Development of deep learning approaches for drought forecasting: a comparative study in a cold and semi-arid region

Basit Öğe Kaydı
dc.authorid0000-0001-7387-0224
dc.authorid0000-0002-2898-3681
dc.authorid0000-0003-1796-4562
dc.authorid0000-0002-4767-6660
dc.contributor.authorGharehbaghi, Amin
dc.contributor.authorGhasemlounia, Redvan
dc.contributor.authorVaheddoost, Babak
dc.contributor.authorAhmadi, Farshad
dc.date.accessioned2026-02-08T15:15:02Z
dc.date.available2026-02-08T15:15:02Z
dc.date.issued2025
dc.departmentBursa Teknik Üniversitesi
dc.description.abstractDrought is an intricate natural disaster that substantially menace the world. Its exact forecasting has a remarkable impact in several parts such as food production, business, industry, etc. In this study, in order to assess the drought procedure in Mahabad River basin, the temporal meteorological reconnaissance drought index (RDIMRB) in four diverse time scales including 3, 6, 9, and 12-month are computed using 576 monthly climatic datasets recorded from Sep 1974 to Aug 2022. To predict the time series RDIMRB, different standalone deep neural network (DNN) models including LSTM, GRU, Bi-directional LSTM (Bi-LSTM), and Bi-directional GRU (Bi-GRU) with the sequence-to-one regression module of forecasting (seq2one) are developed. For sake of this aim, the first 70% of data (395 months) and the last 30% of data (169 months) chronologically are used in the calibration and validation parts, respectively, to feed in the models development process. So as to achieve the most advantageous models' structure, a lot of scenarios are adopted by tuning the meant meta-parameters including NHU (number of hidden units), SAF (state activation function), and P-rate (learning dropout rate). According to the performance assessment criteria, total learnable parameters (TLP) criterion, and comparison plots, the Bi-GRU model is verified as the most satisfactory model, and best results are obtained in RDIMRB-12. It under the epitome meant meta-parameters achieved (i.e., NHU = 120, P-rate = 0.5, and softsign as the suitable SAF) results in the RMSE, MBE, NSE, PBIAS, and R2 of 0.17, 0.011, 0.92, -2.02%, and 0.86, respectively, nonetheless for the GRU model are gotten 0.64, 0.071, 0.23, 17.97%, and 0.65, respectively.
dc.identifier.doi10.1007/s12145-024-01650-7
dc.identifier.issn1865-0473
dc.identifier.issn1865-0481
dc.identifier.issue1
dc.identifier.scopus2-s2.0-85212688143
dc.identifier.scopusqualityQ1
dc.identifier.urihttps://doi.org/10.1007/s12145-024-01650-7
dc.identifier.urihttps://hdl.handle.net/20.500.12885/5569
dc.identifier.volume18
dc.identifier.wosWOS:001381233800001
dc.identifier.wosqualityQ2
dc.indekslendigikaynakWeb of Science
dc.indekslendigikaynakScopus
dc.language.isoen
dc.publisherSpringer Heidelberg
dc.relation.ispartofEarth Science Informatics
dc.relation.publicationcategoryMakale - Uluslararası Hakemli Dergi - Kurum Öğretim Elemanı
dc.rightsinfo:eu-repo/semantics/closedAccess
dc.snmzWOS_KA_20260207
dc.subjectReconnaissance drought index
dc.subjectTime series prediction
dc.subjectDeep neural network models
dc.subjectMahabad River basin
dc.subjectTLP criterion
dc.titleDevelopment of deep learning approaches for drought forecasting: a comparative study in a cold and semi-arid region
dc.typeArticle

Dosyalar

Koleksiyon

WoS İndeksli Yayınlar Koleksiyonu
Scopus İndeksli Yayınlar Koleksiyonu