Models reference guide

Buildings

Extracting of rooftops of buildings from imagery of high resolution. High performance deep learning model is trained to detect the buildings roofs. It is trained primarily on the territory of Russia, including cities, small towns and rural territories and performs great on all of them.

However, the results for the territories with significantly different landscape and urbanization pattern (like dense building blocks, skyscrapers or mountain areas) can be not so impressive.

Note: The building candidates with area less than 25 sq.m. are removed to avoid clutter

The model does not extract the footprints directly, because they are not clearly visible in the images, but we can obtain them, just like human cartographers, by moving the roof to the bottom of the wall (see Additional options).

Additional options:

Classification by types of buildings – typology of buildings is represented by the main classes (see reference).
Building heights - building height estimation by the length of the shadow and the visible part of the wall. This option also provides building footprints instead of roofs. See our article for some details on the technology.
Simplification - the algorithm corrects the irregularities of the contours of our model. The irregular geometries are replaced with rectangles, circles or arbitary polygons with 90 degree angles, which fits better to the original shape. Also the corrected buildings are rotated to align with the nearest roads. This option produces much more map-friendly shapes which look better and are easier to edit, but some shape accuracy can be lost. See our blog post for more information and some visuals.
Merge with OSM - some of the areas have great coverage of OpenStreetMap data, and if you prefer human-annotated data, you can select this option. In this case, we check for each building whether it has a good corresponding object in OSM (Jaccard index more than 0.7) and if there is one, we replace our result with OSM contour. This makes the result not based on the image, so the buildings can be shifted from actual positions, and some changes that have occurred after OSM mapping may be lost.

Processing results samples

A sample of processing result with different options for Prague, Chech Republic.

Processing result of buildings model — Result without postprocessing: irregular building shapes, but best fit to the actual rooftop contour seen in the image.

Buildings (Aerial imagery)

This model is specifically designed to make use of very high resolution aerial imagery (10 cm per pixel) for extraction of small buildings and structures. It is best suited for rural and suburban residential areas.

We do not recommend using this model in areas with high urban residential or commercial buildings. Use Buildings model instead, even for aerial imagery.

Processing results samples

Processing example rural residential area with a Building model (Aerial photo)

Standard model for buildings segmentation, with polygon simplification

Objects that have been detected in an aerial image by the Building (Aerial imagery) model as opposed to the standard model Buildings.

High-density housing

Our “high-density housing” AI model is designed for areas with terraced or otherwise densely built buildings, common in the Middle East, parts of Africa, etc. This model, just like the regular building model, detects the building roofs.

Firstly, the building blocks are segmented as a whole, and then each block is divided into individual houses with rectangular grid or Voronoi diagram, based on the detected individual roof markers.

Processing results samples

Processing result sample for dense urban development area (Tunisia, Africa):

Processing result of high-density housing model — Standard model for buildings segmentation, with polygon simplification

Forest

Forest Segmentation. The model is trained on high-resolution data (2m ) for central and boreal Russia in summer period.

The result includes all areas covered with tree and shrub vegetation, including sparse forest and shrublands. Model resolution does not allow it to detect individual trees and narrow tree lines, and draw a strict border for the forested areas, but suits well for building a general analytical map.

The model is robust to region change, and performs well not only for Russia, but also in other countries and continents. The image should be taken in active vegetation period, because leafless trees or vegetation covered with snow are not target class. Postprocessing:

Additionally we use models for density and height estimation, dividing the forested area into the following classes:

Shrubs lower than 4 meters high(sparse or dense);
Forest from 4 to 10 meters high, sparse;
Forest from 4 to 10 meters high, dense;
Forest more than 10 meters high, sparse;
Forest more than 10 meters high, dense.

This model can be used as a decision support for the forest growth clearing.

This postprocessing is available for processings 50 sq.km and more.

Processing results samples

Processing result of forest model — Processing results for central Russia (Tatarstan)

Constructions

This model outlines the areas in the satellite image that contain construction sites and buildings under construction. Very high resolution imagery (0.3-0.5 m) for the territory of Russia is used. See our blog post on the model development and motivation.

Processing results samples

Processing result of construction model — Processing result sample for a rapidly developing area with a lot of construction sites.

Agriculture fields

Model for fields segmentation allows to detect the agricultural fields and delineate the nearby fields from each other, if there is a visual boundary (forest line, road, different crop stage). The model is trained on the high resolution data (1-1.2 m), primarily for Europe, Russia. It performs better with larger fields with active vegetation. Smaller and terrace fields (typical for Asia) are delineated not so good. Fields without vegetation, especially in winter period, are not target class.

Processing results samples

Processing result of agriculture fields model — Processing result sample for Europe (Belgium)

Roads

Model for road segmentation in high resolution imagery (0.3 - 0.5 m)

The model is trained primarily for rural and suburban areas. Multi-task learning is applied in order to improve the road mask connectivity, especially in the spots obscured by trees or buildings. Best suited for areas with low urbanization, and can fail in cities where wide roads with sidewalks and complex crossroads are present. We extract the road central line in order to decrease the clutter and optimize the extracted road network, and then the road lines are inflated back to polygonal object.

In version 1.1 we added the road graph postprocessing:

geometry simplification;
merging of the gaps;
removal of double edges;
removal of detached and too short segments;

Processing results samples

Processing result of roads model — Optimal conditions for the model: rural/suburban territory, Russia