Drought prediction and modeling by hybrid wavelet method and neural network algorithms

Subject Areas : Natural resources and environmental management

Jahanbakhsh  Mohammadi ¹ (PhD Student, Department of Remote Sensing and GIS, Faculty of Natural Resources and Environment, Science and Research Branch, Islamic Azad University, Tehran, Iran)
Alireza  Vafaeinezhad ² (Associate Professor, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran)
Saeed  Behzadi ³ (Assistant Professor, Faculty of Civil Engineering Shahid Rajaee Teacher Training University, Tehran, Iran)
Hossein  Aghamohammadi ⁴ (Assistant Professor, Department of Remote Sensing and GIS, Faculty of Natural Resources and Environment, Science and Research Branch, Islamic Azad University, Tehran, Iran)
Amirhooman  Hemmasi ⁵ (Professor, Department of Natural Resources Engineering, Faculty of Natural Resources and Environment, Tehran science and Research Branch, Islamic Azad University, Tehran, Iran)

Received: 2021-11-17 Accepted : 2022-01-03 Published : 2022-10-23

Keywords: Standardized Precipitation Index (SPI), Generalized Regression Neural Network (GRNN), Radial base function (RBF), Multilayer Perceptron neural network (MLP), drought, Neural network,

Abstract :

Background and Objective A drought crisis is a dry period of climate that can occur anywhere globally and with any climate. Although this crisis starts slowly, it can have a serious impact on health, agricultural products, the economy, energy, and the environment for a long time to come. Drought severely threatens human livelihood and health and increases the risk of various diseases. Therefore, modeling and predicting drought is one of the most important and serious issues in the scientific community. In the past, mathematical and statistical models such as simple regression, Auto-regression (AR), moving average (MA), and ARIMA were used to model the drought. In recent years, machine learning methods and computational intelligence to model and predict drought have been of great interest to scientists. Computational intelligence algorithms that have been previously considered by scientists to model drought include multilayer perceptron neural network, RBF neural network, support vector machine, fuzzy, and ANFIS methods. In this research, the purpose of modeling and predicting drought is by using three neural network algorithms, including multilayer perceptron, RBF neural network, and generalized regression neural. The drought index used in this research is the standardized precipitation index (SPI). In this research, the wavelet technique in combination with artificial neural network algorithms for modeling and predicting drought in 10 synoptic stations in Iran (Abadan, Babolsar, Bandar Abbas, Kerman, Mashhad, Rasht, Saqez, Tehran, Tabriz, and Zahedan) have been used in different climates and with suitable spatial distribution throughout Iran.Materials and Methods This study, initially using monthly precipitation data between 1961 and 2017, SPI drought index in time scales of 3, 6, 12, 18, 24, and 48 months through programming in soft environment MATLAB software implemented. The results of this step were validated using the available scientific software MDM and Drinc. Then, prediction models were designed using the Markov chain. In this study, a total of six computational intelligence models, including three single models of multilayer perceptron neural network (MLP), radial basis function neural network (RBF), and generalized regression neural network (GRNN), and three hybrids wavelet models with these three models (WMLP-WRBF-WGRNN) have been used to model and predict the SPI index in 10 stations of this research. In implementing all these six models, the MATLAB software programming environment has been used. In this study, four types of discrete wavelets were used, including Daubechies, Symlets, Coiflets, and Biorthogonal. Due to the better performance of the Dobbies wavelet, this type of wavelet was used as a final option in the research. In the Daubechies wavelet used between levels 1 to 45, level 3 showed the best performance among different SPI time scales; therefore, the Daubechies level 3 wavelet was used in all hybrid models of this study. After training all six algorithms used, the evaluation criteria of coefficient of determination (R2) and root mean square error (RMSE) was used to measure the difference between actual and estimated values.Results and Discussion The results of this study showed that computational intelligence methods have high accuracy in modeling and predicting the SPI drought index. In the first stage, the results showed that the individual MLP, RBF, and GRNN models, if properly trained, have close results in modeling and predicting the SPI drought index. In the next step, it was observed that the wavelet technique would improve the modeling results. In using the wavelet technique in combination with three single models MLP, RBF, and GRNN, the choice of wavelet type is also more effective in modeling, so in this research, the first of the four types of discrete wavelets Daubechies, Symlet, Qoiflet, and Biorthogonal in combination with Three single models of this research were used and the results of these four types of wavelets showed the relative superiority of the Daubechies wavelet over the other three wavelets. In using the Daubechies wavelet, since this wavelet has 45 times and the choice of order was also effective in modeling, it was observed by testing the wavelet 45 times that the 3rd wavelet, in general, has higher accuracy in all time scales of SPI index, 3, 6, 12, 18, 24 and 48 months and also in all three algorithms MLP, RBF, and GRNN. Therefore, in this research, the third-order Daubechies wavelet was used in all three algorithms of this research, as well as in all time scales. The results showed that combining the wavelet technique with all three models MLP, RBF, and GRNN will improve the results. The research graphs showed that for the quarterly time scale, the values obtained from the single model prediction in MLP and RBF modeling have a somewhat one-month phase difference compared to the hybrid model, while in the GRNN model, this prediction difference is negligible. The modeling results for both single and hybrid modeling modes indicate that there is no phase difference between the single and hybrid modeling methods in time scales of 6, 12, 18, 24, and 48. For the 12- and 24-month time scales, the single GRNN model had more fluctuations and errors in SPI monthly modeling and forecasting, while the hybrid model in these two-time scales had much better behavior in monthly modeling and forecasting. Distribution diagrams of data related to observational SPI of Abadan station showed that the modeling results for single and hybrid modes in 3 and 6-month time scales are less accurate than other time scales and fit line separation, and its uncertainty is higher than others. However, in all neural network models and in all time scales, the hybrid method has shown more accuracy. The numerical results of the study indicate that in all SPIs and stations under study, the differential values of R2 are positive, which indicates higher values of R2 in the hybrid model than in single neural network modeling, which indicates an improvement in hybrid modeling compared to individual models. Also, the differential values of RMSE are negative in all studied models and stations, which indicates that the amount of RMSE in predicting hybrid models is lower than individual neural network models. In the research graphs, it can be seen that the amount of differences in RMSE and R2 indicates a greater difference in time scales 3 and 6 than the time scales 12, 18, 24, and 48, which somehow goes back to the nature of the data of these time scales. The most significant improvement in R2 and RMSE is from the 3-month low to the 48-month high, respectively.Conclusion From the findings of this study, it can be concluded that artificial neural network algorithms are efficient methods for modeling and predicting the SPI drought index. The use of wavelets in all three models of artificial neural networks will also improve the results. It can also be concluded that for better modeling of the SPI drought index, it is necessary to select the optimal wavelet type and order. From the results of this study, it can be concluded that the wavelet technique has a greater impact on the lower time scales, i.e., 3 and 6 months, than the higher scales, i.e., 24 and 48 months.

References:

Abeysingha N, Rajapaksha U. 2020. SPI-based spatiotemporal drought over Sri Lanka. Advances in Meteorology, 2020. doi:https://doi.org/10.1155/2020/9753279.

Azimi S, Moghaddam MA. 2020. Modeling short term rainfall forecast using neural networks, and Gaussian process classification based on the SPI drought index. Water Resources Management: 1-37. doi: https://doi.org/10.1007/s11269-020-02507-6.

Bhunia P, Das P, Maiti R. 2020. Meteorological drought study through SPI in three drought prone districts of West Bengal, India. Earth Systems and Environment, 4(1): 43-55. doi:https://doi.org/10.1007/s41748-019-00137-6.

Diop L, Bodian A, Djaman K, Yaseen ZM, Deo RC, El-Shafie A, Brown LC. 2018. The influence of climatic inputs on stream-flow pattern forecasting: case study of Upper Senegal River. Environmental earth sciences, 77(5): 1-13. doi: https://doi.org/10.1007/s12665-018-7376-8.

Foroumandi E, Nourani V, Sharghi E. 2021. Climate change or regional human impacts? Remote sensing tools, artificial neural networks, and wavelet approaches aim to solve the problem. Hydrology Research, 52(1): 176-195. doi:https://doi.org/10.2166/nh.2020.112.

Ghritlahre HK, Prasad RK. 2018. Exergetic performance prediction of solar air heater using MLP, GRNN and RBF models of artificial neural network technique. Journal of environmental management, 223: 566-575. doi: https://doi.org/10.1016/j.jenvman.2018.06.033.

Hadi SJ, Tombul M. 2018. Streamflow forecasting using four wavelet transformation combinations approaches with data-driven models: a comparative study. Water Resources Management, 32(14): 4661-4679. doi: https://doi.org/10.1007/s11269-018-2077-3

Hannan SA, Manza R, Ramteke R. 2010. Generalized regression neural network and radial basis function for heart disease diagnosis. International Journal of Computer Applications, 7(13): 7-13. doi: https://doi.org/10.5120/1325-1799.

Hosseini-Moghari SM, Araghinejad S. 2015. Monthly and seasonal drought forecasting using statistical neural networks. Environmental Earth Sciences, 74(1): 397-412. doi:https://doi.org/10.1007/s12665-015-4047-x.

 Khan MMH, Muhammad NS, El-Shafie A. 2020. Wavelet based hybrid ANN-ARIMA models for meteorological drought forecasting. Journal of Hydrology, 590: 125380. doi: https://doi.org/10.1016/j.jhydrol.2020.125380.

KISI Ö. 2006. Generalized regression neural networks for evapotranspiration modelling. Hydrological Sciences Journal, 51(6): 1092-1105. doi:https://doi.org/10.1623/hysj.51.6.1092.

Kisi O, Tombul M, Kermani MZ. 2015. Modeling soil temperatures at different depths by using three different neural computing techniques. Theoretical and applied climatology, 121(1): 377-387. doi: https://doi.org/10.1007/s00704-014-1232-x.

Komasi M, Sharghi S. 2020. Drought Forecasting Using Wavelet-Support Vector Machine and Standardized Precipitation Index (Case Study: Urmia Lake-Iran). Journal of Environmental Science and Technology, 22(7): 83-101. doi:https://doi.org/10.22034/jest.2020.9578. (In Persian).

Lazri M, Ameur S, Brucker JM, Lahdir M, Sehad M. 2015. Analysis of drought areas in northern Algeria using Markov chains. Journal of Earth System Science, 124(1): 61-70. doi: https://doi.org/10.1007/s12040-014-0500-6.

Li L, She D, Zheng H, Lin P, Yang Z-L. 2020. Elucidating diverse drought characteristics from two meteorological drought indices (SPI and SPEI) in China. Journal of Hydrometeorology, 21(7): 1513-1530. doi:https://doi.org/10.1175/JHM-D-19-0290.1.

Lin G-F, Chen L-H. 2004. A non-linear rainfall-runoff model using radial basis function network. Journal of Hydrology, 289(1-4): 1-8. doi: https://doi.org/10.1016/j.jhydrol.2003.10.015.

Lippmann R. 1994. Book Review:" Neural Networks, A Comprehensive Foundation", by Simon Haykin. International Journal of Neural Systems, 5(04): 363-364. doi: https://doi.org/10.1142/S0129065794000372.

Mahmoudzadeh H, Azizmoradi M. 2019. Deforestation modeling using artificial neural network and GIS (Case study: forests of Khorramabad environs). Journal of RS and GIS for Natural Resources, 10(4): 74-90. http://girs.iaubushehr.ac.ir/article_670420.html. (In Persian).

McKee TB, Doesken NJ, Kleist J. 1993. The relationship of drought frequency and duration to time scales. In: Proceedings of the 8th Conference on Applied Climatology, vol 22. Boston, pp 179-183.

Mehdizadeh S, Ahmadi F, Mehr AD, Safari MJS. 2020. Drought modeling using classic time series and hybrid wavelet-gene expression programming models. Journal of Hydrology, 587: 125017. doi: https://doi.org/10.1016/j.jhydrol.2020.125017.

Mirdashtvan M, Saravi MM. 2020. Influence of non-stationarity and auto-correlation of climatic records on spatio-temporal trend and seasonality analysis in a region with prevailing arid and semi-arid climate, Iran. Journal of Arid Land, 12(6): 964-983. doi: https://doi.org/10.1007/s40333-020-0100-z.

Ozan Evkaya O, Sevinç Kurnaz F. 2021. Forecasting drought using neural network approaches with transformed time series data. Journal of Applied Statistics, 48(13-15): 2591-2606. Doi: https://doi.org/10.1080/02664763.2020.1867829.

Paulo AA, Pereira LS. 2007. Prediction of SPI drought class transitions using Markov chains. Water resources management, 21(10): 1813-1827. Doi: https://doi.org/10.1007/s11269-006-9129-9.

Pei Z, Fang S, Wang L, Yang W. 2020. Comparative analysis of drought indicated by the SPI and SPEI at various timescales in Inner Mongolia, China. Water, 12(7): 1925. doi: https://doi.org/10.3390/w12071925.

Raziei T. 2017. Köppen-Geiger climate classification of Iran and investigation of its changes during 20th century. Journal of the Earth and Space Physics, 43(2): 419-439. doi: https://doi.org/10.22059/jesphys.2017.58916. (In Persian).

Rhif M, Ben Abbes A, Martinez B, Farah IR. 2021. An improved trend vegetation analysis for non-stationary NDVI time series based on wavelet transform. Environmental Science and Pollution Research, 28(34): 46603-46613. doi: https://doi.org/10.1007/s11356-020-10867-0.

Taylan ED, Terzi Ö, Baykal T. 2021. Hybrid wavelet–artificial intelligence models in meteorological drought estimation. Journal of Earth System Science, 130(1): 1-13. doi:https://doi.org/10.1007/s12040-020-01488-9.

Won J, Choi J, Lee O, Kim S. 2020. Copula-based Joint Drought Index using SPI and EDDI and its application to climate change. Science of the Total Environment, 744: 140701. Doi: https://doi.org/10.1016/j.scitotenv.2020.140701.

Xu D, Zhang Q, Ding Y, Huang H. 2020. Application of a Hybrid ARIMA–SVR Model Based on the SPI for the Forecast of Drought—A Case Study in Henan Province, China. Journal of Applied Meteorology and Climatology, 59(7): 1239-1259. doi: https://doi.org/10.1175/JAMC-D-19-0270.1.

Zadeh MR, Amin S, Khalili D, Singh VP. 2010. Daily outflow prediction by multilayer perceptron with logistic sigmoid and tangent sigmoid activation functions. Water resources management, 24(11): 2673-2688. doi: https://doi.org/10.1007/s11269-009-9573-4.