Elsevier

Journal of Hydrology

Volume 599, August 2021, 126373
Journal of Hydrology

Research papers
Intercomparison of downscaling methods for daily precipitation with emphasis on wavelet-based hybrid models

https://doi.org/10.1016/j.jhydrol.2021.126373Get rights and content

Highlights

  • Wavelet-based hybrid methods are proposed for downscaling daily precipitation.

  • Wavelet-based hybrid models are compared with key traditional and modern methods.

  • Wavelet-based methods perform well in capturing extreme events and overall accuracy.

Abstract

Downscaling of local daily precipitation from large-scale climatic variables is required for assessing the impact of climate change on hydrology and water resources. This study proposes wavelet transform (WT)-based Feed-Forward Neural Network (FF-NN) and Nonlinear Auto Regressive with exogenous inputs Network (NARX-NN) models for downscaling daily precipitation. The models are applied to a large river basin, the Krishna River basin, in the Indian subcontinent. Several climatic variables, including geo-potential heights, wind direction, vorticity, humidity, air temperature, mean sea level pressure, meridional velocity at surface, and 500hpa and 850hpa levels, are considered based on their statistical correlations. The results are evaluated using different performance measures and the ability of the models to capture the extreme events at five selected grid points (in different locations) having varying climatic characteristics is assessed. The performance of the proposed wavelet-based models is also compared with that of four different traditional and recent downscaling methods: Multiple Linear Regression (MLR), Statistical Downscaling Model (SDSM), Genetic Programming (GP), and Artificial Neural Networks (ANNs). The results reveal that the wavelet-based neural network models (WT-FF-NN and WT-NARX-NN) are robust compared to the other methods in terms of their ability to capture the regional precipitation patterns and the extreme events. The improvement in the wavelet-based models can be attributed to their ability to unravel the hidden relationship between the predictors and precipitation. It is also observed that there is considerable increase in the correlation between precipitation and the decomposed climatic variables. All these results suggest that wavelets aid in unravelling the relationship between local precipitation and large-scale climatic variables and improving the overall performance of the downscaling models.

Introduction

Studying the impact of global warming-induced climate change on water resources has become a key area of research in recent decades. To project future climate change scenarios, General Circulation Models (GCMs), which are computer models working at the global scale, are commonly used. These models generally produce climate outputs (e.g. precipitation, temperature) at coarser spatial scales, typically having horizontal grid size ranging from 100 km to 600 km. Such coarse-scale climate outputs, however, are not appropriate for most hydrologic studies, since hydrologic studies generally require data at much finer spatial scales corresponding to catchments. Therefore, to properly study the impact of climate change on hydrology and water resources, it is required to downscale the coarse-scale climate outputs to catchment-scale hydrologic variables (Sehgal et al., 2018, Wilby et al., 1998).

Downscaling of GCM outputs assumes that the local climate is a function of both the global atmospheric features (large-scale climatic conditions) and the local conditions (topography, land surface characteristics). Generally, deriving local climatic characteristics from GCM outputs through downscaling is a multistep process involving a multitude of assumptions and approximations. The methods for downscaling can be broadly classified into (1) Dynamic downscaling, and (2) Statistical downscaling.

Dynamical downscaling uses an appropriate GCM to generate regional climate models (RCMs) to be used at regional scales (Srinivas et al., 2014). The RCMs consider the coarse-scale GCM outputs as boundary conditions and detailed physical processes (complex terrain and land surface characteristics) at regional scale to model the regional-scale climate. The RCMs have better ability in capturing the basic regional patterns in time and space. Contrary to this, statistical downscaling models use empirical relationships (Wilby et al., 1998, Wilby et al., 1999) between predictors (atmospheric processes) and predictands (local climatic variables, such as rainfall) to generate the data at finer spatiotemporal resolutions. Even though dynamical downscaling has shown robust performances, it requires large amount of data, high level of expertise to analyze the results, and is also computationally intensive, which make it beyond the reach of institutions in most countries (Fowler et al., 2007, Hay and Clark, 2003, Sachindra et al., 2014, Mahmood et al., 2016). The number of experiments for climate scenarios is limited by the intensive computational requirements, which is a function of resolution, domain size, and accuracy. Further, the RCMs are limited by systematic errors branching from the GCMs, as the RCMs are strongly dependent on the outputs from the GCMs.

Compared to the RCM, statistical downscaling methods have gained much wider acceptance, due to their simplicity and low computational burden (Okkan and Inan, 2015, Rashid et al., 2015, Sachindra et al., 2018, Sachindra et al., 2016, Sehgal et al., 2018). However, establishing robust linkage between local climatic variables and large-scale atmospheric processes requires long historical data (Hay and Clark, 2003). Depending on the assumed relationship, statistical downscaling consists of a heterogeneous group of methods that vary in sophistication and applicability. These methods are classified into three categories: stochastic weather generators, weather classification-based approaches, and regression-based methods. Stochastic weather generators involve fitting probability distribution function to the variable. The weather classification-based methods involve classification of the state variables into weather types using the k-means algorithm in addition to the stochastic weather generation. The weather generator methods are based on the statistical properties of the climatic variables. They are sensitive to outliers in the training set and involve the generation of an ensemble of time series and statistical analysis of the results. Furthermore, the weather generator and weather classification methods are data-intensive, need large computational resources, and do not have prediction capability outside the historical data range. These issues led the regression-based methods to gain popularity, especially due to their ease in application (Sachindra et al., 2018).

The regression-based downscaling methods primarily depend on developing a predictor-predictand relationship. Some of these methods include Multiple Linear Regression (MLR) (Sachindra et al., 2014, Joshi et al., 2015, Duhan and Pandey, 2015, Ravansalar et al., 2017), Generalized Linear Model (GLM) (Asong et al., 2016, Beecham et al., 2014), Artificial Neural Networks (ANNs) (Yeditha et al., 2020, Chithra et al., 2016, Hassan-Esfahani et al., 2015), Genetic Programming (GP) (Coulibaly and Burn, 2004, Sachindra et al., 2018), Random Forests Regression (RFR) (Hutengs and Vohland, 2016, Pang et al., 2017, Mishra et al., 2017), Support Vector Machines (SVMs) (Kannan and Ghosh, 2011, Srinivas et al., 2014), and Extreme Learning Machines (ELMs) (Sachindra et al., 2018, Zhu et al., 2019). These studies have shown that the regression-based downscaling methods are robust and accurate in capturing the local climatic patterns. Further, the results from the past studies indicate that techniques based on artificial intelligence (AI), which take into account the complexity and nonlinearity in the relationship between predictors and predictands, perform better in comparison with the other (linear) methods.

Many past studies (Kurths et al., 2019, Guntu et al., 2020, Maheswaran and Khosa, 2015, Rashid et al., 2016) have reported that the relationship between rainfall and atmospheric variables is a function of time-scale. For instance, Maheswaran and Khosa (2015), studying the Cauvery River Basin in India, showed that the relationship between climatic indices and rainfall varied within the time–frequency domain. Therefore, the popular traditional approach to downscaling based on single-scale models might not be able to capture the variability at different time scales (Rashid et al., 2016). In light of this limitation with single-scale models, there has been a special emphasis on the multi-scale dynamics of the climate system involving interactions and feedbacks among different processes at different temporal and spatial scales (Agarwal et al., 2017, Agarwal et al., 2018, Agarwal et al., 2019, Carey et al., 2013, Maheswaran and Khosa, 2012a, Maheswaran and Khosa, 2012b, Addison, 2005, Torrence and Compo, 1998). In this regard, wavelets have become a highly popular tool. With their inherent advantages, including time–frequency localization and multiscale resolution, wavelets decompose the given time series into scale-specific components as proxies of the physical processes at those scales. Several studies (Kaheil et al., 2008, Adamowski and Sun, 2010, Maheswaran and Khosa, 2012a, Maheswaran and Khosa, 2012b, Nourani et al., 2014, Nourani et al., 2018, Rathinasamy et al., 2013, Niu and Sivakumar, 2013, Beecham et al., 2014, Shafaei and Kisi, 2017, Nourani et al., 2009, Nourani and Farboudfam., N, , 2019, Baghanam et al., 2019, Rashid et al., 2018) have shown that wavelet transform-based modelling has aided in unravelling the multiscale dynamics in different processes. A recent study by Sun et al. (2019) investigated the efficiency of the wavelet-based models over the single (stochastic, ANN, regression) models for daily streamflow forecasting and reported better performance of the wavelet-based models. On similar lines, Ravansalar et al. (2017) used wavelet transform-based linear genetic programming (WLGP) models for predicting monthly streamflow in Iran. They found that the wavelet-based models significantly improved the flow prediction in comparison with GP, ANN, and MLR models. Djerbouai and Souag-Gamane (2016) showed that the accuracy of drought prediction in the Karkheh basin in Iran improved with wavelet pre-processing of inputs when compared to single models without wavelets.

Most of the above studies have combined the wavelets with other models in developing hybrid models for hydrologic forecasting applications. Even though there has been extensive research on the application of wavelets for hydrologic forecasting, there have only been very few studies on the application of wavelets for statistical downscaling. Among such studies, Kaheil et al. (2008) combined wavelets and SVM for downscaling and forecasting evapotranspiration. Rashid et al. (2015) used the wavelet GAMLSS model for downscaling of precipitation in the Onkaparinga River catchment in South Australia. The results showed that the WT-GAMLSS models were accurate in reproducing the time series properties when compared to the orthodox GAMLSS models. More recently, Lakhanpal et al. (2017) used a wavelet-based second-order Volterra model for downscaling monthly temperature in the Krishna River Basin in India and found that the wavelet-Volterra model performed better when compared to other models, including ANNs and MLR. Sehgal et al. (2018) used a similar model for downscaling monthly precipitation using the NCEP data and showed that the model results were closer to the local precipitation.

It is important to note that most of the above-mentioned studies have focussed on downscaling of precipitation (and other variables) at the monthly scale and there have hardly been any studies on evaluating the hybrid models for precipitation downscaling at the daily scale, despite the fact that downscaling of precipitation at the daily scale (and even sub-daily scales) is very important, especially in the context of extreme events. Furthermore, thus far, there has not been any serious attempt to conduct a detailed investigation to compare different methods for downscaling of daily precipitation.

In view of these, the present study proposes the application of wavelet-based hybrid methods for statistical downscaling of daily precipitation. The proposed wavelet-based hybrid models are applied for downscaling daily precipitation in five selected grid points (in different locations) in the Krishna River Basin in India, with varying climatic characteristics. The performances of these models are also compared with some popular models, including MLR, SDSM, GP, and single-scale ANN models.

Section snippets

Study area and data

The Krishna River Basin, being the fourth-largest basin in the Indian peninsula, originates at Mahabaleshwar in the state of Maharashtra, India and joins the Bay of Bengal at Hamasaladeevi (near Koduru) in Andhra Pradesh (Fig. 1). The catchment of the basin spreads between 73° to 82° E and 13° to 19° N, covering a total area of 260,401 km2. Fig. 2(a–d) shows the topography, mean annual rainfall distribution, Land Use Land Cover (LULC), and climate classification, respectively, of the Krishna

Methodology

In this study, several traditional and modern methods are applied for downscaling daily precipitation in the Krishna River Basin and for comparing their performances. These methods include Multiple Linear Regression (MLR), Statistical Downscaling Model (SDSM), Genetic Programming (GP), Artificial Neural Networks (ANNs), and Wavelet–Neural Network (WNN) hybrid model. The first four models are single-scale models, while the last one is a multiscale model. A brief description of these methods is

Model application

In any downscaling strategy, the selection of the predictors and its spatial domain plays a major role. Based on the climatic characteristics of the Krishna River Basin and reports by some previous studies (Lakhanpal et al., 2017, Sehgal et al., 2018), a total of 16 probable predictors is considered, as listed in Table 1. The list of probable predictors considered here is similar to that used by Sehgal et al. (2018) for the Krishna River Basin, Sachindra et al. (2018) for Australia, and Yang et

Discussion

In this study, wavelet-based hybrid models were proposed for downscaling daily precipitation and their performances were compared with those of some key traditional and other modern methods, including MLR, SDSM, GP, and ANN models. Among the nine downscaling methods applied in this study, machine-learning methods were found to generally outperform the basic MLR and SDSM. Among the single-scale machine-learning models, the NARX-NN model outperformed the FF-NN model and the GP model. The better

Conclusions

In this study, nine downscaling models, including MLR, SDSM, GP, ANN, and wavelet-based hybrid models, were developed and applied for downscaling daily precipitation at five selected grid points in different locations in the Krishna River Basin in India. Using the NCEP data and the IMD data over the period 1948–2017, calibration and validation of the models were carried out. Based on the results obtained for the five selected grids, in terms of several performance measures (NRMSE, NSE, R2, and

CRediT authorship contribution statement

Yeditha Pavan Kumar: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing - original draft. Rathinasamy Maheswaran: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Resources, Software, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing. Ankit Agarwal: Investigation, Methodology, Resources, Software, Validation, Visualization, Writing - original draft, Writing - review

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

RM gratefully acknowledges the funding received through the Inspire Faculty Award (IFA-12/ENG 28) from the Department of Science and Technology, India. AA acknowledges the funding support provided by the Indian Institute of Technology Roorkee through Faculty Initiation Grant number IITR/SRIC/1808/F.I.G and COPREPARE project funded by UGC and DAAD under the IGP 2020-2024. BS acknowledges the support from the IIT Bombay seed grant (RD/0519-IRCCSH0-027). The authors would like to thank the three

References (87)

  • R. Maheswaran et al.

    Wavelet Volterra Coupled Models for forecasting of nonlinear and non-stationary time series

    Neurocomputing

    (2015)
  • R. Maheswaran et al.

    Comparative study of different wavelets for hydrologic forecasting

    Comput. Geosci.

    (2012)
  • R. Maheswaran et al.

    Wavelet-Volterra coupled model for monthly stream flow forecasting

    J. Hydrol.

    (2012)
  • J.E. Nash et al.

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

    J. Hydrol.

    (1970)
  • J. Niu et al.

    Scale-dependent synthetic streamflow generation using a continuous wavelet transform

    J. Hydrol.

    (2013)
  • V. Nourani et al.

    A combined neural-wavelet model for prediction of Ligvanchai watershed precipitation

    Eng. Appl. Artif. Intell.

    (2009)
  • V. Nourani et al.

    Applications of hybrid wavelet-Artificial Intelligence models in hydrology: A review

    J. Hydrol.

    (2014)
  • V. Nourani et al.

    Data-driven ensemble model to statistically downscale rainfall using nonlinear predictor screening approach

    J. Hydrol.

    (2018)
  • V. Nourani et al.

    Rainfall time series disaggregation in mountainous regions using hybrid wavelet-artificial intelligence methods

    Envrionmental Research

    (2019)
  • M.M. Rashid et al.

    Statistical downscaling of CMIP5 outputs for projecting future changes in rainfall in the Onkaparinga catchment

    Sci. Total Environ.

    (2015)
  • M. Rathinasamy et al.

    Multiscale streamflow forecasting using a new Bayesian Model Average based ensemble multi-wavelet Volterra nonlinear method

    J. Hydrol.

    (2013)
  • M. Ravansalar et al.

    Wavelet-linear genetic programming: a new approach for modeling monthly streamflow

    J. Hydrol.

    (2017)
  • Weiwei Ren et al.

    A probabilistic method for streamflow projection and associated uncertainty analysis in a data sparse alpine region

    Glob. Planet. Change

    (2018)
  • D.A. Sachindra et al.

    Statistical downscaling of precipitation using machine learning techniques

    Atmos. Res.

    (2018)
  • V. Sehgal et al.

    Application of multi-scale wavelet entropy and multi-resolution Volterra models for climatic downscaling

    J. Hydrol.

    (2018)
  • Benny Selle et al.

    Testing the structure of a hydrological model using Genetic Programming

    J. Hydrol.

    (2011)
  • Daniel Svozil et al.

    Introduction to multi-layer feed-forward neural networks

    Chemom. Intell. Lab. Syst.

    (1997)
  • R.L Wilby et al.

    SDSM – a decision support tool for the assessment of regional climate change impacts

    Environ. Model. Softw.

    (2002)
  • R.L. Wilby et al.

    A comparison of downscaled and raw GCM output: Implications for climate change scenarios in the San Juan River Basin

    Colorado. J. Hydrol.

    (1999)
  • J. Adamowski et al.

    Comparison of multivariate regression and artificial neural networks for peak urban water-demand forecasting: Evaluation of different ANN learning algorithms

    J. Hydrol. Eng.

    (2010)
  • P.S. Addison

    Wavelet transforms and the ECG: A review

    Physiol. Meas.

    (2005)
  • A. Agarwal et al.

    Network-based identification and characterization of teleconnections on different scales

    Sci. Rep.

    (2019)
  • A. Agarwal et al.

    Wavelet-based multiscale similarity measure for complex networks

    Eur. Phys. J.-B

    (2018)
  • A. Agarwal et al.

    Multi-scale event synchronization analysis for unravelling climate processes: a wavelet-based approach

    Nonlinear Process. Geophys.

    (2017)
  • K. Ahmed et al.

    Multilayer perceptron neural network for downscaling rainfall in arid region: A case study of Baluchistan

    Pakistan. J. Earth Syst. Sci.

    (2015)
  • A. Anandhi et al.

    Downscaling precipitation to river basin in India for IPCC SRES scenarios using support vector machine

    Int. J. Climatol.

    (2008)
  • A.H. Baghanam et al.

    Conjunction of wavelet-entropy and SOM clustering for multi-GCM statistical downscaling

    Hydrol. Res.

    (2019)
  • S. Beecham et al.

    Statistical downscaling of multi-site daily rainfall in a South Australian catchment using a Generalized Linear Model

    Int. J. Climatol.

    (2014)
  • S.K. Carey et al.

    Use of color maps and wavelet coherence to discern short and longer-term climate influences on streamflow variability in northern catchments

    Water Resources Research.

    (2013)
  • Chen, S., Yu, P., Tang, Y., 2011. Support Vector Machines and Multivariate Analysis...
  • P. Coulibaly et al.

    Wavelet analysis of variability in annual Canadian streamflows

    Water Resour. Res.

    (2004)
  • S. Djerbouai et al.

    Drought Forecasting Using Neural Networks, Wavelet Neural Networks, and Stochastic Models: Case of the Algerois Basin in North Algeria

    Water Resour. Manag.

    (2016)
  • D. Duhan et al.

    Statistical downscaling of temperature using three techniques in the Tons River basin in Central India

    Theor. Appl. Climatol.

    (2015)
  • Cited by (15)

    • Heterogeneous spatiotemporal streamflow response to large-scale climate indexes in the Eastern Alps

      2022, Journal of Hydrology
      Citation Excerpt :

      To determine and investigate the crucial scales of variability we apply wavelet transform (WT) techniques. Wavelet analysis has shown to be a very promising tool to analyze non-stationary hydro-climatic time series (Marcolini et al., 2017; Sang et al., 2018; Yeditha et al., 2021; Zolezzi et al., 2009). In the past decades, it has found application in a wide range of climatological studies based on oscillation indexes (Das et al., 2020; Fu et al., 2012; He et al., 2021; Jevrejeva et al., 2003; Massei et al., 2010; Ranzi et al., 2021) and has been applied to investigate streamflow variability in selected rivers worldwide (Agarwal et al., 2016; Labat, 2008; Labat, 2010; Labat et al., 2004; Massei et al., 2011; Nalley et al., 2012; Nalley et al., 2016; Nalley et al., 2019; Pérez Ciria et al., 2019; Rossi et al., 2009).

    • A novel statistical downscaling approach for analyzing daily precipitation and extremes under the impact of climate change: Application to an arid region

      2022, Journal of Hydrology
      Citation Excerpt :

      Pour et al. (2016) combined random forests (RF) and SVM for daily rainfall projection, results showed that the hybrid model could downscale rainfall with Nash-Sutcliff efficiency in the range of 0.90 ∼ 0.93. In the previous studies, the performance of different statistical downscaling methods for ordinary precipitation was compared; nevertheless, few works were reported on assessing the statistical downscaling methods for extreme events of daily precipitation (Kumar et al., 2021; Yang et al., 2018a). Moreover, since not all small-scale variability can be explained by the large-scale predictors, predictions of precipitation (especially for extreme precipitations) generally have less variance than the observed local variables (Maraun, 2013; Pryor and Schoof, 2019).

    • Climate indices and hydrological extremes: Deciphering the best fit model

      2022, Environmental Research
      Citation Excerpt :

      Runoff is the response of a catchment to its rainfall input. Recently, there has been a significant development in satellite products that provide gridded precipitation data for the entire globe (Kumar et al., 2021), but this is not true for runoff, and its data is not as accurate and readily available. For India, Central Water Commission maintains gauges sites to observe daily runoff.

    • Downscaling of the flood discharge in a probabilistic framework

      2022, Journal of Hydro-Environment Research
      Citation Excerpt :

      Results show that the developed model can downscale flood discharge in all parts of the hydrograph (rising, falling and crest) well. Although many statistical downscaling methods fail to capture peak and extreme values (Kumar et al., 2021; Vandal et al., 2019; Gusain et al., 2019; Shashikanth et al., 2018), the developed method can properly downscale extreme discharge (peak). Hourly downscaled data follow hourly observations in validation period with the same accuracy as the calibration one.

    View all citing articles on Scopus
    View full text