@article{Li2026, 
author = {Bin-quan Li and Yi-jie Xia and Si-ji Tao and Yun-yao Chen and Jian-fei Zhao and Zhong-min Liang},
title = {Integrating process-based and deep learning models for flood simulation in karst basins},
year = {2026},
journal = {Water Science and Engineering},
volume = {19},
number = {1},
pages = {23-34},
keywords = {Karst basin, Karst-Xin'anjiang model, Long short-term memory neural network, Bayesian model averaging, Flood simulation},
url = {https://www.sciopen.com/article/10.1016/j.wse.2025.11.005},
doi = {10.1016/j.wse.2025.11.005},
abstract = {Flood process simulation in karst basins is challenging due to complex runoff generation and concentration mechanisms, often resulting in low accuracy. This study investigated two typical karst basins (the Maiweng and Liudong river basins) in Guizhou Province, China, and developed two hydrological models for flood simulation: the karst-Xin'anjiang (Karst-XAJ) model, a modified Xin'anjiang (XAJ) hydrological model adapted for karst runoff characteristics, and the long short-term memory (LSTM) deep learning model. Their performances were compared, and their results were integrated using Bayesian model averaging (BMA). The Karst-XAJ model accurately simulated flood peak time and runoff depth but showed limited peak flow accuracy. The LSTM model performed well within a 2-h computational window, with accuracy declining for longer computational windows (3—4 h) yet maintaining a Nash—Sutcliffe model efficiency coefficient above 0.7. The BMA approach further enhanced simulation accuracy beyond individual models. Overall, both models effectively captured flood dynamics in karst basins, with the LSTM model achieving superior precision. This study offers a novel framework for simulating flood processes in karst regions with complex runoff processes.}
}