Yan et al. (2026) Identification of spatiotemporal sensitivity of karst spring discharge to precipitation and evaporation: an LSTM–Transformer Decoder framework

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-22
• Authors: Xiping Yan, Meihong Ma, Yonghong Hao, Chunmei Ma, Wenrui Zhang, Qi Wang, Xiaogang Shi, Huiqing Hao, Xinyao Luo, Tianchyi Jim Yeh
• DOI: 10.1016/j.jhydrol.2026.136083

Research Groups

• Department of Hydrology, Peking University
• Institute of Karst Geology, Chinese Academy of Sciences
• School of Earth and Space Sciences, University of Science and Technology of China

Short Summary

This study develops an LSTM–Transformer Decoders (LSTM–TDs) framework to predict spring discharge in karst systems by linking discharge-difference prediction with attention-based hydroclimatic sensitivity identification. The framework outperformed benchmark models in predicting the Niangziguan Springs' discharge.

Objective

• Investigate the feasibility of using LSTM–Transformer Decoders (LSTM–TDs) for predicting spring discharge in karst systems and identifying hydroclimatic sensitivity patterns.

Study Configuration

• Spatial Scale: Local to regional scale, focusing on the Niangziguan Springs catchment area.
• Temporal Scale: Daily to seasonal scale, with a focus on simulating short-term to medium-term discharge dynamics.

Methodology and Data

• Models used: LSTM–Transformer Decoders (LSTM–TDs) framework, incorporating Long Short-Term Memory (LSTM) networks and Transformer Decoder pathways.
• Data sources: Historical precipitation and evaporation data from satellite observations and reanalysis datasets.

Main Results

• The LSTM–TDs framework achieved high accuracy in predicting spring discharge at the Niangziguan Springs, with NSE = 0.85, RMSE = 0.34 m³/s, MAE = 0.26 m³/s, and MAPE = 3.88%.
• Attention-based sensitivity identification revealed hydrologically plausible spatial and temporal patterns of precipitation and evaporation relevance.
• The model-derived sensitivity turning points were approximately 42.0 mm for precipitation and 130.2 mm for evaporation.

Contributions

• This study provides a novel framework (LSTM–TDs) for predicting spring discharge in karst systems while offering interpretable diagnostics of hydroclimatic sensitivity patterns.
• The results demonstrate the potential of using attention-based mechanisms to identify key drivers of discharge dynamics in complex karst aquifers.

Funding

• National Natural Science Foundation of China (Grant No. 41922019)
• Key Research and Development Program of Guangdong Province (Grant No. 2019B121205001)

Citation

@article{Yan2026Identification,
  author = {Yan, Xiping and Ma, Meihong and Hao, Yonghong and Ma, Chunmei and Zhang, Wenrui and Wang, Qi and Shi, Xiaogang and Hao, Huiqing and Luo, Xinyao and Yeh, Tianchyi  Jim},
  title = {Identification of spatiotemporal sensitivity of karst spring discharge to precipitation and evaporation: an LSTM–Transformer Decoder framework},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136083},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136083}
}