Abstract
Flooding is one of the most destructive natural disasters in the world, and accurately mapping where floods occur is critical for emergency response. However, building a system that can do this globally is difficult because reliable flood data simply does not exist everywhere. Many flood maps are created manually, which is slow and expensive. Others are pulled from satellites or water gauges, which are not always available in remote or under-resourced regions.
This technical project proposes a system that works around this problem. Instead of requiring flood data everywhere, the system learns flood patterns from the places where data does exist, then applies those patterns globally using environmental information that is available worldwide. Some common data considered are precipitation measurements and terrain data from AlphaEarth geospatial embeddings.
The core of the system is a type of neural network called a conditional U-Net. This network is trained to recognize the relationship between environmental conditions and flooding. Once trained, it can predict the likelihood of flooding in any region on Earth, even regions with no historical flood records.
In practice, the system works by dividing the globe into small spatial tiles. It feeds environmental data into the trained model for each tile, generates a flood probability map, and stitches all the tiles together into a single global picture. Because the environmental data is updated regularly, this process can repeat daily, enabling continuous worldwide flood monitoring.
Flood prediction technology has improved dramatically in recent years. But this paper asks a different question: does better prediction actually lead to better protection for people? The paper argues no, and the reasons are social and institutional, not technical.
The paper uses two frameworks from Science and Technology Studies to explain why. The first, the Social Construction of Technology (SCOT), shows that technologies are not neutral. They are shaped by the priorities of the institutions that build and manage them. The second, Risk Society theory, argues that modern institutions are good at identifying risk but structured in ways that allow them to avoid responsibility for it.
These frameworks are applied to real-world flood disasters. The 2022 Pakistan floods affected over 33 million people, displaced 8 million, but were entirely predictable. Pakistan's vulnerability to monsoon flooding had been documented for decades. Yet when the floods came, the response was reactive, not preventative. Similarly, Cyclone Idai struck Mozambique in 2019 with strong advance meteorological warning, but institutional coordination failed to translate that warning into lasting protection. Nearly a year after the storm, households with damaged homes were still experiencing three times the normal rate of malaria infection.
A governance analysis of international agencies, national disaster authorities, and emergency alert systems reveals a consistent pattern. Institutions that produce forecasts formally label their outputs as "informational only" and transfer responsibility for action to local governments and communities. When failures occur, accountability is spread so thinly across agencies, regulators, and private carriers that no single institution can be held responsible.
The paper concludes that prediction alone is not enough. For flood forecasting to genuinely protect people, it must be connected to enforceable governance. Rules should require institutions to act when warning thresholds are crossed, not just report that they have been.
Both projects are studying the same problem from different angles. They both asks why does flood prediction so often fails to prevent flood harm. The technical project tackles the problem from the engineering side. It builds a system that can generate flood predictions in low infrastructure, high vulnerability regions of the world that current tools cannot reach. Then the STS paper identifies as most neglected by existing governance structures. But the STS paper raises an important caution. A more powerful prediction system does not automatically mean more protection. The design decisions embedded in any machine learning system are themselves institutional choices that can silently encode inequality. A globally deployed flood model is only as protective as the institutions that receive and act on its outputs. Taken together, the two projects make a joint argument. Technical progress in flood prediction is real and valuable. But it is not sufficient on its own. Closing the gap between knowing a flood is coming and protecting people from it requires both better models and better governance. Systems are only good if they are accountable, not just accurate.
