Outline’s proprietary Hyper Dimensional Machine Learning Algorithms are enhancing the production of high accuracy digital elevation models from stereo imagery.
It’s a mouthful, but what does it mean?
Traditionally, each pixel in a stereo image can be used to triangulate and calculate its position in space (x,y,z). The raw output of these calculations can then be used to generate a digital surface model (DSM) of the ground, like a relief map. However, more often that not, what is needed is the actual ground level with the above ground features like vegetation, buildings and vehicles etc, removed. To do this we need to teach a computer how to identify these features electronically and then have them ‘removed’ from the model – to do this we use Machine Learning. So, out of the 4 different bands of light typically captured by our sensors (Red, Green, Blue and Near Infra Red) we can begin to identify patterns that relate to features and thereby start to ‘recognize them’. With advances in computing power and our own R&D we are now able to broaden the scope of recognition from 4 bands to any combination of the 4 bands (>256 combinations), enter the word ‘Hyper‘. For example, we can now train a computer to recognize things like water surfaces, grassy areas, treed areas, buildings, rocks and many other obvious features and then analyse a myriad combination of light bands to find the best correlation and then apply this ‘learning’ across the target area. We can even add variables relating to the DSM such as slope, aspect or roughness to refine the result. The machine (computer) keeps adding to its ‘knowledge’ gained from the ‘learning data’, the results are astounding. Almost by magic, the generation of digital elevation models (DEM – the one without the surface features) is done very cleverly leaving a vastly improved solution from what was previously achievable.
Outline is continuously enhancing this technology and applying it to a range of different image analysis challenges.