Machine Learning-Based Hydrological Drought Prediction Integrating Teleconnections and Hydrological Memory in a Semi-Arid Basin, Algeria
Atmosphere, cilt.17, sa.7, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 17 Sayı: 7
- Basım Tarihi: 2026
- Doi Numarası: 10.3390/atmos17070670
- Dergi Adı: Atmosphere
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, Geobase, INSPEC, Directory of Open Access Journals, Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Earth, Atmospheric, & Aquatic Science Collection (ProQuest)
- Anahtar Kelimeler: Algeria, atmospheric teleconnection, drought, precipitation, semi-arid regions
- Erzincan Binali Yıldırım Üniversitesi Adresli: Evet
Özet
Hydrological drought forecasting in semi-arid basins is challenging due to the combined influence of meteorological forcing, large-scale atmospheric teleconnections, and basin memory processes, which are rarely jointly analysed within a leakage-free predictive framework. This study addresses this gap by evaluating gradient-boosted trees and neural forecasting models for one-month-ahead prediction of the Standardized Runoff Index (SRI) in two sub-basins of the Wadi Sahaouat Basin, Algeria. The models include gradient-boosted regression trees (GBRT), A-N-BEATS, A-N-HiTS, and TiDE, representing distinct forecasting architectures. Predictors consist of the Standardised Precipitation Index (SPI), seven teleconnection indices (NAO, AO, EAWR, SCAND, MEI, SOI, WeMO), and their one- to three-month lags. Two scenarios are tested: Scenario 1 uses SPI and teleconnection lags only, while Scenario 2 additionally includes lagged SRI values (SRI_lag1–3) to represent hydrological memory. A train-only Variance Inflation Factor (VIF > 10) procedure is applied to remove multicollinearity without data leakage. In Basin 1, SRI lags were excluded due to strong collinearity with SPI lags (r = 0.984), resulting in identical inputs for both scenarios. In Basin 2, SRI lags were retained to assess their predictive contribution. GBRT achieved the best overall performance across both basins and scenarios, with mean RMSE, NSE, and KGE values of 0.0682, 0.9907, and 0.8945, respectively. TiDE ranked second overall, with a mean RMSE of 0.1166, followed by A-N-HiTS in third place with a mean RMSE of 0.1203 and A-N-BEATS with the weakest overall performance, with a mean RMSE of 0.2159. These results indicate that gradient-boosted trees remain highly competitive with neural models for small monthly hydrological datasets and that the value of hydrological memory is basin-dependent and varies according to its independence from concurrent meteorological forcing.