Application of Hybrid Neural Network Techniques for Drought Forecasting

Introduction

Among many natural disasters, drought is one of the major disasters causing severe havocs in regions around the world. It is one of the most complicated natural hazards and has broadly adverse effect on water resources, economy, tourism, agriculture, and ecosystem (Wambua et al., 2016; Maca and Pech. 2015). It is a natural portion of climate which happens in both low and high precipitation regions and virtually all climatic systems (Wilhite and Buchanan, 2005, Wilhite, 2009). Temperature and precipitation are very significant natural elements or variables; and several studies revealed that they are important aspects affecting drought intensity (Sun and Ma, 2015; Easterling et al., 2007). Prediction of upcoming dry events in an area is very essential to find viable answers regarding assessment of risk related to drought occurrences and management of water (Bordi and Sutera, 2007). Moreover, forecasting drought conditions play a vital part in mitigating impact of drought on water management (Kim and Valdes, 2003). Occurrence of aridity is a perpetual climatic factor, whereas drought is a temporary irregularity (Zhang and Lin, 2016; Ndehedehe et al., 2016).

An objective drought condition assessment in a specific region is preliminary step for water resources planning for preventing and mitigating adverse effects of future happenings. Temporal and spatial severity and extent of drought can be found out with support of these measures (McKee et al., 1993; Palmer, 1995; Guttmann, 1998; Edwards and Mckee, 1997; Hayes, 2000). McKee et al. (1993) developed SPI, which is an efficient drought index having numerous benefits compared to others. SPI calculation is very easy than more complicated drought index, for example Palmer Drought Severity Index (PDSI; Palmer, 1965), as SPI necessitates rainfall data only, while PDSI utilizes many parameters. SPI classifies different kinds of drought as environmental, agricultural, or hydrological and has been broadly utilized for analysis of drought events occurring at many parts around the world.

Some specific applications of ANN in water resources involve forecasting river flows (Dibike and Solomatine, 2001; Imrie et al., 2000; Mohanta et al., 2020a), model evapotranspiration (Trajkovic et al., 2003; Sudheer et al., 2002). modelling sediment yield and runoff (Agarwal et al., 2006; Mohanta et al., 2020b; Samantaray and Sahoo,

2020a, d; Samantaray et al., 2019a; Samantaray et al., 2020a), water quality modelling (Milot et al.. 2002; Schmid and Koskiaho, 2006; Keskin et al., 2015), groundwater level prediction (Maiti and Tiwari, 2014; Daliakopoulos et al., 2005; Samantaray et al., 2019b; Samantaray et al., 2020b; Sridharam et al., 2020a), estimating sediment concentration (Nagy et al., 2002; Samantaray and Sahoo, 2020b), and rainfall-runoff process (Jeong and Kim, 2005; Samantaray and Sahoo, 2020c, Jimmy et al., 2020). Mishra et al. (2007) urbanized an amalgam model, conjoining a linear stochastic model and a non-linear ANN model for forecasting drought conditions in River Kansabati, India, and compared its performance with individual stochastic and ANN models using SPI. Observation of findings reveals that proposed hybrid model produced drought forecasts with superior accuracy. Ali et al. (2017) investigated application of multilayer perceptron (MLP) algorithm for drought forecasting on the basis of SPI at Northern Area and KPK, Pakistan. On the basis of different evaluation criteria, results demonstrated that MLP has potential capability for SPI drought forecasting. Moghari and Araghinejad (2015) applied direct multi-step MLP, recursive multi-step MLP. direct multi-step RBF, recursive multi-step RBF, direct multi-step generalized regression neural network (DMSGRNN), recursive multi-step GRNN (RMSGRNN), and SPI time series approach for providing drought forecasting of Gorganroud basin situated in northern Iran. Outcomes based on performance indicators revealed that RBF and GRNN performed pre-eminent in drought index forecasting and drought class. Santos et al. (2009) applied ANN models to forecast drought of three areas in River San Francisco, Brazil. Results revealed that employed technique was proficient to forecast SPI. Ваша et al. (2010) developed aggregated drought index (ADI), RMSNN, and DMSNN to present a drought forecasting approach of River Yarra in Victoria (Australia) utilizing monthly time step. Findings from study indicated that DMSNN produced slightly better results compared to RMSNN. Djerbouai and Souag-Gamane (2016) explored use of ANN model along with combined W-ANN to forecast drought events in Algerois River, Algeria, and compared with traditional stochastic models. Outcomes showed that W-ANN model gave best performance for all SPI time series than simple ANN model. Morid et al. (2007) scrutinized usability of ANN method for both likelihood and severity of drought forecasting with different timescales. Ваша et al. (2012) developed DMSNN and RMSNN to classify drought condition of Yarra River in Victoria, Australia, based on non-linear ADI and compared with traditional ARIMA model. Results revealed that both the neural network models performed superiorly than ARIMA model. Rezaeianzadeh et al. (2016) established an ANN model on the basis of hydro-climatic parameters for forecasting inflow volume on monthly basis of succeeding month and evaluating Markov chain outcomes to forecast drought events of Doroodzan watershed in Fars Province, Iran. Findings from study revealed that both models predicted drought events accurately at proposed study area. Borji et al. (2016) assessed efficacy of ANN and SVM models to predict streamflow drought index (SDI) of Latian catchment situated in Iran. SVM approach provided better efficiency and reliability compared to ANN in forecasting long-term droughts. Belayneh and Adamowski (2012) compared and examined efficiency of ANN, SVM, and W-ANN to forecast drought events based on SPI in Awash Basin of Ethiopia. Forecasting results indicated that W-ANN gave best SPI forecasting values over multiple time series at desired study area.

In past decades, ANN was a successive tool commonly applied to the prediction of various hydrologic parameters in different watersheds (Mohanta, 2020a; Sahoo et al. 2019, 2020a, b; Samantaray and Ghose, 2019; Sridharam et al„ 2020b). Mishra and Desai (2006) calculated SPI for multiple timescales and compared RMSNN and DMSNN with ARIMA for drought forecasting of River Kansabati located in Purulia district of West Bengal. India. Findings from the study revealed that RMSNN produced better results for a short lead time, while DMSNN model outperformed RMSNN and ARIMA models for a longer lead time to forecast drought conditions. Bacanli et al. (2009) explored applicability of adaptive neuro-fuzzy inference system (ANFIS) to forecast drought events and quantitative value of SPI at Central Anatolia, Turkey. Results demonstrated that ANFIS provided more precise and consistent results in forecasting drought events and can be effectively utilized. Shirmohammadi et al. (2013) developed and assessed potential of W-ANFIS and W-ANN models for drought forecasting in Ajabshir Plain, Iran, based on SPI and compared with simple ANN and ANFIS models. Comparative findings showed that W-ANFIS gave most accurate and reliable results follow'ed by W-ANN, ANFIS, and ANN. Mokhtarzad et al. (2017) investigated usability of SVM, ANFIS, and ANN techniques for finding a suitable drought forecasting model in Bojnourd city of Khorasan Province. Tehran, based on SPI. Results showed that SVM model produced more precise and consistent values for drought forecasting. Rezaeian-Zadeh and Tabari (2012) explored the ability of MLP model to forecast SPI values in different timescales of drought at five stations in Iran. Outcomes revealed that MLP4 gave better prediction results with superior efficacy compared to other MLPs. Hosseini-Moghari et al. (2017) applied recursive MLP and recursive SVM for multistep ahead drought forecasting in Gorganrood, Iran, on the basis of monthly time series of SPI and compared the accuracy of obtained results with ARIMA. Findings from study in accordance to performance indices suggested that ANN models outperformed traditional ARIMA model. Bari Abarghouei et al. (2013) applied ANN model to predict drought conditions in Ardakan region of Yazd province computing different time series of SPI. Results revealed that ANN proved to be a potential model in drought prediction with greater accuracy and reliability. Jalalkamali et al. (2015) employed MLP, ANFIS, SVM, and ARIMA models to develop a suitable drought forecasting model using SPI for Yazd Province, Iran. Results demonstrated that ARIMA model produced SPI values and drought forecasting outcomes with better precision compared to other ANN models. Kousari et al. (2017) utilized ANN model and SPI for developing a regional drought forecasting model for Fars Province of Iran.

Belayneh et al. (2014) compared efficiency of ARIMA, ANN, SVM, W-ANN, and W-SVM to forecast long-term drought events in River Awash located in Ethiopia. Results indicated that combined ANN models performed better than simple ANN and ARIMA models for long-term drought forecasting. Komasi et al. (2018) proposed W-SVM for drought forecasting utilizing SPI Urmia Lake located in Iran and also evaluated the effectiveness of cuckoo search based SVM model to model and forecast SPI time series. Obtained results indicated that W-SVM and CS-SVM models performed better compared to simple SVM model in SPI time series forecasting. Zahraie and Nasseri (2011) utilized SVM model to develop seasonal SPI forecasting models which acts as an indicator to severity of drought events for Tehran city in Iran. Results showed that SVM model predicted SPI values with good accuracy and reliability that can be utilized for long-term drought forecasting. Zhang et al. (2020) investigated and compared forecasting capabilities of WNN, ARIMA, and SVM models to forecast drought conditions in Sanjiang Plain. China, based on SPI. Comparative results based on different performance evaluation criteria revealed that ARIMA model performed better than proposed data-driven models for desired study area. Masinde (2014) proposed hybridization of ANN and effective drought index (EDI) approach for short- and long-term drought forecasting with severity of drought conditions in Kenya. Proposed hybrid model was found to be an effective and consistent model for drought forecasting with enhanced results.

Forecasting drought events and their severity plays a significant part in drought mitigation and its impact on water resources management. Since SPI is most extensively utilized techniques linked to drought among many other applied approaches, precise and consistent SPI estimation is very much vital.

Major objective of present research is to investigate potential of various data- driven methods for drought forecasting. For solving aforementioned purpose, this study aims at developing W-ANN and W-SVM models for forecasting SPI for different prediction time steps. Also, W-ANN and W-SVM models were compared with simple ANN and SVM models in the Bharuch, Porbandar, Surendranagar, India. Performance of developed models was assessed and compared utilizing standard statistical performance measures.