Google’s Machine Learning Model for Flood Forecasting Expands Coverage
When it rains, it pours, but Google’s breakthrough Machine Learning model warns you before the torrent knocks your doors
“What if people could predict natural disasters before they happen?”
This is a question that the team behind Google’s AI For Social Good initiative has been steadfastly looking to answer for years, in theory, and in praxis.
What appears to be turning the tides is the fact that it may be very close to an answer.
In 2018, Google embarked on the endeavour to harness artificial intelligence (AI) and computational resources to place humanity one step ahead of erratic, torrential floods of debilitating destructive potential, and never looked back. The flood forecasting pilot project was initiated in the Patna region of India, which has now expanded coverage to the larger Indian subcontinent and parts of Bangladesh, protecting over 250 million people through precise forewarning systems that provide greater response time during the onset of disaster, when every second matters.
The pan-Asian expansion comes as a precursor to a more widespread coverage to countries all over the world, as the tech giant plans to make the forewarning systems globally accessible.
Here’s a peek into Google’s novel approach to flood forecasting.
The Convergence of Physics and Machine Learning
Google’s proprietary morphological inundation model interlaces physics-based modelling and machine learning (ML) to create precise inundation models, involving running and re-running thousands of simulations to identify risk-prone areas.
High-resolution satellite imaging contributed in the development of elevation maps, upon which physics-based models simulated the flow of water. The vast terrain data obtained from the firm’s satellite partners covers hundreds of millions of square kilometres for inundation modelling.
Now, three impediments stood in the way of precision inundation modelling.
The first pertains computational complexity. The size and scale of the large areas involved and the resolution required for precise simulation of such models entail high computational complexity, which, in large-scale computation is the cost associated with solving a problem. This implies a significantly large number of operations and computing memory is required to model with precision.
The second is inadequate data. According to Google, “most global elevation maps don’t include riverbed bathymetry, which is important for accurate modelling”.
The third and the final is errors in the existing data, which may come as errors in gauge measurement, missing features in elevation maps, and in other forms, which require correction.
Google’s novel approach referred to as the ‘morphological model’ however, offers an innovative workaround encapsulating the best of machine learning and physics-based modelling.
In lieu of modelling complex hydrodynamics of a given water body, the model simulates inundation through ‘simple physical principles’.
In an April article, I covered Google’s ‘physics-free’ approach involving machine learning algorithms in its project ‘Machine Learning For Precipitation Nowcasting from Radar Images’ and how its proprietary model had been challenging the industry standard HRRR atmospheric model. In a similar approach, Google will train a pure machine learning model with (no access to extrapolate physics-based information) for the estimation of a unidimensional river profile from gauge measurements.
The input for the ML model is a stream gauge installed which monitors the water level. An increase in the reading of the gauge at one point will point towards an increase in the water level at all points, since the gauge assumes the level increases monotonically. It is also reasonably assumed that the elevation subsequently decreases as the river flows downstream.
The synthetic elevation map thus obtained is used to model flood behaviour by running a flood-fill algorithm.
The morphological model, therefore, allows circumvention of classical physics-based models with hard-to-counter impediments and limitations whilst simultaneously providing large scale flood forecasts.
“This morphological model improves accuracy by 3%, which can significantly improve forecasts for large areas, while also allowing for much more rapid model development by reducing the need for manual modelling and correction” added Google.
Predictive analytics almost extensively utilize machine learning for modelling data owing to its insurmountable ability to recognise patterns and process gargantuan chunks of data to deduce meaningful insights.
While we’re only witnessing the tip of the ginormous iceberg that represents ML’s full potential, it appears that Google has found a way to prevent floods from dampening our progress and our spirits alike.
Subscribe to the Blog
Join for the latest tech, design and industry news straight to your inbox.
Our UX team designs customer experiences and digital products that your users will love.