Original papers
Multi-stage committee based extreme learning machine model incorporating the influence of climate parameters and seasonality on drought forecasting

https://doi.org/10.1016/j.compag.2018.07.013Get rights and content

Highlights

  • Multi-stage committee extreme learning machine (Comm-ELM) model is constructed.

  • SPI forecast performed against committee PSO-ANFIS and multiple linear regression.

  • Climate parameters and seasonality factor incorporated for enhanced accuracy.

  • Comm-ELM model is evaluated in a drought prone, agricultural region of Pakistan.

  • Comm-ELM model is a versatile data intelligent tool for monthly drought modelling.

Abstract

Drought forewarning is an important decisive task since drought is perceived a recurrent feature of climate variability and climate change leading to catastrophic consequences for agriculture, ecosystem sustainability, and food and water scarcity. This study designs and evaluates a soft-computing drought modelling framework in context of Pakistan, a drought-stricken nation, by means of a committee extreme learning machine (Comm-ELM) model in respect to a committee particle swarm optimization-adaptive neuro fuzzy inference system (Comm-PSO-ANFIS) and committee multiple linear regression (Comm-MLR) model applied to forecast monthly standardized precipitation index (SPI). The proposed Comm-ELM model incorporates historical monthly rainfall, temperature, humidity, Southern Oscillation Index (SOI) at monthly lag (t − 1) and the respective month (i.e., periodicity factor) as the explanatory variable for the drought’s behaviour defined by SPI. The model accuracy is assessed by root mean squared error, mean absolute error, correlation coefficient, Willmott’s index, Nash-Sutcliffe efficiency and Legates McCabe’s index in the independent test dataset. With the incorporation of periodicity as an input factor, the performance of the Comm-ELM model for Islamabad, Multan and Dera Ismail Khan (D. I. Khan) as the test stations, was remarkably improved in respect to the Comm-PSO-ANFIS and Comm-MLR model. Other than the superiority of Comm-ELM over the alternative models tested for monthly SPI forecasting, we also highlight the importance of the periodicity cycle as a pertinent predictor variable in a drought forecasting model. The results ascertain that the model accuracy scales with geographic factors, due to the complexity of drought phenomenon and its relationship with the different inputs and data attributes that can affect the overall evolution of a drought event. The findings of this study has important implications for agricultural decision-making where future knowledge of drought can be used to develop climate risk mitigation strategies for better crop management.

Introduction

Drought is a socio-economic hazard posing severe threats to groundwater reservoirs, leading to a scarcity of water resources, crop failure and socio-economic challenges (Mpelasoka et al., 2008, Riebsame et al., 1991, Wilhite et al., 2000, Deo et al., 2009). A general scarcity of sufficient palatable water due to rapid growth of human populations and an expansion of agricultural, energy and industrial sectors is also a growing concern (IPCC, 2012, McAlpine et al., 2007). Climate change further escalates the potential severity and frequency of drought events (Deo et al., 2009, McAlpine et al., 2009). Large-scale climate mode indices including the Southern Oscillation Index (SOI) is significantly correlated with fluctuations in rainfall and onset of drought (Mishra and Singh, 2010, Morid et al., 2007, Nguyen-Huy et al., 2017, Özger et al., 2012). Hence drought forecasts that incorporate climate factors can assist hydrologists, agriculturalists and resource planners in strategic decisions to address socio-economic challenges posed by a drought, particularly in cases of prolonged events (Bates et al., 2008, Deo et al., 2016a, Mishra and Singh, 2010, Mishra and Singh, 2011, Wilhite and Hayes, 1998).

Traditionally, drought indices are used to measure, monitor and forecast drought. Several different indices have been developed according to their appropriateness for a given geographic region (Mishra and Singh, 2010, Mishra and Singh, 2011). The Palmer (1965) Drought Severity Index (PDSI) measures the overall dryness based on precipitation and temperature datasets. PDSI is particularly useful in pointing out long-term drought events and it is not appropriate for a region with generally high surface run-off (Mishra and Singh, 2010, Mishra and Singh, 2011). To tackle the complexities associated with PDSI, researchers have developed the Crop Moisture Index (CMI), particularly for assessing agricultural drought events (Palmer, 1968). CMI is based on the rank of the precipitation records and aims to compute both positive and negative precipitation anomalies. However, CMI in the short-term is insufficient to offset long-term issues. Byun and Wilhite (1999) developed the Effective Drought Index (EDI) based on daily precipitation data. EDI is a good index for operational monitoring of meteorological and agricultural drought, although the consideration of precipitation alone does not take into account other environmental parameters that also cause drought (e.g., the impact of temperature). Several drought indices have been developed built on PDSI to take into account additional data on precipitation and crop moisture, however the standardized precipitation index (SPI) is used universally as a standard metric (Deo et al., 2017a).

The importance of modelling SPI is derived from the notion that: (1) SPI is able to assess the water shortage situations built on a statistical distribution of rainfall that can enable both short-term (i.e. monthly) and long-term (i.e., seasonal and annual) drought assessments (ranging from 1 up to 48 months). (2) SPI is presented as a normalized standard metric of rainfall surpluses/deficits in relation to a benchmark climatological period (Hayes et al., 1999, McKee et al., 1993, Yuan and Zhou, 2004), and therefore, it can enable a comparison of the drought behaviour in geographically and climatologically diverse regions. (3) SPI has been explored and validated for drought mitigation studies in diverse climatic regions (Almedeij, 2016, Choubin et al., 2016, Svoboda et al., 2012). (4) SPI has the potential to represent both short (1 and 3 months) and long-term (6–12 months) drought in a probabilistic fashion, largely on multiple timescales. In the case of agricultural drought, the SPI makes it possible to examine the soil moisture status with respect to precipitation anomalies on a comparatively short timescale using hydrological reservoirs to replicate long-term climate anomalies (Svoboda et al., 2012). Due to the advantageous features, the SPI is an ideal metric for the management of not only hydrological, but also for agricultural drought events (Deo et al., 2017a, Guttman, 1999). In this study, SPI based on 1 month is utilized for drought forecasting as there is no study on a monthly drought in the selected study regions, although the study of Ali et al. (2018a) designed a study using ensemble-ANFIS model for long term (3–12 months) drought forecasting and Ali et al. (2018b) developed a multi-stage hybridized online sequential extreme learning machine integrated with Markov Chain Monte Carlo copula-Bat algorithm for rainfall forecasting, both focussed in Pakistan.

In the existing literature, data-driven models have been used to model SPI for drought forecasting. For example, an SPI-based methodology was designed by Cancelliere et al. (2007) to forecast probabilistic drought alterations in Sicily, Italy. Jalalkamali et al. (2015) in Yazd, Iran conducted a study to forecast the SPI using several model variants, including a multilayer perceptron artificial neural network (MLP ANN), an adaptive neuro-fuzzy inference system (ANFIS), support vector machines (SVM), and an autoregressive integrated moving average (ARIMA) multivariate model. In another study, models based on ANFIS and ANN Wavelet tools were adopted by Shirmohammadi et al. (2013) to forecast SPI in Azerbaijan. An SPI-based forecasting study was also performed by Santos et al. (2009) using an ANN model for San Francisco, USA. Drought forecasts using SPI have an extensive history in the current literature (Adamowski et al., 2012, Bonaccorso et al., 2003, Cancelliere et al., 2006, Choubin et al., 2016, Deo et al., 2017b, Guttman, 1999, Hayes et al., 1999, Jalalkamali et al., 2015, Moreira et al., 2015, Moreira et al., 2008, Paulo and Pereira, 2007, Sönmez et al., 2005). However, SPI based drought forecasts are yet to be explored for agricultural regions in Pakistan where the influence of drought is a major impeding factor for crop productivity and farmers’ livelihoods.

Drought hazard continues to severely affect the agriculturally dependent nation of Pakistan (Report, 1950-2015). The severe drought event of 1998 led to a significant reduction in Pakistan’s national agricultural productivity by 2.6% over 2000–2001 (Ahmad et al., 2004). In spite of such pressing issues, drought models for local agricultural zones in Pakistan have been very limited: (1) Khan and Gadiwala (2013) aimed to investigate drought behaviour using SPI at multi timescales for the province of Sindh: (2) Xie et al. (2013) applied a spatiotemporal variability analysis based on the SPI data to also forecast drought behavior, and most recently, (3) the study of Ali et al. (2017) forecasted drought based on Standardized Precipitation-Evapotranspiration Index (SPEI) where a multilayer perceptron-based artificial neural network model was employed. (4) The study of Ahmed et al. (2016) has utilized antecedent SPI data for the characterization of future seasonal drought events in Balochistan, Pakistan and (5) Ali et al. (2018a) implemented an ensemble strategy based on the ANFIS model to forecast the SPI using the historical SPI to forecast future SPI. While it was not specifically on drought forecasting, a recent study of Ali et al. (2018b) has forecasted rainfall in Pakistan using a multi-stage online sequential extreme learning machine integrated with Markov Chain Monte Carlo copula-Bat algorithm. However, there has been no study in Pakistan specifically on future drought models or drought indices utilising different climatological parameters.

Considering a lack of drought models for developing nations like Pakistan, the aim of this research is to develop and evaluate Comm-ELM (a data-intelligent model) with universal approximation capabilities. The specific objectives are:

  • (i)

    To develop a committee based extreme learning machine (Comm-ELM) model following its successful application elsewhere, e.g., (Barzegar et al., 2018, Prasad et al., 2018) and evaluate its preciseness for forecasting future SPI incorporating the relevant hydro-meteorological dataset (i.e., temperature, rainfall, humidity, Southern Oscillation Index) and the seasonality metric (i.e., periodicity) as predictors for the period of 1081–2015. Here, a multi-stage model strategy is adopted incorporating antecedent climate-based variables at (t − 1) and corresponding periodicity in the first stage, and the forecasted SPI.

  • (ii)

    To elucidate the importance of periodicity as a pertinent factor in drought forecast models and seasonal influences on drought progression using data monitoring for Pakistan, a nation vulnerable to significant agricultural and water management issues.

  • (iii)

    To compare the performances of Comm-ELM in respect to Comm-PSO-ANFIS and Comm-MLR models for SPI-forecasting.

Section snippets

Standardized precipitation index (SPI)

The SPI drought forecasting metric, centred on normalized probabilities of dryness relative to a base climatology, provides a depiction of irregular wetness and dryness situation. Before developing the forecast model for the designated regions in Pakistan, the monthly SPI index was calculated using precipitation (PCN) data (Deo et al., 2017a, McKee et al., 1993). Pearson Type III distribution/gamma distribution function is given by the following mathematical expression:g(PCN)=1βαΓ(α)PCNα-1e-x/β

Climate data

The climate data utilized in this study includes precipitation (PCN), mean monthly temperature (T) and mean monthly relative humidity (H). These parameters were obtained from the Pakistan Meteorological Department (PMD, 2016) for the period 1981–2015. Prior to developing the drought forecasting model, the monthly SPI index was computed using the rainfall data from 1981 to 2015.

As machine learning models rely on a prognostic features in historical data to forecast the future drought, the climate

Results

A scatterplot diagram (i.e., Fig. 5) was constructed using the forecasted and observed SPI data in the testing phase to compare the performance of the Comm-ELM vs. the Comm-PSO-ANFIS and Comm-MLR model with periodicity factors included for all three study sites. Further, to analyse the model performance more closely, a scatterplot showing the goodness-of-fit and its correlation coefficient (r) is shown to depict the extent of agreement between forecasted and observed monthly SPI. The Comm-ELM

Further discussion

Data-driven models are likely to become important decision-support tools in major decision-making for hydrology, agriculture and water science, aiming to address drought-related issues arising from the imminent global warming. The increasing risk of water scarcity remains a problem in first world nations as well as developing nations such as Pakistan. This work has developed the most accurate drought forecasting models to date for three locations in Pakistan utilizing Comm-ELM against the

Conclusion

Due to serious impacts of drought peril, the SPI based drought forecasting can be considered as a promising tool to deal with future hazards implicated on horticulture, water administration, water demand, pricing and policy. In this study, Comm-ELM model was benchmarked with two alternative techniques: the Comm-PSO-ANFIS and Comm-MLR based models where the monthly SPI forecasting model was constructed and evaluated for three agricultural regions in Pakistan where farming activity is a critical

Acknowledgment

This study was supported by the University of Southern Queensland Office of Graduate Studies Postgraduate Research Scholarship (2017–2019). Data were acquired from Pakistan Meteorological Department, Pakistan, which are duly acknowledged. The authors express their sincere gratitude to all reviewers and the Editor for their constructive comments that has improved the final version of this paper.

References (123)

  • M.K. Goyal et al.

    Modeling of daily pan evaporation in sub tropical climates using ANN, LS-SVR, Fuzzy Logic, and ANFIS

    Expert Systems with Appl.

    (2014)
  • G.-B. Huang et al.

    Extreme learning machine: theory and applications

    Neurocomputing

    (2006)
  • B. Karthika et al.

    Prediction of air temperature by hybridized model (Wavelet-ANFIS) using wavelet decomposed data

    Aquatic Procedia.

    (2015)
  • A.K. Mishra et al.

    A review of drought concepts

    J. Hydrol.

    (2010)
  • A.K. Mishra et al.

    Drought modeling–A review

    J. Hydrol.

    (2011)
  • K. Mohammadi

    A new hybrid support vector machine–wavelet transform approach for estimation of horizontal global solar radiation

    Energy Convers. Manage.

    (2015)
  • E.E. Moreira et al.

    SPI-based drought category prediction using loglinear models

    J. Hydrol.

    (2008)
  • J.E. Nash et al.

    River flow forecasting through conceptual models part I—A discussion of principles

    J. Hydrol.

    (1970)
  • P.C. Nayak et al.

    A neuro-fuzzy computing technique for modeling hydrological time series

    J. Hydrol.

    (2004)
  • T. Nguyen-Huy et al.

    Copula-statistical precipitation forecasting model in Australia’s agro-ecological zones

    Agric. Water Manag.

    (2017)
  • R. Prasad et al.

    Ensemble committee-based data intelligent approach for generating soil moisture forecasts with multivariate hydro-meteorological predictors

    Soil Tillage Res.

    (2018)
  • C.R. Reeves

    A genetic algorithm for flowshop sequencing

    Comput. Oper. Res.

    (1995)
  • M. Şahin et al.

    Comparison of ANN and MLR models for estimating solar radiation in Turkey using NOAA/AVHRR data

    Adv. Space Res.

    (2013)
  • 2016. Climate change impacts on capital Islamabad. The Times of...
  • J. Abbot et al.

    Application of artificial neural networks to rainfall forecasting in Queensland, Australia

    Adv. Atmos. Sci.

    (2012)
  • J. Adamowski et al.

    Comparison of multiple linear and nonlinear regression, autoregressive integrated moving average, artificial neural network, and wavelet artificial neural network methods for urban water demand forecasting in Montreal, Canada

    Water Resources Research.

    (2012)
  • M. Afzal et al.

    Influence of North Atlantic oscillations and Southern oscillations on winter precipitation of Northern Pakistan

    Pakistan J. Meteorol.

    (2013)
  • S. Ahmad et al.

    Drought Mitigation in Pakistan: Current Status and Options for Future Strategies

    (2004)
  • K. Ahmed et al.

    Characterization of seasonal droughts in Balochistan Province, Pakistan

    Stochastic Environ. Res. Risk Assess.

    (2016)
  • M. Ali et al.

    An ensemble-ANFIS based uncertainty assessment model for forecasting multi-scalar standardized precipitation index

    Atmos. Res.

    (2018)
  • M. Ali et al.

    Multi-stage hybridized online sequential extreme learning machine integrated with Markov Chain Monte Carlo copula-Bat algorithm for rainfall forecasting

    Atmos. Res.

    (2018)
  • Z. Ali

    Forecasting drought using multilayer perceptron artificial neural network model

    Adv. Meteorol.

    (2017)
  • J. Almedeij

    Long-term periodic drought modeling

    Stochastic Environ. Res. Risk Assess.

    (2016)
  • Amir, I., 2012. Tough times ahead for farmers in southern KP....
  • Apaydın, A., Kutsal, A., Atakan, C., Ankara, 1994. The Statistics in practice. Hacettepe...
  • Bates, B., Kundzewicz, Z.W., Wu, S., Palutikof, J., 2008. Climate change and water: Technical paper vi....
  • M. Behmanesh et al.

    Chaotic time series prediction using improved ANFIS with imperialist competitive learning algorithm

    Int. J. Soft Comput. Eng.

    (2014)
  • B. Bonaccorso et al.

    Spatial variability of drought: an analysis of the SPI in Sicily

    Water Resources Manage.

    (2003)
  • H.-R. Byun et al.

    Objective quantification of drought severity and duration

    J. Clim.

    (1999)
  • A. Cancelliere et al.

    A non-parametric approach for drought forecasting through the standardized precipitation index

    Metodi statisticie matematici per l Analisi delle serie idrologiche.

    (2006)
  • A. Cancelliere et al.

    Drought forecasting using the standardized precipitation index

    Water Resour. Manage.

    (2007)
  • T. Chai et al.

    Root mean square error (RMSE) or mean absolute error (MAE)?–Arguments against avoiding RMSE in the literature

    Geoscientific Model Develop.

    (2014)
  • G.-C. Chen et al.

    Particle swarm optimization algorithm

    Inform. Control-Shenyang-.

    (2005)
  • G. Civelekoglu et al.

    Prediction of bromate formation using multi-linear regression and artificial neural networks.

    Ozone: Sci. Eng.

    (2007)
  • R.C. Deo et al.

    An extreme learning machine model for the simulation of monthly mean streamflow water level in eastern Queensland

    Environ. Monitor. Assess.

    (2016)
  • R.C. Deo et al.

    Application of effective drought index for quantification of meteorological drought events: a case study in Australia

    Theor. Appl. Climatol.

    (2016)
  • R.C. Deo et al.

    An extreme learning machine model for the simulation of monthly mean streamflow water level in eastern Queensland

    Environ. Monitor. Assess.

    (2016)
  • R.C. Deo et al.

    Forecasting effective drought index using a wavelet extreme learning machine (W-ELM) model

    Stoch. Env. Res. Risk Assess.

    (2016)
  • R.C. Deo et al.

    Forecasting effective drought index using a wavelet extreme learning machine (W-ELM) model

    Stochastic Environ. Res. Risk Assess.

    (2017)
  • R.C. Deo et al.

    Impact of historical land cover change on daily indices of climate extremes including droughts in eastern Australia

    Geophys. Res. Lett.

    (2009)
  • Cited by (61)

    View all citing articles on Scopus
    View full text