Subhadarsini et al. (2026) EXtreFormer: a general deep learning framework for forecasting compound extreme events: experience with dry-hot extremes and vegetation response

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-28
• Authors: Suchismita Subhadarsini, D. Nagesh Kumar, Rao S. Govindaraju
• DOI: 10.1016/j.jhydrol.2026.135404

Research Groups

• Department of Civil Engineering, Indian Institute of Science Bangalore, Bengaluru 56012, India
• Lyles School of Civil and Construction Engineering, Purdue University, West Lafayette, IN 47907, USA

Short Summary

This study introduces EXtreFormer, a novel deep learning framework for long-horizon forecasting of compound dry-hot extreme events and vegetation response, demonstrating superior performance and interpretability in the Godavari River Basin. The framework effectively captures the complex interplay between temperature, soil moisture, and vegetation across diverse land use types, achieving high predictive accuracy for extreme events.

Objective

• To develop and evaluate EXtreFormer, a deep learning framework, for long-horizon forecasting of compound dry-hot extreme events (temperature and soil moisture) and their impact on vegetation (Normalized Difference Vegetation Index - NDVI) in a multivariate space.

Study Configuration

• Spatial Scale: Godavari River Basin, India.
• Temporal Scale: Long-horizon forecasting of extreme events.

Methodology and Data

• Models used: EXtreFormer (a deep learning framework incorporating advanced embedding (RoPE) and attention (mTAN) mechanisms). A Modified Transformer was used for comparison.
• Data sources: Data from the Godavari River Basin, including temperature, soil moisture, and Normalized Difference Vegetation Index (NDVI).

Main Results

• EXtreFormer achieved superior performance in forecasting compound extreme events over multiple horizons.
• The model demonstrated versatility across dominant land use and land cover (LULC) types (grasslands, broadleaf croplands, savannas, and deciduous broadleaf forests).
• Tail-Kling-Gupta Efficiency (tKGE) scores up to 0.86 were achieved, indicating robustness and effectiveness in capturing unique dynamics across LULC categories.
• Attention maps revealed that regular events maintained balanced variable interactions, while compound extreme events tended to prioritize temperature and soil moisture.

Contributions

• Development of EXtreFormer, a novel deep learning framework specifically designed for long-horizon forecasting of compound extreme events.
• Integration of advanced embedding (RoPE) and attention (mTAN) mechanisms into EXtreFormer, enhancing both predictive performance and interpretability.
• Demonstrated the model's robustness and effectiveness across diverse land use and land cover types, a critical aspect for real-world applicability.
• Provided insights into the differing variable interactions during regular versus compound extreme events through the use of attention maps.

Funding

• Not specified in the provided text.

Citation

@article{Subhadarsini2026EXtreFormer,
  author = {Subhadarsini, Suchismita and Kumar, D. Nagesh and Govindaraju, Rao S.},
  title = {EXtreFormer: a general deep learning framework for forecasting compound extreme events: experience with dry-hot extremes and vegetation response},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135404},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135404}
}