On Data-Driven Science in Hydrology

In a traditional approach, we lean on centuries of scientific thought, painstakingly piecing together processes—like runoff, infiltration, groundwater flow, and sediment transport—into large, complex models. Now, with an explosion of data and machine learning methods, there’s a push to let the data itself guide our understanding. Instead of relying solely on a stack of equations, we stitch together relationships through patterns found in the data. It feels like a new kind of science, one where we feed in enough observations and let flexible algorithms fill in the gaps.

But as exciting as this is, I also worry about what happens when we step into completely uncharted territory. Data-driven models might predict well within the range of what we’ve seen before enough, but how do they hold up in never-before-seen conditions? That’s where these methods might struggle. We might need more than just neural networks connecting the dots, we might need models that learn underlying mechanisms, not just correlations. The dream is some hybrid approach that understands processes at a very fundamental level while still leveraging the power of massive datasets?

At the same time, there’s never been a better moment to dig deeper into data. With decades’ worth of satellite imagery, sensor networks, and massive archives of measurements, we’re definitely equipped to get creative. We need to look beyond just predicting tomorrow’s river flow and start asking bigger questions about our water resources—how they change, what they carry, and where they’re heading. If we can figure out how to generalize these methods, to make them robust against uncertainty and new scenarios, the payoff could be huge. It might take quantum computing or entirely new algorithms to get there, but the vision is clear: blending data-driven insights with fundamental processes could open up horizons in hydrology we’re only just starting to imagine.