More than 95% of small to medium-sized watersheds worldwide lack monitoring data, posing a long-standing scientific challenge in hydrology for predicting runoff and floods in these data-scarce regions. Recently, a team led by Dr. Ouyang Chaojun from the Institute of Mountain Hazards and Environment at the Chinese Academy of Sciences introduced a novel AI-based runoff flood prediction model called ED-DLSTM. By incorporating watershed static attributes and meteorological drivers, the model was trained using data from over 2000 hydrological stations globally to address runoff prediction issues in both gauged and ungauged watersheds. The research findings have been published online in Innovation.
The ED-DLSTM model proposed in this study, tailored for watershed runoff prediction, features a spatial attribute encoding module that utilizes convolutional layers and spatial pyramid pooling layers to map the static attributes of all watersheds to a uniformly sized latent space. This enables the model to abstractly "understand" the hydrological response characteristics of different watersheds.
The training dataset used in the study comprises 2089 watersheds from the United States, United Kingdom, Central Europe, Canada, and other regions, ensuring dataset diversity given the significant distribution differences among these watersheds. By training the model on historical data from these watersheds and testing its predictive accuracy and reliability for future periods, the researchers evaluated the experimental results using the Nash-Sutcliffe Efficiency coefficient (NSE). They found that 81.8% of watersheds had an average NSE greater than 0.6, indicating better predictive accuracy compared to traditional hydrological models and other AI models.
Building upon the pre-trained model (Northern Hemisphere), the researchers conducted predictions for 160 new watersheds in Chile (Southern Hemisphere) without using any monitoring data to assess the model's predictive capabilities in ungauged watersheds. The results from different pre-trained models exhibited strong spatial distribution consistency. In the best-case scenario, over 76.9% of ungauged watersheds had NSE values greater than 0, showcasing the significant potential of AI in predicting water runoff and floods in ungauged watersheds.
For more information on the related paper, visit: https://doi.org/10.1016/j.xinn.2024.100617
