Ijraset Journal For Research in Applied Science and Engineering Technology
Authors: Angel R, Kokila Ramesh, Anita Chaturvedi
DOI Link: https://doi.org/10.22214/ijraset.2024.66196
Certificate: View Certificate
Drought has resulted in the massive destruction of lives, livestock and economy, locally and globally.Underthecurrentcircumstances,thedryweatherconditionsareprevalentleadingto water stress resulting in imbalance of the ecosystem. The persistent dry weather eventually resultsindroughtconditions,whichindirectlya?ectstheeconomyofthecountry.Asaresult, it has become imperative to assess and mitigate the e?ects of the droughts. With the advancements in the machine learning techniques, there has been a substantial improvement in the level of accuracy of the prediction of drought events. The motivation of this paper is to review those machine learning techniques such as Extreme Gradient Boost, ARIMA, Long Short-Term Memory (LSTM), Random Forest used in forecasting drought events.Consideringtheextensivelyuseddroughtindices,thatis,theStandardPrecipitation Index(SPI) and Standardized Precipitation Evaporation Index (SPEI) together with the various meteorological parameters, the influence of the climate variables on drought forecasts are given more importance in the present study. This study also includes review on single and hybrid models used for forecasting droughts. Further the gaps in the reviews have been identifiedbycriticallyobservingthestudiesintheresearcharticleswhichwillenhancethe e?ciency of the predictive models.
Droughtsarethemostexpensivenaturalevent.Withthecontinuouschangeintheclimate,the global average surface temperature is on the rise. “There is a 66% chance that the annual average temperature of the near surface global temperature is 1.50 C above the pre-industrial level for at least one year between 2023 and 2027” [1]. This temperature increase leads to dryness of the weather, consequently increases the evapotranspiration and this results in the water stress. The incidence of drought events have increased in the past decades.The World Health Organization reports that around 55 million people have been a?ected by drought globally each year [2]. It a?ects the livelihood and livestock. Water shortages have impacted about 40% of the global population. There is a possibility that by 2030, 700 million people will be displaced due to drought [2].
AccordingtoWhilhiteandGlantz[3],“therecannotbeauniversaldefinitionofdrought”[3]. The absence of a single definition for drought has handicapped the scientists to identify the
onsetofdrought.Therearevariousdefinitionsacceptedbythedi?erentinstitutions.National Integrated Drought Information Systems (NIDIS), has defined drought as “A deficiency of precipitation over an extended period of time, resulting in a water shortage” [4]. The consequence of droughts, causes an imbalance in the demand and the supply of water, a?ectingwaterneedsoftheecosystem[25].Hencequantificationofdroughtbecomespivotal to manage droughts.
MishraandSingh[5]haveclassifieddi?erenttypesofDroughts.
Meteorological drought- It occurs due to the e?ect of the mereological variables like precipitation,rainfall,temperature,evapotranspiration,cloudcoveretc.Itoccurswhenthe dry weather pattern exists.
Agriculturaldrought–Thisoccurswhenthereisinsu?cientsupplyofwatertothecrops,due to lack of water in the surface and the subsurface or ground level[5]. This results in the reduction in soil moisture, where the water supply does not meet the demand of the water requirement of the crops. The soil moisture during this period is very less or nil.
Hydrological drought – This is the result of decrease of water levels in the natural water resourcessuchasrivers,ponds,lakesetc.andintheotherstoragetanksandtheirrigation canals, dams, reservoirs, groundwater, etc. [5].
Socio -economic Drought-The non-availability of waterforan extended period of time will lead to malnutrition in the population, lack of fodder for cattle which results in the death of thesame,massmigration,andothereconomicactivitiessuchastourism,hospitalityindustry will be a?ected.
“The drought indicators are the variables which describe the drought conditions. The commonindicatorsassumedareprecipitation,temperature,soilmoisture,groundand reservoir water, streamflow and snowpack” [6].
DroughtIndiceshelptoconverttheindicatordatawithrespecttoacommonscale,whichcan be used to compare the droughts across di?erent geographical locations and across di?erent time scales.There are more than 150 drought indices listed based on the drought indicators in the hand book of indicators and indices[6]. One of the earliest indices that was developed is the - “Palmer Drought Severity Index (PDSI) [7], China Z Index (CZI) [8], Rainfall Anomaly Index (RAI) [9], Rainfall deciles [10], Crop Moisture Index (CMI) [11], Aridity Anomaly Index (AAI) [12], Standardised Precipitation Index (SPI) [13], Standardised precipitation Evapotranspiration Index (SPEI)” and many more. Among all the indices, the World Meteorological Organisation (WMO) has recommended SPI. The SPI index was developed by Mackee et.al.[15], uses only precipitation as the input variable. It is a simple index compared to the other complex drought indices. The droughts can be better monitored and compared using SPEI rather than SPI, as temperature is a fundamental variablein drought occurrence together with precipitation [16].
It is observed in the existing literature that di?erent modelling techniques are available to forecastthedroughts.Themodelscanbedescribedasdynamicalordatadriveninnature.The dynamical models are based on physical laws involving hydrological cycle- land, ocean and atmosphere cycle. If the initial and the boundary conditions are accurately considered, then the prediction is bound to be accurate. Since the natural cycles are not yet completely understood, converting them into mathematical equations or di?erential equations is still a challenge. While the data driven models are constructed on the past data and its interaction with the other atmospheric variables. These interactions are based on the empirical relationshipsbetweenthevariablesandarederivedbasedonthemeasuresoftherelation.Itis less time consuming and the ease to handle the complex data has made it more popular in predicting the weather data. [17] classified the drought modelling techniques and are as follows. Stochastic models -It is the time series model which includes Auto Regressive Integrated Moving Average (ARIMA) modelsand Seasonal Auto Regressive Integrated Moving Average (SARIMA) models. Probabilistic modelling that includes the Copula and Markov models, The Artificial Neural Network (ANN) models have performed extremely well when compared with the other models. Hybrid models- which is the combination of di?erent models- such as copulas with ANN, ARIMA with ANN. so that it can be tailored according to the needs.So that the e?ciency of each of the models can be used.This contributes to developing a model with a good performance.
Inastudyof[18],thetrainingdatausedwerefrom1901to2010,thetestingdatafrom2011 to2018wereobtainedfromtheClimateResearchUnitforNewSouthWales,Australia.The indicator variables used were precipitation, mean temperature, vapour pressure, cloud cover and potential evapotranspiration [18]. The droughts were predicted using Long Short-Term Memory (LSTM) at two di?erent time scales SPEI1, SPEI3. Receiver Operating Characteristic-based Area Under the Curve (ROC-AUC) approach was used to identify the droughtclasses.Here,thehydrometeorologicalvariablesalonewereconsidered,anddidnot include the climate variables. The deep learning method was able to e?ectively predict theresultsevenwithouttheclimatevariables.Theindicatorssuchasvapourpressureandcloud cover are observed to beinfluencing factors of drought, like the temperature.
Thedroughtforecasts[19]usingSPEIandSPIforvaryingtimescaleswasanalysed.Thedata sets from 1959 to 2013 were used to train the data using LSTM and the year 2014 was to test the data under the used modelling technique. The input variables considered were the precipitation,minimum temperature and maximum temperature, windspeed, sunshine, humidity and evapotranspiration [19]. The correlation analysis was performed to identify the indicatorsa?ectingdroughtsindexSPIandSPEI.Thecorrelationoftheindicators-humidity and temperature were considered with the indices. While the correlation existed between humidity and SPEI and not with SPI. For the sake of uniformity, the indicators humidity and temperature were considered.A combination of 4 models were developed- SPI with temperature, SPI with humidity, SPEI with temperature and SPEI with humidity for varying timescales-1,6,12, using ARIMA.
The performance of the ARIMA model for 1month-timescaleusingSPIandSPEIwashigher,whilefor6-and12-monthtimescaleswere not very convincing. The accuracy of LSTM models for 1-month timescale is 99% for SPI AND SPEI. LSTM outperformed ARIMA model in the case of 6, 12-month timescale. The multivariate approach using LSTM has shown better results compared to the univariate method of ARIMA.Inanotherstudyof[20],comparisonbetweenmachinelearningalgorithms-XGB-Extreme Gradient Boost and RF- Random Forest andthe Deep Learning models namely, LSTM - LongShort-TermMemoryandCNN-ConvolutionalNeuralNetwork.Sevenscenarioswith di?erent choices of the input variables- “precipitation(P), Temperature average(T), temperature minimum, Temperature maximum, wind(W), relative humidity (RH), sunshinehours (SH) and solar radiation (SR) were analyzed for two di?erent timescales SPEI 3 and SPEI-6 [20]. The study area was Tibetan plateau, China. The data from 1980 to 2019 were extracted.Theinputdatawasnormalisedforbetterunderstandingofthestructure.Agnew’s approachwasusedtodeterminethedroughtclasses.Thevariouscharacteristicsofdrought, such as drought area, drought severity and drought duration were quantified. Principal component Analysis was used to locate drought patterns. The performance of machine learning models- XGB and RF were considered to be better compared to the deep learning models CNN and LSTM. It has been observed that there existed a significant correlation between wind speed and relative humidity.
Thecasestudy[21]wasconductedinJiangsuprovinceinChina,withthree di?erentmethods-ARIMA,PROPHET-(isaPythonbasedopen-sourcesoftware-e?ectiveinthecase of missing data, with outliers and reckoning patterns) and LSTM drought prediction. Remote sensing data such as NDVI, LST- Land Surface Temperature, Climate data such as SPEI, Evapotranspiration-ETP, humidity, precipitation, wind speed and pressure and Biophysical data, soil moisture were considered [21]. Droughts were forecasted using SPEI values for the three models. Maps were created. For LSTM, the e?ciency using Root Mean Square Error (RMSE) metric was 0.001 and concluded that LSTM was the best model as compared to ARIMA and PROPHET using the metrics, R2, RMSE, Mean Absolute Error (MAE).
Duetothecomplexnatureoftheweatherdata,thetraditionalmethods[22]offorecastingthe drought, is more suitable for the linear trend, and fails to cope with the non-linear trend existing in the data. With the development of the Artificial neural network, the drawbacks of the traditional methods have been overcome to a certain extent. This paper utilizes 6 models forecasting the drought using datasets from 613 stations from 1980 to 2019 with themultivariate drought index, SPEI. ARIMA was used to predict SPEI 24 and was more accuratecomparedtoSPEI1.Theprecipitationandtemperaturedatasetsfromthestations were used. The deep learning model considered was LSTM and Support Vector Machine (SVR)forvarioustimescales.SupportvectorRegression(SVR)modelwasobservedtobe better suited for the prediction of the long-term drought prediction. Hybrid models with di?erent combinations of standalone models such as ARIMA-LSTM, ARIMA-SVR andLS-SVR were developed [22]. The prediction using the ARIMA-LSTM model was far more e?cient compared to the individual models and other hybrid models. In the ARIMA-LSTM model, the linear patterns were analysed by the ARIMA model. The residue between predicted and actual observations are then fed into the LSTM model. The nonlinear features wereextractedandforecasted.Finally,thepredictedlinearfeatureandthenon-linearfeatures wereintegratedtogetthecompletepredictionofthehybridmodels.Theaimofthispaperis to improve the accuracy of short-term prediction. Of the considered models, theARIMA-LSTM,performsbetterwithlongleadtime.
Lagged climate variables and stacked LSTM [23] were used to forecast the long lead time drought The hydro-meteorological variables - precipitation, potential evapotranspiration, temperature,cloudcoveralongwith“seasurfacetemperature(SST)indicessuchasPacific Decadaloscillation,SouthernoscillationIndices,IndianoceanDipoleIndices,NinoIndices and southern Annular mode” were used as the input variable, along with SPEI [23]. The lag periods were determined for SPEI and the variables, using Cross correlation with various climate indices.The optimum lag period obtained was as large as 8 months for Pacific Decadal Oscillation and zero for Indian Dipole Oscillation. The lagged variables used for prediction considerably, improves the forecasting for long lead time.It was observed that cloudcoverandprecipitationresultedinhighcorrelation.Deeplearningalgorithm-Stacked LSTM,wasusedtoforecast.TheLSTMmodelsperformbetterthanthetraditional models.
The meteorological drought was predicted using the Wavelet - ARIMA- LSTM model using China Z-score Index (CZI)[24]. The Grey prediction model is used for the drought characterization of the forecasted values. The precipitation data of 51 years is used for training and testing of data. This data is first normalized. This method involved the decompositionandreconstructionofdata.First,thedataisclassifiedintohighfrequency,low frequency components using wavelet decomposition [24]. ARIMA estimates the low frequency coe?cients, while LSTM estimates the high frequency coe?cients which arenon-linear. The predicted coe?cients are then used to reconstruct the data. Comparison was made between ARIMA, LSTM, Wavelet-ARIM-LSTM. Then the forecasted values of ARIMAandLSTMareutilizedtoreconstructthedata.TheCZIwasusedtocharacterizethe drought. Finally, the Discrete Grey model was used to categorise the drought classes. Grey models work well when the sample size is small. The model predicted better in the humid regions.
Droughts are one of the least understood natural calamities of nature resulting in devastating e?ects on the ecosystem. The frequency of droughts in the current times is alarming. Hence there is an urgent need for a reliable forecasting model. The e?ciency of the forecasting models with long lead time is the pressing priority. The fundamental requirements of a forecastingmodelarehistoricaldata,asuitableindexandaforecastingmodel.Thehistorical data can be of gridded data or the remote sensing data of the required parameters. Secondly there are several drought indices available and the recommended meteorological drought indices by WMO are SPI - which considers only precipitation as the parameter and SPEI - considers temperature and precipitation as the parameter. The performance of SPI and SPEI have been significantly higher, as compared to the other drought indices. Nevertheless, the prediction of droughts can be further enhanced by the use of climate indices such as Indian Ocean Dipole, Southern Oscillation Index, ENSO, etc. The meteorological index along with theclimateindiceshaveresultedinabetteraccuracy.Anotherelementinthepredictionofthe drought, is the type of models used. Among the data driven models, the hybrid models such asARIMA-LSTM,ARIMA-SVR,Wavelet-ARIMA-LSTMetc,havehadhigherperformance accuracy than the single models such as LSTM, ARIMA, XGB, Random Forest etc.Thus, the essential components of the modelling should be chosen carefully for the accurate prediction of droughts.
The RMSE for SPI -1 with humidity and that of SPEI-1 with humidity is 0.011 and 0.2 respectively, which gives a good accuracy for the region of Hyderabad [19]. It may be observed that in both the cases the RMSE is low and it may not be the only variable included for predicting the drought events, but this may be one of the steps.This accuracy was achievedusingthedeeplearningLSTMwhichperformedbetterthantheARIMAmodel.[18] used SPEI with LSTM to predict the droughts for New South Wales of Australia and RMSE was found to be 0.018 for SPEI-1 and 0.027 for SPEI-3. Here the only indices used areSPEI-1 and SPEI-3 and are not connected to other atmospheric indices such as Southern Oscillationindex(SOI),IndianDipoleIndex,snowcoverandsoon.SinceLSTMisusedto predict drought events, it may perform during the training period, but theactual challenge might arise during forecasting. The machine learning models, namely Extreme Gradient Boost,RandomForest,ConvolutionneuralNetworkandtheLSTMmodelswereusedinthe prediction in Tibet plateau [20]. Nash-Sutcli?e model e?ciency for SPEI -3 was more than0.75 for Extreme Gradient Boost models and RF models. However, the inclusion of atmospheric variables might improve the e?ciency. [23] has included the atmospheric parameters by considering Southern Oscillation Index, Indian Ocean Dipole, along with the meteorological parameters considering the stacked LSTM model. It was able to consider the nonlinearrelationshipbetweenthevariablesalongwithENSO,buttheforecastinge?ciency needs to be estimated as per the model, which is not considered in this article. [24] used the hybrid model of WAVELET-ARIMA-LSTM over north east China with the CZI index. The correlation coe?cient was found to be as close as 0.880 between the model and the precipitation estimates. To e?ectively model, predict and forecast the extreme events in any country, one has to concentrate on the e?ciency of the model through the training period. If in the training period e?ciency crosses 95%, then there is a possibility of predicting and forecasting with better accuracy and can be relied upon.With the availability of various machine learning models, and considering all possible correlated variables of the drought, drought forecast can be improvised. This will provide the researchers to proceed in a correct path to come out with better accuracy models to forecast drought events with better confidence levels.
Drought has a massive impact on the ecosystem and has inevitably compromised human existence. Considering the climate changes, the frequency of occurrence of droughts is inclinedtoincreaseinthefuture.Thispaperreviewsthevariousdroughtindices,atmospheric variables and the models with di?erent methodology used to predict droughts. It highlights the di?erent drought indices used in the models.The e?ectiveness of various parameters such as the meteorological parameters and inclusion of the climatic parameters in theprediction are examined through the articles reviewed. The comparison of di?erent models used such as the individual models and the hybrid models are included. This review provides su?cientinformationtotheresearchersforusingapropermethodologytoovercomethegaps in the existing models.
[1] WorldMeteorologicalOrganisation.,(2023,March17),GlobalTemperaturessetto reach new records in next five years. https://wmo.int/news/media-centre/global-temperatures-set-reach-new-records-next-five-years. [2] World Health Organisation, (n.d.). Drought,https://www.who.int/health-topics/drought/#tab=tab_1 [3] Wilhite, Donald A. and Glantz, Michael H., \"Understanding the Drought Phenomenon:TheRoleofDefinitions\"(1985).DroughtMitigationCenterFaculty Publications. 20. http://digitalcommons.unl.edu/droughtfacpub/20 [4] NationalintegratedDroughtInformationSystem,(n.d.),DroughtBasics, https://www.drought.gov [5] Mishra,A.K.,&Singh,V.P.(2011).Droughtmodeling–Areview,Journalof Hydrology. [6] Handbook of Drought Indicators and Indices (n.d.), pg. 3,https://www.droughtmanagement.info/literature/GWP_Handbook_of_Drought_Indicators_and_Indices_2016.pdf. [7] Palmer,W.C.,1965:MeteorologicalDrought.ResearchPaperNo.45,USWeather Bureau, Washington, DC. [8] Wu,H.,M.J.Hayes,A.WeissandQ.Hu,2001:AnevaluationoftheStandardized Precipitation Index, the China-Z Index and the Statistical Z-score. International Journal of Climatology, 21: 745–758. [9] vanRooy,M.P.,1965:ARainfallAnomalyIndexindependentoftimeandspace. Notos, 14: 43–48. [10] Gibbs,W.J.andJ.V.Maher,1967:RainfallDecilesasDroughtIndicators.Bureauof Meteorology Bulletin No. 48, Melbourne, Australia. [11] Palmer,W.C.,1968:Keepingtrackofcropmoistureconditions,nationwide:thenew Crop Moisture Index. Weatherwise, 21: 156–161. DOI:10.1080/00431672.1968.9932814 [12] WorldMeteorologicalOrganization,2012:Chapter6,AgrometeorologicalForecasting, 6–15 [13] Guttman,N.B.,1998:ComparingthePalmerDroughtIndexandtheStandardized Precipitation Index. Journal of the American Water Resources Association, 34: 113–121. DOI: 10.1111/j.1752-1688.1998.tb05964.x. [14] Vicente-Serrano,S.M.,S.BegueriaandJ.I.Lopez-Moreno,2010:Amulti-scalar drought index sensitive to global warming: the Standardized Precipitation Evapotranspiration Index. Journal of Climate, 23: 1696–1718. DOI:10.1175/2009JCLI2909.1. [15] McKee, T. B., N. J. Doesken, and J. Kleist, 1993: The relationship of drought frequencyanddurationoftimescales.EighthConferenceonAppliedClimatology, American Meteorological Society, Jan17-23, 1993, Anaheim CA, pp.179-186. [16] Ashok K. Mishra, Vijay P. Singh, Drought modeling – A review, Journal of Hydrology,Volume403,Issues1–2,2011,Pages157-175,ISSN0022-1694, https://doi.org/10.1016/j.jhydrol.2011.03.049. [17] Ashok K. Mishra, Vijay P. Singh, A review of drought concepts, Journal of Hydrology,Volume391,Issues1–2,2010,Pages202-216,ISSN0022-1694, https://doi.org/10.1016/j.jhydrol.2010.07.012. [18] DikshitA,PradhanB,HueteA.AnimprovedSPEIdroughtforecastingapproach using the long short-term memory neural network. J Environ Manage. 2021 Apr 1;283:111979. doi: 10.1016/j.jenvman.2021.111979. Epub 2021 Jan 19. PMID:33482453. [19] Poornima,S.,&Pushpalatha,M.(2019).DroughtpredictionbasedonSPIandSPEI with varying timescales using LSTM recurrent neural network. Soft Computing, 23, 8399 - 8412. [20] A.Mokhtaretal.,\"EstimationofSPEIMeteorologicalDroughtUsingMachine Learning Algorithms,\" in IEEE Access, vol. 9, pp. 65503-65523, 2021, doi: 10.1109/ACCESS.2021.3074305. [21] H. Balti et al., \"Big data based architecture for drought forecasting using LSTM, ARIMA,andProphet:CasestudyoftheJiangsuProvince,China,\"2021International Congress of Advanced Technology and Engineering (ICOTEN), Taiz, Yemen, 2021,pp. 1-8, doi: 10.1109/ICOTEN52080.2021.9493513. [22] Xu D, Zhang Q, Ding Y, Zhang D. Application of a hybrid ARIMA-LSTM model based on the SPEI for drought forecasting. Environ Sci Pollut Res Int. 2022 Jan;29(3):4128-4144.doi:10.1007/s11356-021-15325-z.Epub2021Aug17.PMID: 34403057. [23] Abhirup Dikshit, Biswajeet Pradhan, Abdullah M. Alamri, Long lead time drought forecasting using lagged climate variables and a stacked long short-term memory model,ScienceofTheTotalEnvironment,Volume755,Part2,2021,142638,ISSN0048-9697, https://doi.org/10.1016/j.scitotenv.2020.142638. [24] Wu,X.,Zhou,J.,Yu,H.,Liu,D.,Xie,K.,Chen,Y.,…Xing,F.(2021). The developmentofahybridwavelet-Arima-lstmmodelforprecipitationamountsand drought analysis. Atmosphere, 12(1). https://doi.org/10.3390/ATMOS12010074 [25] Iban Ortuzar, Ana Serrano, Àngels Xabadia, Macroeconomic impacts of water allocationunderdroughts.Accountingforglobalsupplychainsinamultiregional context, Ecological Economics,Volume 211, 2023, 107904, ISSN 0921-8009, https://doi.org/10.1016/j.ecolecon.2023.107904.
Copyright © 2025 Angel R, Kokila Ramesh, Anita Chaturvedi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Paper Id : IJRASET66196
Publish Date : 2024-12-30
ISSN : 2321-9653
Publisher Name : IJRASET
DOI Link : Click Here