Explainable hybrid deep learning for monthly precipitable water vapor forecasting over Türkiye
Meteorology and Atmospheric Physics, cilt.138, sa.5, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 138 Sayı: 5
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s00703-026-01179-y
- Dergi Adı: Meteorology and Atmospheric Physics
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Environment Index, Geobase, INSPEC, Natural Science Collection (ProQuest), Earth, Atmospheric, & Aquatic Science Collection (ProQuest), Technology Collection (ProQuest)
- Çukurova Üniversitesi Adresli: Evet
Özet
Accurate prediction of precipitable water vapor (PWV) is important for weather forecasting, hydrological modeling, and climate diagnostics. This need is particularly relevant in regions with complex topography and limited observational coverage. Existing models often struggle to capture temporal dependencies, maintain predictive accuracy across heterogeneous regions, and provide interpretable forecasts. To address these limitations, this study proposes an explainable hybrid deep learning framework in which each module targets a specific forecasting challenge. One-dimensional convolutional neural networks (1D-CNNs) extract localized month-to-month patterns from the meteorological sequences. The adaptive scaled dot-product attention (ASDPA) mechanism uses learnable scaling and temperature parameters to refine the weighting of informative temporal features. Bidirectional recurrent modules capture dependencies across the 12-month input window. Three model variants were developed using the bidirectional gated recurrent unit (BiGRU), bidirectional long short-term memory (BiLSTM), and a hybrid BiGRU–BiLSTM module. To improve robustness and spatial generalization, a stacking ensemble combines the proposed deep model with four machine learning regressors. Shapley additive explanations (SHAP) and gradient-weighted class activation mapping (Grad-CAM) provide model interpretation by identifying influential predictors and relevant temporal segments, respectively. The models were trained using monthly ERA5-derived data from eight representative cities in Türkiye. Their performance was evaluated on a national-scale dataset covering 81 cities. Among the individual architectures, CNNASDPA-BiGRU provided the best balance between predictive accuracy and computational efficiency at the reference stations. It achieved and an RMSE of 1.058 kg m. At the national scale, the stacking ensemble achieved the best overall performance, with an RMSE of 1.358 kg m. SHAP identified air temperature and dew point temperature as the dominant predictors. Grad-CAM highlighted temporal segments associated with pronounced moisture variability and dry-wet transitions. Overall, the proposed framework provides an accurate, interpretable, and scalable approach for monthly PWV forecasting over Türkiye.