A top-down deep learning model for predicting spatiotemporal dynamics of groundwater recharge
Xin Huang, Lei Gao, Nan Zhang, Russell S. Crosbie, Long Ye, Jian Liu, Zhaoxia Guo, Qinghua Meng, Guobin Fu, Brett A. Bryan
Abstract
This paper proposes s-LSTM, a top-down deep learning model, for efficiently modelling and predicting the spatiotemporal dynamics of groundwater recharge. The model's effectiveness was evaluated and compared with three bottom-up machine learning models using a case study of 246 bores in South Australia. The results demonstrate that s-LSTM outperformed the three bottom-up models by 3.7%–28.9% in prediction performance (based on root mean squared error) and reduced training time by 37.4%–99.5%. Furthermore, s-LSTM exhibited superior capabilities in terms of interpolation/upscaling and identifying influential factors to recharge prediction at the regional scale. Groundwater extraction, mean number of wet days per year, seasonal minimum temperature, seasonal rainfall, and seasonal actual evaporation were identified by s-LSTM as the top five influential factors for groundwater recharge prediction in the study region. Practitioners are encouraged to consider the adoption of s-LSTM when addressing spatiotemporal data modelling and prediction challenges.