If you live in a metropolitan city in India, most of your commute concerns involve issues with mapping road networks and navigation. You’re either struggling with by-passing new metro construction sites that are cropping up every day or finding shorter & quicker routes through the heart of the city or confused about the sudden closure of roads that you previously frequented. Along with this, over time mapping new road networks and updating the golden database of mapping road networks for everyone’s benefit. Not just commuters, but mobility service providers and last-mile delivery. If we look at Bengaluru, a city with an average population of 13.1 mil and rapidly growing, squeezed into 2196 sq. kilometers, the urban geography poses a good challenge for regularly mapping road networks.
Why mapping road networks?
Under our earth observation project, we wanted to enrich the golden database of road networks. This can add value to several industries requiring road network data, not just the end consumer driving on the roads. However, there are many gaps to be closed in road network mapping to make life easier for service providers and the end consumer.
- Limited data available from mobility service providers functioning in select areas in a city
- Unavailability of telemetry data – GPS data from vendors and service providers, in low volume areas
- Negative areas that limit visibility from satellite imagery
- Cost reduction of updating road maps – using satellite data to analyze change detection of any kind can drive up the costs.
- Creating a map that is customizable, more perceptive, and responsive to updates and displays accurate features like building footprints, road networks, toll plazas etc.
There were 2 sources of data we used in our algorithms for mapping road networks
- Telemetry data is collected from service providers and vendors that function within the city – stressing that this is anonymized aggregate data and the only data that is received and worked on is telemetry data
- Using satellite data for 2 dates, the current project is analyzing satellite data between two years. It is also possible to update road n/w by using the latest date imagery to extract roads from our DL model and compare against Golden DB. Any new roads found would be QCed and updated)
There are certain roadblocks that need to be overcome –
When we look at satellite data – there may be cloud cover or tree cover that limits our visibility of roads. In some cases, high rise buildings can obscure road visibility due to the orientation of the satellites there. This can be overcome by using telemetry data from vendors and mobility service providers, that is available for highly frequented roads within the city.
For low-volume areas like residential streets and new infrastructure on the outskirts of cities, satellite imagery will have to be used to update the road networks. Usage of satellite data will be done on a priority basis, studying to rapidly expanding territories first vs those territories with stagnant growth. Therefore, using satellite imagery only where it is absolutely necessary.
We’ve come up with some algorithms to do this and are prototyping them.
The first step is creating and updating a continuous area coverage map (a vector map and a satellite base map) of the city using very high-resolution satellite data. Further, temporal updating is done by prioritizing specific areas and not the entire city. This is enabled by our in-house, course level change detection techniques using freely available satellite data.
Then we used data provided by vendors and service providers. There are service boundaries for service providers so we can work with the data provided within those areas alone. There are new datasets that are available fortnightly, so we’ve used all those data sets in the 2 years were considering, to calculate this change detection. On the backend, there is a road network detector algorithm running, and it will map all the roads available with the selected boundary using satellite imagery. There is also a delta mapper team working, their job is to perform quality control and small course corrections that are required area-wise, to conclude that the changes detected are accurate and move on to the next area.
Using these data sets in combination we were able to discover new roadways that were not previously mapped and serviced. The results were found to be 82% accurate and are as shown below. The changes shown are for infrastructure in general. (Below map holds sensitive info like our service bounds for Bengaluru city. Our competitors may strategize against this, also it contains infrastructure whereabouts in the city. Construction companies, allied hardware, and building material supply folks can benefit from this)
Above we see the polygons representing new roadways that previously were not mapped and serviced. You can see the course level change detection that we’ve carried out. The Road network ground truth (in cyan) and prediction (in green) was obtained from our Deep Learning Model using a sample mosaic around Bengaluru Airport
We’ll have to procure very high-resolution satellite data (It is expensive irrespective of developed regions) for regions that develop faster. These are on priority to be mapped vs the regions in the city that don’t develop at the same pace. (All satellite data comes from vendors, we don’t have our own constellation yet, but this is our moon shot, we’re open to partnering with satellite data vendors). Cost reduction is a big factor here. We’ve currently been able to successfully obtain results on 30% of the cost proposed had we used satellite data for the entire city.
Why can’t we use telemetry data alone – sometimes when you consider mapping road networks in residential and outskirts, there is a coverage issue from vendors and mobility service providers?
In cases of inadequacies in satellite imagery we pull in telemetry data from a recent timeline, using that data in our algorithm – (that we’re prototyping, in due course we will be deploying it) that allows us to map the incompletely mapped roads and further help improving the mapping quality in updating the roads. The chances of quality mapping roads with a combination of these 2 approaches are higher and solving a mobility problem in cities that have incomplete mapping on roads that are traveled frequently. There are 2 algorithms we use – an offline and an online model. The offline model gives us more accurate results but is time-consuming. The online model can be run with real-time data, as and when the mobility service providers are traveling through the cities we can perform mapping, but the problem is it is not very accurate. The way to increase the quality of road mapping is with constant updates on a daily basis. All these daily updates will only be possible with online real-time mapping, and quality control using the offline model as a gold standard.
If you want to extend service boundaries and expand the reach of your services, if you want to open new stores and depots, and if you want to place assets on the ground will need updated road maps. We are on our way to creating our own maps – allowing real-time updates with exact features, which we can achieve side-by-side with our Earth Observation project. The value addition and use cases of spatial technology is immense. Imagine the changes in standards of living, ease of logistics, field force, and last-mile delivery that a well-designed and accurate road map can provide.
Mapping the road ahead
Spatial intelligence can provide businesses the flexibility to explore network expansion opportunities based on valuable location and demographic insights. This approach takes the guesswork out of the equation, so businesses can hone in on profit-making locations and make decisions that are most likely to succeed. Find out how GIS solutions can help you gain leverage in your industry, talk to our experts today!
This project is the proud endeavor of Puneeth Shankar, our Principle Data Scientist!
Don’t forget to follow us on LinkedIn, Facebook, Medium and Twitter!
The post Mapping Road Networks – Acute Change Detection Analysis appeared first on GeoSpoc.