Multi-Point Geostatistical Generation of Water-Level Changes of Urmia Lake

Abstract:

Urmia Lake, which is located in the north-west part of Iran, has been affected by droughts during the last decades. Its water level and area have decreased significantly. Forecasting future water levels is crucial for navigation, water-resource management, and agricultural purposes. Moreover, exploring the uncertainty of water-level forecasts is essential for optimal water-management decisions. Unlike most time-series prediction methods, such as Autoregressive Integrated Moving Average (ARIMA), Generalized Autoregressive Conditional Heteroskedasticity (GARCH), and Support Vector Machines (SVM), geostatistical simulation methods quantify the uncertainty by generating various equally probable realizations.

Water-level simulation is related to factors such as temperature and precipitation (i.e., auxiliary information). In this study, to have better forecasting of water level, radar altimetry time series of Urmia Lake for more than 20 years, as well as an ensemble of auxiliary information, are used to simulate the water level time series. Quick Sampling (QS), a multiple-point geostatistics (MPS) method, is used to forecast and simulate multivariate time series of lake levels data based on training data coming from historical measurements.

In order to validate the ability of the proposed method in forecasting, the data set is divided into training and validation. Visual comparison and quantitative criteria, such as Root Mean Square Error (RMSE), demonstrate that QS can forecast water-level time series while preserving non-linear patterns better than sparse-kernel-machines methods such as Least Square-SVM (LS-SVM).