Modeling and Predicting the Drought Indices Time Series Using Machine Learning Methods In Order To Managing Hazards(Case Study: Eastern District of Isfahan)

The drought has been known as a complex and perilous phenomenon at the whole of the world especially in Iran. Determining and predicting its severity can be effective at managing the hazards due to it. To determine the drought severity, the indices have been used that can be divided into two broad categories of meteorological (M) and remotely-sensed (RS) indices. The most important M index has been the standardized perception index (SPI), and the common RS indices have been those extracted from the vegetation index (NDVI) and land surface temperature (LST) index. For modeling time series behavior of these indices and also predicting their future values, the machine learning methods can indicate the high efficiency. This paper also aims to evaluate the performance of four important machine learning methods, i.e. neural network (NN), support vector regression (SVR), least squares support vector machine (LSSVM) and also an adaptive neuro fuzzy inference system (ANFIS) for modeling the M and RS indices of Eastern district of Isfahan during 2000 to 2014 and predicting their values at 2015 and 2016. The data used in this paper are the NDVI and LST time series of MODIS, and the rainfall time series of TRMM satellite of study area. At first, the vegetation condition index (VCI) and temperature vegetation index (TVX) have been built by NDVI and LST and 12-month SPI has been built by rainfall data. Next, the time series behavior of three these indices has been modeled by four aforementioned methods that according to the results, SVR has a highest efficiency and NN has a lowest efficiency among these methods. The speed performance of LSSVM and then ANFIS have been higher than the other methods. Finally by designing a fuzzy inference system (FIS), the drought severity at spring and summer of 2000 to 2016 has been monitored that the results have shown the normality of the spring in all years except 2000 and 2011 and severe drought in the summer in all years except for the four years 2000, 2010, 2011 and 2014. In fact, this research has aimed to present a strategy for modeling drought behavior and predicting and monitoring it at future using machine learning methods and the remotely-sensed and meteorological time series data and fusing them in a FIS system.