Enter Big Tech: Google takes on flood forecasting

Last year, Google rolled out a pilot flood forecasting service in Patna, India. For an unnecessarily dramatic introduction, look no further than this video:

Naturally, there are grand plans for expansion "anywhere in the world".

But's first let's take a look at how this forecasting system works.

Forecasting floods in river basins is a matter of taking predictions about water in (usually from rainfall forecasts and upstream water measurements) and get a forecast of water out (how much of it there will be where and when, and for how long).

Did Google decide that "Don't Be Evil" was too hard? Yes. Do they have many, many shiny gifs of inundation models? Also yes.

Google uses a hydraulic model describing the flow of water in the river based on real time water level data from the Central Water Commission to generate a forecast rate of discharge, which becomes the input for an inundation model, predicting how the water will flow in the floodplain. This is based on Google's own highly detailed digital elevation models developed by using machine learning to clean up satellite images. The model is corrected against historical flood extent data collected from satellites, and the output is converted into risk maps, made available on Google products.

An example of the kind of risk map that Google's model produces.

Google are aiming to produce very high-resolution forecasts. This has always been a challenge with urban flood forecasting: terrain and landcover vary on a micro scale which makes accurate modelling a nightmare; and low spatial resolution forecasts are not terribly useful in cities. Unsurprisingly, obtaining this level of detail requires huge amounts of computational power to solve the equations in the hydraulic model; inevitably, Google has that covered. So far the model has produced maps that are 75% accurate at a resolution of 300m; not an enormously high resolution, but it's a work in progress.

Can it be replicated everywhere else? Given the high computational demands, I'm doubtful, though it might be possible in a few years. But this model requires real-time and historical data, which isn't available for most river basins. Data-poor basins require new conceptual models, and approaches towards downscaling and remote estimation of water levels; this might be in the future of the project, but it's not there yet.

What then are the wider implications of Google's involvement in flood forecasting? As world leaders in machine learning, they have resources that governments and universities could only dream of. If they can help people through leveraging their expertise, isn't that a good thing?

Quite apart from Google's chequered history making me cynical about how they will eventually find a way to profit from this, I feel a bit uncomfortable when tech behemoths get involved in providing public services. Governments are prepared to give big firms a lot of leeway, but that rarely comes with increased scrutiny and public accountability.

Indian rescue services evacuate residents of Patna during flooding last month. The floods were blamed on a failure of the urban drainage and pumping systems. Photo: PTI

Ultimately any forecast is only as useful as the way it's communicated. Smartphone penetration in India is relatively high (26% in 2018), but the majority of people won't have access to the risk maps: multiple low-tech avenues will be needed to ensure that warnings are communicated in a clear and actionable manner to everybody at risk. Google are partnering with local NGOs to achieve this, and claim to be actively soliciting feedback from communities, which is a good sign; I hope that local residents are able to have a real say in developing a warning system that works for them.