DIGITIZE THE WORLD USING AI
VISICOM BUILDINGS FOOTPRINTS
Today, Artificial Intelligence (AI) has entered the everyday life of many companies in the world, from sorting parcels to writing program code and launching rockets into space.
Visicom team are also in trend, and therefore developed and implemented the technology of recognizing satellite images based on AI, or to be more accurate — machine learning (ML), which analyzes each pixel and determines the class to which a particular pixel belongs and accordingly transfers data from the raster to a vector with corresponding attributes. Below is more about our achievements.
AI BUILDING FOOTPRINTS — PRODUCT FEATURES:
· Automated production (99% of buildings > 25 sq.m matched automatically by the machine learning algorithm)
· Completeness (achieved 100% coverage due to manual post-processing and validation)
· 3m SE90 accuracy
· Rapid production of countrywide building footprints dataset
· Available worldwide
· Based on up-to-date satellite images with 0,3–0,5m resolution
1. Preliminary stage:
For each territory with the related building, the type must be used its own database of etalons and input parameters with the precise coefficients that worked well for a similar area.
|1st output from the recognition|
· Image segmentation process (U-Net convolutional network for satellite/aerial image segmentation)
· Applying the threshold coefficients for identification of buildings layer
· Buildings pattern recognition using deep learning algorithms
· Raster outputs by a neural network
· Validation of the resulting output by engineer
· Calibration and changing the input parameters for the neural network (changing the coefficients, color patterns, geometric patterns, adding more etalons in the database).
· The cycle could be repeated a few times from the start to achieve the desired 3/5m CE90 Planimetric accuracy.
3. Generation of vector building footprints from raster
|Vector building footprints generated from raster|
5. Adding height values into building footprints — Assignment of building heights into building footprints.
6. Manual postprocessing of recognized building footprints: final verification, correction, and checking of the buildings dataset both, geometry and heights.
|Finalized building footprints in vector|
7. Depending on the project’s needs (technical accuracy requirements) the LOD1 building footprints could be updated into LOD2 models representing buildings and roof details with multiple heights.