<img height="1" width="1" style="display:none;" alt="" src="https://dc.ads.linkedin.com/collect/?pid=371178&amp;fmt=gif">
Sivakumaran S
Sivakumaran S

Detecting New Construction Sites

Through my own experiences, I know the frustrations of manually undertaking certain tech tasks. Computers now have the intelligence to perform these tasks for us, which is proving very useful within the realm of satellite imagery.


Large-scale problems are still being tackled manually, particularly within the construction industry.  For example, tracking the progress of large-scale infrastructure projects within a city. In a fast-growing city, the number of projects you may wish to track at any one time could realistically cross well into the 100s (this may include your own or competitor activity).


For example, in Shenzhen, China, there are at least 60 ongoing construction projects of buildings over 150m tall and the city covers an area of around 2000 km2. Given a conservative estimate of about 300 projects, an analyst would require about one hour to evaluate each project on site = 300 hours evaluation time and could also involve an additional 100 hours or so of driving. Let's break this down....working 10 hours a day, it would take one analyst 40 working days to cover the whole city. Even with a six-day week, it would take one and half months to cover the list of construction sites. Even with employing more analysts to tackle the job, it's a huge task (never mind the extra expenses.) What if we could teach a computer to detect these changes using satellite imagery? Using satellite images has the advantages of frequency, consistency, and alignment.


Many similar problems exist in the rest of world, some of which have been solved at scale using satellite imagery. At Bird.i we have solved the issue of detecting new construction sites that emerge in a region. Bird.i provides access to the world’s most up to date, high-quality images from multiple satellite imagery vendors.  Above this, we have built an intelligence service that takes an image of a region and compares it to an older image (say from 6 months ago) and determines if a construction site is present where the land was empty, at the same place, six months ago. We achieve this by defining various stages of construction.



Open land

Empty ground, no evidence of any structure, building, vehicles or cranes


Preparation Stage

Still predominantly empty but there are signs of the start of construction. For example: the ground is dug up, presence of cranes, building materials piled up in the area.


Construction Stage

Construction work is in progress and the image contains evidence of a clear structural form (This could be a building, an interchange, or a dam)


Completed Stage

All construction work is completed. The absence of cranes and incomplete structural forms are typical evidence of this stage.


 automatic detection of construction sitesautomatic detection of construction sites

 automatic detection of construction sitesautomatic detection of construction sites


We have created a model using a training dataset of these four categories and then used that model, with some additional math, to determine whether a new construction site has appeared in an area.


With this technology, there are, of course, some challenges. Satellite imagery over urban areas does not always neatly fall into these four categories. For example, open land and a completed building could appear in the same image. Also, an image could be on a borderline between categories 1 and 2, or between 3 and 4. Marking the ground truth for a dataset like this is tricky. With this model in the pipeline, we can obtain two predictions: one from an image of a location taken 6 months ago and another from an image of the exact same location taken as recently as possible. When this pair of predictions score a 1 in the first stage and a 2 in the second stage, it clearly indicates that a new construction site has appeared in the location.


Now, using Bird.i's tool, a search of the entire area of Shenzhen can be taken in about 2 days. So, a task that takes a human one and a half months can be reduced to just 2 days! The generated report shows the sites listed, along with a measure of confidence in the prediction, which can assist greatly with planning. 


Human intelligence should not be wasted on tasks that can be done by computers. In the past, only tasks that could be automated, purely on the basis of deterministic rules, were completed by computers. But with the advent of Machine Learning, tricky problems like these can be solved using automation, which can be scalable. Discover the types of problems that Bird.i can help to solve using the world’s satellite resources and Machine Learning with our Intelligence Service.


Subscribe to Get the Latest Insights