number of vanishing moments of a Daubechies wavelet function which is half of its starting length, and n stands for the number of data points in a time series. (for annual scale). Therefore, three, four, and five levels, and two, three, and four levels, and one and two levels of decomposition were tried for monthly, seasonal, and annual time series, respectively.
According to the above-mentioned theories, four and five levels, and two and three levels of decomposition were used for monthly and seasonal time series, respectively. All smooth mother wavelets (db5-db10) were applied to the signals to determine the more appropriate one in terms of the lowest mean relative error (MRE). The MRE can be calculated as (de Artigas et al. ): where n is the number of records of a signal with x j original data value, and a j is the approximation component of x j .
For the monthly timescales of the hydro-environmental datasets, the lowest MRE was generally obtained for five levels of decomposition when the different db functions for three, four, and five levels of decomposition for each signal were applied. Therefore, five levels of decomposition were chosen to decompose the signals with a monthly horizon using DWT (db types vary from one variable to another).
Thereafter, MRE was used to determine the best db to analyze the datasets of each variable. For the monthly dataset of LST, for example, applying different db mother wavelets, produced the lowest MRE using db6 (MRE ¼ 0.46).
Similar procedures were utilized for the seasonal datasets in order to find the best methods to analyze the signals in terms of the lowest MRE. Two, three, and four levels of decomposition were applied to the datasets with the seasonal horizon and four levels of decomposition admitted the lowest MRE. Therefore, four levels of decomposition were applied to the datasets to analyze them using DWT (db types vary from one variable to another). For annual datasets, also, one and two levels of decomposition were applied and the lowest MRE was obtained for two levels of decomposition. Thereafter, the same steps as for monthly and seasonal horizons were applied to find the best smooth db which varied from one dataset to another.
As the second step of WTMK, the MK trend tests were employed to determine potential trends in approximation and detail components' subseries, as well as the detail components' combination with relevant approximations.

Correlation coefficient
The correlation coefficient (CO) is widely used in data analysis to show the relationship between two datasets.
The current study used CO as well as Z-value to determine where n and m are the average of the variables n and m, respectively.

RESULTS AND DISCUSSION
Reviewing the datasets The current paper used RS tools' instruments to acquire hydro-environmental data over the Lake Urmia basin and Two MK tests were utilized to detect potential trends in the original time series and the components. The MK1 was used to study the datasets without a significant autocorrelation and MK2 was employed to study the datasets with a statistically significant lag-one autocorrelation.

Results of WTMK technique
The approximation (A), and detail components (D) of the hydro-environmental variables with five, four, and two levels of decomposition for monthly, seasonal, and annual datasets, respectively (the mother wavelets were different from each dataset to another), were collected (an example is given in Figure 5). It is noteworthy that due to the space According to the results (Table 2)

Seasonal data analysis
The presence of annual and seasonal cycles in hydroenvironmental signals in monthly-based data analysis showed the need to study the seasonal-based data. Each dataset with seasonal timescale was decomposed into five components including an approximation (A4) and four detail sub-series, 6-month periodicity (D1), 12-month periodicity (D2), 24-month periodicity (D3), and 48-month periodicity (D4). The D2 sub-series with 12-month periodicity could help to explain the trends which were found in the monthly-based analysis, in annual timescale.
According to the results (Table 3), the water level had a significant negative trend in the original dataset

Annual data analysis
The possibility of longer-time periodicities' presence which was obvious in monthly-and seasonally-based data analysis as well as the fact that higher-frequency components affect many of the monthly-and seasonally-based datasets, led the study to investigate the lower-frequency components of the time series. Therefore, using DWT, each annuallybased dataset was decomposed into two levels with an approximation (A2), and two detail sub-series (D1 and D2). D1 and D2 corresponded to the 2-year and 4-year variations, respectively.
According to the results (Table 4)

CONCLUSION
The current paper applied ANNs, wavelet transforms, and Mann-Kendall trend tests to RS-based datasets of the water level, precipitation, LST, NDVI, and ET over the Lake Urmia basin from 1995 to 2019 in order to analyze the time series to determine the most important periodicities of the hydro-environmental variables over the basin and to understand the climate change and human-landscape drivers of the major decline in the water level of Lake Urmia.
There were some limitations in this study, which solving them may reveal new aspects of basins to researchers. Studying the water inflow into a lake from rivers which the current study did not have access to could strengthen future studies. As well, investigating groundwater level to study its interactions with the water-level fluctuations of a lake is a suggestion for future studies.
Given the stable condition of the precipitation during the study period, it was concluded that the change in precipitation is not the main reason for the decline in the water level. Instead, a significant positive trend was seen in LST. The increasing temperature will lead to increasing evaporation from the lake's surface and, consequently, it will intensify the decline of the water level. Also, the positive trend in NDVI and ET time series in the conditions that the precipitation was not increasing supported the hypothesis that the human interventions over the basin in terms of agriculture and agricultural activities are the major reasons for the decline in the water level. These conclusions are in agreement with other research experiences over other parts of the world that found human land-use and climate change as the major reasons for the water resources problems.
Although the periodic components that had the major effects on the trends were not the same for all of the hydroenvironmental variables, a general conclusion was made.

DATA AVAILABILITY STATEMENT
Data cannot be made publicly available; readers should contact the corresponding author for details.