Capability and robustness of novel hybridized models used for drought hazard modeling in southeast Queensland, Australia

Highlights

• An integrated agricultural-meteorological drought was spatially modeled.

• The hybridized drought hazard model ANFIS-BA performed best.

• Hybrid models prevent overfitting and can improve ANFIS performance.

• The most important factors for spatial drought hazard were determined.

Abstract

Widespread detrimental and long-lasting droughts are having catastrophic impacts around the globe. Researchers, organizations, and policy makers need to work together to obtain precise information, enabling timely and accurate decision making to mitigate drought impacts. In this study, a spatial modeling approach based on an adaptive neuro-fuzzy inference system (ANFIS) and several metaheuristic optimizations (ANFIS-BA, ANFIS-GA, ANFIS-ICA, ANFIS-PSO) was developed to predict the spatial occurrence of drought in a region in southeastern Queensland, Australia. In this approach, data describing the distribution of eight drought-contributing factors were prepared for input into the models to serve as independent variables. Relative departures of rainfall (RDR) and relative departures of soil moisture (RDSM) were analyzed to identify locations where drought conditions have occurred. The set of locations in the study area identified as having experienced drought conditions was randomly divided into two groups, 70% were used for training and 30% for validation. The models employed these data to generate maps that predict the locations that would be expected to experience drought. The prediction accuracy of the model-produced drought maps was scrutinized with two evaluation metrics: area under the receiver operating characteristic curve (AUC) and root mean square error (RMSE). The results demonstrate that the hybridized models (ANFIS-BA (AUC_mean = 83.7%, RMSE_mean = 0.236), ANFIS-GA (AUC_mean = 81.62%, RMSE_mean = 0.247), ANFIS-ICA (AUC_mean = 82.12%, RMSE_mean = 0.247), and ANFIS-PSO (AUC_mean = 81.42%, RMSE_mean = 0.255)) yield better predictive performance than the standalone ANFIS model (AUC_mean = 71.8%, RMSE_mean = 0.344). Furthermore, sensitivity analyses indicated that plant-available water capacity, the percentage of soil comprised of sand, and mean annual precipitation were the most important predictors of drought hazard. The versatility of the new approach for spatial drought modeling and the capacity of ANFIS model hybridization to improve model performance suggests great potential to assist decision makers in their formulations of drought risk, recovery, and response management, and in the development of contingency plans.

Introduction

Drought is Earth’s most widespread hazard (Mastrangelo et al., 2012). Whether naturally occurring or a product of human-induced climate change, drought adversely effects various environments and social systems (IPCC (Intergovernmental Panel on Climate Change), 2013, Wang et al., 2018, Di Baldassarre et al., 2017). Drought arises from abnormal temporal or spatial reductions of precipitation, abnormal rates of evaporation, or increased extraction of water resources by people (Huang and Chou, 2008, Huang et al., 2015). The spatial extent of drought is often greater than any other natural event, and thus the corresponding damages attributable to a drought event can be expected to be large (Mishra and Singh, 2010, Mishra and Singh, 2011, Xu et al., 2014). Therefore, drought monitoring and prediction have attracted the attention of policy makers to acquire greater insight or improve response to drought events (Luo and Wood, 2007).

Droughts are often classified as either meteorological, agricultural, hydrological, or socioeconomic events (American Meteorological Society, 2004). A reduction in precipitation results in meteorological drought. This is generally defined as occurrence of lower than normal precipitation amounts over a given period. Meteorological drought affects soil moisture and leads to agricultural drought, a shortage of water available for plant growth. Hydrological drought refers to deficiencies of surface and/or subsurface water supplies. And socioeconomic droughts occur when there is insufficient supply to meet human water demands. It is often difficult to separate these droughts from each other, as they may occur sequentially or simultaneously and are often interconnected (Mo, 2008, Mo and Lettenmaier, 2014, Hao and Singh, 2015).

From a drought management viewpoint, it is necessary to define and understand the terms ‘drought hazard’, ‘drought vulnerability’, and ‘drought risk’. IPCC (2012) defines a hazard as “the potential occurrence of a natural or human-induced physical event that may cause loss of life, injury, or other health impacts, as well as damage and loss to property, infrastructure, livelihoods, service provision, and environmental resources”. Drought vulnerability pertains to the characteristics of a place or system that make it susceptible to suffering the consequences of drought (Naumann et al., 2014). Drought vulnerability includes both biophysical and socio-economic drivers of drought impacts and requires a determination of the degree of susceptibility to a drought hazard and also the capacity to cope with drought (Biazin and Sterk, 2013). Risk is “the likelihood over a specified time period of severe alterations in normal functioning of a community or a society due to hazardous physical events interacting with vulnerable social conditions, leading to widespread adverse human, material, economic, or environmental effects that require immediate emergency response to satisfy critical human needs and that may require external support for recovery” (IPCC, 2012). Therefore, drought risk is a function of drought vulnerability to hazardous conditions and exposure to a drought hazard (Wilhite, 2000).

As a result of the insidious nature of drought events, numerous studies have assessed drought hazard around the world, mostly using time series-based drought indices to monitor and evaluate meteorological drought characteristics (e.g., Mishra and Singh, 2010, Dayal et al., 2017, Deo et al., 2017a). In a comprehensive review, Mishra and Singh (2011) describe drought modeling approaches that range from simplistic to more complex models, and they discuss their corresponding advantages and limitations. Drought risk has also been assessed using statistical multivariate joint distribution in a time series approach, e.g., using copula (Zhang et al., 2013, Nguyen-Huy et al., 2017, Ali et al., 2018). A number of drought indices have been applied in different studies to characterize and predict different types of drought. The most commonly used of these drought indices are the standardized precipitation index (SPI) (e.g., McKee et al., 1993, Edwards, 1997, Uddameri et al., 2019), the soil moisture drought index (SMDI) (e.g., Hollinger et al., 1993, Sohrabi et al., 2015, Carrão et al., 2016), the soil moisture deficit index (SMDI) (e.g., Narasimhan and Srinivasan, 2005, Yang et al., 2017), the soil wetness deficit index (SWDI) (e.g., Keshavarz et al., 2014), the standardized runoff index (SRI) (e.g., Shukla and Wood, 2008), the standardized precipitation evapotranspiration index (SPEI) (Vicente-Serrano et al., 2010), the linearly combined drought index (LDI) (Mo and Lettenmaier, 2014, Hao et al., 2016), the aggregated drought index (ADI) (Keyantash and Dracup, 2004), and the multivariate standardized drought index (MSDI) (Hao and AghaKouchak, 2013). This approach uses drought indices based on interpolation techniques for generalizing weather station data (McLeman et al., 2010). However, such indices are station-based and computed on a point-to-point scale, and they are thus not representative of the spatial variations of droughts. Consequently, they generate relatively high uncertainty in interpolated areas (Brown et al., 2008, Swain et al., 2011). This circumstance has yielded combined applications of remote sensing techniques and drought indices for drought monitoring. Rhee et al. (2010) proposed a new remote sensing-based scaled drought condition index (SDCI) for agricultural drought monitoring. This index uses multi-sensor data comprising normalized difference vegetation index (NDVI) data and land surface temperature (LST) data from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor, and precipitation data from the Tropical Rainfall Measuring Mission (TRMM) satellite. Gouveia et al. (2017) used monthly NDVI and SPEI data at different time scales (1–24 months) to analyze drought impacts on vegetative cover in the Mediterranean region. Nicolai-Shaw et al. (2017) applied satellite-derived soil moisture observations to quantify the relationships between soil moisture drought and evapotranspiration, precipitation, temperature, and vegetation, and demonstrated the usefulness of satellite-based soil moisture for drought assessment. However, there is limited information in the remotely sensed soil moisture signal of the deeper root zone which produces inaccurate predictions of drought conditions. Mariano et al. (2018) investigated biomass anomalies using leaf area index (LAI) and MODIS data to detect trends in drought and land degradation in Brazil, and concluded that there are advantages in and great prospects for remote sensing-based approaches. However, there are also some major challenges as regards unquantified uncertainty, data continuity, community acceptability, sensor changes, and short-period data series (e.g., a decade, which is insufficient for meaningful studies of droughts from a climatological perspective) (AghaKouchak et al., 2015, Liu et al., 2016). In another approach, researchers have used a multi-criteria decision analysis (MCDA) method (e.g., analytical heirarcy process (AHP)) to analyze the relationships between predictor variables, which is based on a questionnaire/survey and experts’ opinions (e.g., Chen and Yang, 2011, Palchaudhuri and Biswas, 2016). A major drawback is that the AHP method allows subjective judgments by decision-makers.

Machine learning (ML) algorithms have been attracting attention as methods to extract knowledge in different fields involving natural processes (Rahmati et al., 2019a, Tien Bui et al., 2019, Moayedi et al., 2019). ML models have valuable advantages which can overcome the above-mentioned limitations. According to Pal and Mather (2005), state-of-the-art ML procedures, if implemented correctly, have higher accuracy than conventional parametric approaches. Machine learning approaches have the capacity to effectively model non-linear and highly dimensional data with intricate connections and missing values (Recknagel et al., 2000, Knudby et al., 2010). As a result, ML models have been successfully employed in environmental and natural hazard studies, such as mapping areas susceptible to land subsidence (e.g., Bui et al., 2018a, Rahmati et al., 2019b), landslides (e.g., Trigila et al., 2013, Goetz et al., 2015, Pham et al., 2016), gully erosion (e.g., Pourghasemi et al., 2017, Garosi et al., 2019), and floods (e.g., Darabi et al., 2019). A ML application on time series-based predictors has been used to make drought predictions (e.g., Deo et al., 2017b, Deo et al., 2017c, Prasad et al., 2017, Ali et al., 2018, Dayal et al., 2018). In these models, drought hazard is the dependent variable, while climatological and environmental factors are independent variables. ML algorithms make it possible to analyze complicated and non-linear relationships between drought events and other drought-affecting factors (environmental, topographical, hydrological, etc.). Farokhnia et al. (2011) used sea surface temperature (SST) and sea level pressure (SLP) as inputs to an adaptive neurofuzzy inference system (ANFIS) model to forecast possible droughts in Tehran three, six, and nine months in advance. They assessed the performance of the SST/SLP data sets and the ANFIS model according to a “drought/wet” classification, and concluded that in more than 90% of cases, the ANFIS model detected the drought status correctly or with only a one-class error (Farokhnia et al., 2011). Therefore, the efficiency and feasibility of this approach has been confirmed in the literature and our study is not intended to evaluate this approach (Deo et al., 2017a, Deo et al., 2017b, Prasad et al., 2017). Instead, our study fills a research gap in the ML-based approach for modeling drought hazard used in previous studies that used data from only a few locations (i.e., weather station records) and often ignored the variations in topo-hydrological and the physical properties of the soils in a given study area.

The ANFIS algorithm is a type of neurofuzzy approach that uses an adaptive network for learning and is popular for forecasting time series in water resources management and hydrological modeling (Aqil et al., 2007, Noori et al., 2010). ANFIS is a powerful model that offers great prospects for solving complicated and non-linear problems (Keshtegar et al., 2018, Zare and Koch, 2018). Its main advantages are fuzziness, its handling of temporally or spatially inadequate data, and the ability of the neural network to analyze complicated relationships between dependent and independent variables (Farokhnia et al., 2011). However, numerous studies have also shown that a standalone ANFIS model is prone to a degree of over-fitting, mainly due to intricacies caused by the adjustments of parameters to achieve global optima (Dehnavi et al., 2015, Jaafari et al., 2019b). Recently, hybridized ANFIS models, in which numerous predictive models can collaborate on the same task, have been used in geo-environmental and hydrological studies (e.g., Moretti et al., 2015, Kalantari et al., 2019a). The main reason for developing these hybridized models is that they assimilate a tuning optimization algorithm with the intelligent predictive ANFIS model. However, no previous study has systematically scrutinized the efficiency of hybridized ANFIS models for spatial drought hazard modeling. Specific objectives of this study are thus to: (1) investigate the utility of the ANFIS model for spatially modeling drought hazard; (2) combine the ANFIS model with each of four popular optimization algorithms (Bee algorithm (BA), imperialistic competitive algorithm (ICA), genetic algorithm (GA), and particle swarm optimization (PSO)) to generate novel hybridized models; (3) compare the performance of the standalone ANFIS model and the hybridized models using statistical evaluation metrics; and (4) assess the relative importance of the independent predictor variables. The aim is to acquire in-depth insights into drought hazard mapping by analyzing past drought events.