Tackling annual urban floods - Waterlogging stresses and detection with geospatial intelligence

The current floods in Bangalore are a perfect example of how even a developed city can be inundated in water. After weeks of continuous rainfall and waterlogging, daily commuters have suffered the most. They are either not getting cabs or are stuck in road blockage. Today we rely mostly on the meteorological department for weather prediction but we do not have an accurate solution for the real-time status of waterlogged areas.

At OLA, we are looking at multiple approaches to solve this problem:

• Use of telemetry data from cabs
• Remote sensing
• Combination of both for better accuracy

Using Telemetry data from cabs

Through an extensive network of vehicles that includes 2/3/4 wheelers, we receive anonymized data in the form of location pings containing geographical coordinates.

Here is an example of telemetry data near the Bangalore Airport, The red dots show the geographical coordinates of vehicles passing through a particular lane or road.

With a dataset of regular telemetry data, we can calculate the day-wise general flow of traffic. In case of unusual traffic flow, we can coordinate with weather stations to know if there is waterlogging or flooding.

Water logging detection via remote sensing

Remote sensing typically means the capture of data from satellites of the earth below.

Using Optical imagery

We can log the changes in water extent for a given area using optical imagery. This is an example of two images of Bangalore Airport, taken by the Sentinel 2 Satellite.

Well, both the images look very similar but you can observe muddy water around the airport. Let us have a look at the difference in water logging between these two dates:

The major drawback is clouds and adverse weather conditions. The optical imagery is dependent on clear weather which is rare during floods and heavy rainfall.

Water logging detection using SAR Remote sensing

Synthetic Aperture Radar (SAR) produces energy pulses and records the amount of energy reflected after interacting with the surface. Since each object absorbs a certain amount of energy, we can differentiate between objects and surface conditions.

Steps needed for flood detection using SAR Data:

1. We need to query Sentinel-1 GRD data for the desired area of interest and VV (Vertical Transmit - Vertical Receive) polarization for a set of selected dates at a 10-meter resolution
2. Then capture before and after images for the desired area of interest
3. Make a note of the difference between the two dates
4. Apply binary segmentation using threshold (1.25 -based on analysis ) to identify wet areas( In this step, we use a simple change detection approach, the post-flood visualization is compared with the pre-flood visualization. The bright pixels show a high change, and the dark pixels towards a little change, A threshold of 1.25 (based on analysis) is applied to assign 1 to all values greater than 1.25 and 0 to all values less than 1.25. This gives out the flooded regions)
5. Mask out existing water bodies to only show excess flooded water using historical surface water data.
6. Mask out areas with more than a five percent slope using a Digital Elevation Model
7. Convert raster tiles to vector tiles to get flooded polygons - These are closed areas represented on a map

This can be quickly done using the script produced by UN-SPIDER and running it over the Google earth engine. Complete documentation here

SAR has a unique ability to penetrate atmospheric conditions, providing clear visibility in cloud-covered areas day or night.

Given below are reference images of SAR visualization of the Bangalore airport region.

If we look closely and try to highlight the water pockets on the after-flood image, the blue markings on the images below are what we will observe:

Here is a picture of an optical image overlayed with the water pockets shown in blue obtained from the SAR backscatter thresholding

Image produced by: UN-SPIDER December 2019 via Google Earth Engine

Change detection using Deep learning (DL)

Training the model with data from SAR makes it more accurate to judge the water logging. SAR data analysis needs domain expertise. However, DL eases this requirement by offering a trained predictive model.

A dataset covering different geographies across the globe would make a good model, as it would have trillions of trainable parameters which are required for deep learning.

Use of both Remote sensing and telemetry data

While deep learning and remote sensing are exciting, they have significant drawbacks. The biggest challenge is that it is not in real-time. The data available can range from hours to days from the current time.

Here is an example of Flood Detection done using SAR Remote sensing combined with the telemetry data of the cabs between 8 AM - 9 AM IST on 5th September.

A total of 300 rides were completed during the given time interval in the area above. You can observe the speed using the given index (red is slow, blue is fast). The rides tend to slow down in areas near water logging (indicated by blue polygons)

Here is the visualization of a single ride for better understanding:

This is how data from remote sensing and telemetry information can be combined to provide the best route for your journey, avoiding all the water-logged/ restricted areas.

What Next?

Whats most imperative in customer facing industries is that every action needs to have a positive impact on customer experience. For our GIS enthusiasts, providing you with the best route for your journey without the experience of dragging your vehicle through muddy water.

Our experts continuously work on challenges like these in real-time to improve your experience. We are applying techniques and methods to enhance our routing in dynamic situations such as the Bangalore floods to provide you with the best mobility experience.

If you have some feedback or have found this blog post of your interest, Do Connect with Us!