— new Roads model

We already posted on our blog about the work in progress on the new version of Roads model. Today we are happy to announce that the model is available in our platform! For now it is in experimental beta and the credits cost for its use is very low. Hence it is a better chance to test it, so go to and try it by yourself.

What’s new?

We have analyzed users processings and business requests from our partners and found out that there is a significant demand for mapping of roads in the areas far away from big cities. This is not a surprise, since metropolises are usually covered in public map services like OpenStreetMaps, Google Maps, etc.

So, we decided to train our model to better perform in areas where there is a lack of publicly available mapping data.
Compared to the previous versions of the model for road network extraction we can highlight three main improvements:

  1. The new model can significantly better detect narrow unpaved roads, which are not well represented on many maps.
  2. It can handle various types of landscape patterns including rural areas, forests, northern regions with snow and swamps, agricultural fields, etc.
  3. In addition, the new model greater preserves the topological correctness of the road network by solving the occlusion problem causing by trees, overpasses and over obstacles in the imagery. Which means that the model’s intended to fill in the gaps in the output mask towards the production of the linked road graph — therefore less time for edits is required.

How it works?

While working on the model improvement we faced two main problems that come from the input data complexity and cannot be solved using only classical approaches:

  1. Most unpaved roads have very vague borders, which leads to fuzzy edges of a segmented mask.
  2. There are many occlusions, caused mainly by trees canopies that overlap the road surface in the satellite image, producing undesirable gaps in the output.

Both these difficulties make accurate roads annotation a complicated task even for a professional cartographer, and the machine is also confused. To handle these difficulties caused by specifics of the data we developed a novel approach that is based on the multi-task learning strategy. Basically, we split the task of road extraction into three correlated sub-tasks:

  1. Road surface extraction — it is a classical semantic segmentation problem and most of the other approaches include only this step.
  2. Road boundaries extraction is primarily aimed at refining the segmented surface.
  3. Road center line extraction makes it is easier for the model to understand which road parts should be connected or not.

The right combination of listed training tasks helped us to achieve more solid results in terms of accuracy (precise localization) and topological correctness of the model’s output.

More examples?

We recommend you to test model by yourself at You can also check some of these examples generated with Mapflow:

  1. Roads in forest
  2. Roads in fields
  3. Industrial zone
  4. Suburban zone

What’s next?

At GeoAlert we’re always working on models performance and user experience enhancement. We are permanently enriching our training datasets with new data samples based on the users feedback. We are looking forward to sharing more details about the development of the new postprocessing algorithm for roads linearization, so it’s gonna be even easier and fast enough to edit and integrate into your workflows.

Stay tuned!


  • — The Geoalert platform for AI-mapping

We apply Machine learning to automated analysis over Earth observation data