Annually, 250 million people from all over the world are adversely affected by floods, often resulting in thousands losing their lives and economic damages amounting to nearly 10 billion dollars. Google studies also reveal that an effective global flood forecasting system, that signals early warnings, can prevent 43% of the fatalities and 35% of the economic damages. Floods are of course most common but the damages can be mitigated more easily than say, hurricanes. Early alerts can help evacuate people in low-lying areas by moving them to higher grounds. 

The Central Water Commission has been engaging with Google to send out alerts to people which helped save lives. These alerts provide critical information and are updated regularly that enable Android smartphone users to make informed choices about where they want to be. Particularly this year, when many parts of the country have grappled with flood situations. These notifications are currently in three languages – English, Hindi, and Bengali. A colour-coded map depicting the flooded areas is also shown and clicking on the same will take the user to Google maps where it can be zoomed to gain a better understanding.

Here’s a brief overview of how it all works. 

Stage 1 – Data Collection, forecasts & measurements

• Atmospheric models – weather forecasts, basin conditions, etc.
• Numerical Weather Prediction models (challenges are faced in data-scarce regions).
• Stream gauges – devices inside the river beds that measure water levels. It is considered a reliable source for building hydrologic models but unfortunately, their deployments are concentrated in Europe & America. 
• Elevation models tell us the structure of the ground and how the water flows. 

Stage 2 – Hydrologic Model

Once we have the data from stage 1, it gives us the amount of water that will flow around the river network while taking into consideration a host of factors such as atmospheric moisture, how much water is likely to be absorbed by the soil, plants, etc. 

Stage 3 – Hydraulic Model 

Taking inputs from the earlier stage, the output is a map detailing where and when the overflow of the river is likely to be and demarcate the areas where there’s a potential danger and where inundation is likely to be. The model takes input from something called the Digital Elevation Map which is created with the help of satellite & stereographic imageries. This map through numerous inputs estimates how the ground looks at every possible point. This data is critical to determine how the water will flow. 

Physics-based Models & Their Assumptions 

The assumption here is that the data captured is accurate. But in real life, that isn’t so. A greater part of the world is not mapped and the information is not easily available – at least not upto the resolution that’s desired. The SRTM Elevation Maps with low resolution prove to be useless in many cases. Often, the images cannot tell if the rivers are overflowing or not. Moreover, some rivers flow incredibly fast and change directions which means that these maps have to be updated every year which is not always possible.  

ML-based Systems 

The alternative is a Multi-Asset Space Alignment (MASA). A very large amount of standard optical imagery is being captured all over the world and at all times. But the angles of the cameras aren’t necessarily uniform because they are deployed for very different missions. However, the light reflected from the surface (time taken) can be used to calculate the gradient and thereby arrive at the surface geology. Of course, this requires 100s and thousands of images to arrive at a reasonably accurate map using ML-based tools. Then the Digital Surface Model is created. Convolutional network based on terrain study is fed into the model to arrive at the Digital Terrain Model. What’s the difference? Trees and buildings both have heights but trees have thin barks so the water flows around it whereas buildings distinctly obstruct the water flow. Again, this is possible when we have labelled data. But the accuracy is incredible. These models can be made to work anywhere in the world with 1-metre resolution – particularly in data-scarce regions. The challenge is high resolution requires a very high computational cost as well. However, these models can also work well with lower resolution. In a flood situation, even a 100 m resolution should suffice which means that it’s okay if we know accurately within 100 metres of the flood-affected area (likely). We don’t always require a 1 m resolution. 

In Patna recently, there was a perplexing incident. A part of the area was flooded which was not captured in the model. Interestingly, this area wasn’t even close to the river banks so it got the scientists to think what could be the possible reason. Was it due to rainfall? So a team flew down to Patna and physically inspected the area. What they found out is most interesting. At the embankment, there was a sluice(gate) which someone had opened to let the waters into the agricultural land. Despite the supreme efficacy of these models, there will always be these outliers which will impact accuracy. Accuracy will develop with time and when every known instance is updated in the model. 

When physics-based models can’t be accurately formed due to scarcity of data, these conceptual models work accurately. They can replicate physics-based models with great accuracy which can then be mirrored in other parts of the world with similar conditions (such as soil, moisture, etc.) 

Actions Taken & Averting Disasters 

Much of this work done by Google has been in India. Sella Nevo, Senior Software Engineer, Google Research, Tel Aviv, recently presented some of these ideas at RAISE, the Global AI Conference hosted by GoI where NASSCOM was a supporting partner. He writes in his blog: ”In recent months, we’ve been expanding our forecasting models and services in partnership with the Indian Central Water Commission. In June, just in time for the monsoon season, we reached an important milestone: our systems now extend to the whole of India, with Google technology being used to improve the targeting of every alert the government sends. This means we can help better protect more than 200 million people across more than 250,000 square kilometres—more than 20 times our coverage last year. To date, we’ve sent out around 30 million notifications to people in flood-affected areas.”

What happens to people who aren’t connected to the internet and as we know, they are usually the most affected. Google has been working with local NGOs and they’ve come up with unique ideas. One of them being volunteers on motorcycles displaying different-coloured flags to convey the potential danger in a particular area.

We have mother nature on one side and man & machine on the other trying to mitigate her vagaries. 

A blog like this cannot be complete unless we raise a toast to all those millions of people who lost their livelihoods but fought back tooth and nail to start all over again. A toast to their resilience!