Thoughts
"What language is thine, O sea?"
"The language of eternal question."
"What language is thy answer, O sky?"
"The language of eternal silence."
When I was a kid, I often found myself lost in the drifting clouds and distant stars. Upon closing my eyes, I was always greeted by the vision of a blue, spinning planet, with flows around it - atmospheric circulation, riverine discharge, population mobility, and global capital flows, among other phenomena. 🌏
In 2017, I stepped into the College of Hydrology and Water Resources at Hohai University. Interestingly, my birthday coincides with World Water Day, on the 22nd of March. 💧
In recent years, the successful application of plenty of AI technologies across various sectors has attracted a large number of hydrologists and scholars, stimulating them to integrate these techniques into hydrological research. This is a field brimming with passion and skepticism, one that also forms the main thread of my master's research - exploring the application of deep learning in hydrological modeling. Currently, my thoughts on this subject are mainly concentrated on the following points:
The importance of generalization ability
After an initial period of exploratory research where various models, research areas, and target tasks were simply switched around, the research on AI hydrological models might now be ready for the next step. In my view, this advancement should emphasize the pursuit of superior spatiotemporal generalization, enhanced robustness, and the capability for multi-task learning, as these are fundamental features for the advancement of sophisticated hydrological modeling.The relationship of AI models with physical laws
A common pitfall of many AI models is their emphasis on data while overlooking the constraints imposed by physical laws, which can significantly undermine the confidence of hydrologists in these models. I envisage that, in the context of runoff modeling, an integrated approach that couples deep learning-driven runoff generation models with hydraulics-based runoff routing models might be a promising avenue worth further exploration.The aspect of explainability
Beyond the conventional methods of enhancing explainability in deep learning, another potential strategy could involve transitioning from end-to-end models to models that offer detailed outputs for each process stage. For instance, the credibility of AI-based runoff generation models might be significantly enhanced if they provide detailed, quantified outputs at each stage of the water dissipation in the runoff generation process, which includes interception, infiltration, and evapotranspiration.
Finally, actions speak louder than words. Forming these thoughts and hypotheses is just the beginning; the key lies in learning through practice. While it's thrilling to observe the rapid train of AI hurtling by, I aspire to be a passenger on board rather than a bystander. 🚅