N et al. (2026) A hybrid CNN-XGBoost-Temporal Fusion Transformer framework for explainable multi-variable hydroclimatic extremes forecasting under CMIP6 climate scenarios

Identification

• Journal: Frontiers in Environmental Science
• Year: 2026
• Date: 2026-09-22
• Authors: Mohammed Faizan N, Sujatha V
• DOI: 10.3389/fenvs.2026.1933637

Research Groups

• Department of Mathematics, School of Advanced Sciences, Vellore Institute of Technology, Vellore, Tamil Nadu, India
• Department of Quantum AI, School of Computer Science and Engineering, Vellore Institute of Technology, Tamil Nadu, India

Short Summary

This study developed CXTF-Net, a hybrid CNN-XGBoost-Temporal Fusion Transformer framework for explainable, multi-variable, multi-horizon hydroclimatic forecasting. Applied to the Mahanadi River Basin, it significantly outperformed benchmarks in predicting streamflow and other hydroclimatic extremes, projecting a decline in rainfall and increased drought frequency under CMIP6 scenarios.

Objective

• To develop an explainable, multi-variable, multi-horizon hybrid deep learning framework (CXTF-Net) that integrates spatial and temporal processing for forecasting hydroclimatic extremes, and to apply it for climate impact assessment under CMIP6 scenarios in the Mahanadi River Basin.

Study Configuration

• Spatial Scale: Mahanadi River Basin, India (141,589 km²), with gridded data at 0.25° resolution.
• Temporal Scale: Historical data from 1980–2022 (43 years); lookback window of 90 days; forecast horizons from 1 to 30 days.

Methodology and Data

• Models used:
  • CXTF-Net: Hybrid CNN-BiLSTM encoder, XGBoost feature selection gate, Temporal Fusion Transformer (TFT) decoder.
  • Baselines: ARIMA(2,1,2) with SARIMA, 3-layer Artificial Neural Network (ANN), Support Vector Regression (SVR) with RBF kernel, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), CNN-LSTM, Vanilla Transformer.
• Data sources:
  • ERA5 (ECMWF): 2-m temperature, SST, 850 hPa zonal and meridional wind, 500 hPa geopotential height, soil moisture (layer 1–4), actual evapotranspiration, total precipitation (0.25° daily).
  • CMIP6-ESGF (6 GCMs): Precipitation, 2-m temperature, SST (1.0° monthly).
  • MODIS Terra (NASA): Normalized Difference Vegetation Index (NDVI), land cover (500 m–1 km monthly).
  • Sentinel-2 (ESA Copernicus): Normalized Difference Water Index (NDWI), surface water extent (10 m 5-day).
  • IMD gauge network/India-WRIS: Daily rainfall (60 stations), daily streamflow (4 stations) (Point daily).
  • NOAA ERSST v5/NCEP: Niño 3.4 SST anomaly (ENSO index), Dipole Mode Index (IOD), Madden-Julian Oscillation (MJO) phase (2.0° monthly).

Main Results

• CXTF-Net achieved a Nash-Sutcliffe Efficiency (NSE) of 0.934, Kling-Gupta Efficiency (KGE) of 0.921, and Root Mean Square Error (RMSE) of 18.4 m³/s for 1-day-ahead streamflow forecasting, outperforming seven benchmark models.
• NSE exceeded 0.79 at the 30-day horizon across all four target variables: streamflow, 12-month Standardised Precipitation Index (SPI-12), compound flood risk index (CFRI), and basin water deficit (BWD).
• Multi-task learning improved NSE by 0.026 over single-task streamflow prediction, with the advantage growing to 0.051 at a 30-day lead time.
• DeepSHAP attribution and attention rollout identified Sea Surface Temperature (SST) anomalies, ENSO Niño 3.4 index, and antecedent soil moisture as the three dominant hydroclimatic drivers, with physically consistent horizon-dependent transitions.
• CMIP6 SSP5-8.5 projections indicate a statistically significant decline in mean annual basin rainfall (–120 mm) and a near-tripling in severe drought frequency (from 8% to 22%) by 2080–2100.
• The 80% prediction interval for streamflow showed an empirical coverage of 81.0% with a Mean Prediction Interval Width (MPIW) of 41.2 m³/s, confirming well-calibrated probabilistic outputs.

Contributions

• Developed CXTF-Net, a novel three-stage hybrid deep learning framework integrating a CNN-BiLSTM encoder, XGBoost feature selection gate, and Temporal Fusion Transformer (TFT) decoder for end-to-end spatial-temporal feature extraction, dimensionality reduction, and multi-horizon probabilistic forecasting.
• Generated simultaneous calibrated quantile forecasts for four distinct hydroclimatic target variables (streamflow, SPI-12, compound flood risk index, and basin water deficit) within a unified multi-task learning framework, demonstrating superior performance over single-task models.
• Embedded a dual explainability module combining DeepSHAP attribution and attention rollout, providing horizon-specific, physically validated identification of dominant teleconnections.
• Applied the framework to generate CMIP6-driven projections of hydroclimatic extremes for the Mahanadi River Basin, delivering the first deep-learning-based multi-variable climate impact assessment for this strategically important Indian basin.

Funding

• The authors declared that financial support was not received for this work and/or its publication.

Citation

@article{N2026hybrid,
  author = {N, Mohammed Faizan and V, Sujatha},
  title = {A hybrid CNN-XGBoost-Temporal Fusion Transformer framework for explainable multi-variable hydroclimatic extremes forecasting under CMIP6 climate scenarios},
  journal = {Frontiers in Environmental Science},
  year = {2026},
  doi = {10.3389/fenvs.2026.1933637},
  url = {https://doi.org/10.3389/fenvs.2026.1933637}
}