Creation of geo-specific 3D environments for training and simulation require a lot of information along with Electro-Optical (EO) imagery. Acquiring vectors of different object classes along with attributes for each vector is a fairly labor-intensive process. Another important component for making the 3D environment geo-specific is the depth information obtained from Digital Surface Models (DSM) which is expensive, difficult to acquire, and might be noisy. This paper discusses how Deep Learning (DL) based techniques can be used for the extraction of attributed vectors of different object classes from EO imagery and eventually create geo-specific 3D synthetic environments without DSM data.
The contribution of this work is twofold: firstly, multi-level Deep Learning techniques are used for the extraction of building footprints and attributes (e.g., roof type) for each extracted building. Using extracted and derived features (area, shape, etc.), the building heights are estimated which alleviates the requirement of acquiring expensive and difficult to procure DSM data.
Secondly, the problem of creating huge training datasets required to train Deep Learning models is addressed by using synthetic data generated using Presagis software, to solve the problem of roof type classification. A performance-based comparative analysis of classification techniques on synthetic data with other state-of-the-art techniques like few-shot classification is performed to provide insights on how synthetic/hybrid datasets can be used when labeled training datasets are not available.
Finally, a qualitative comparison of two 3D models is performed where the models are created using Velocity (Presagis’ workflow automation for the creation of 3D synthetic environments). The two models are created using the same EO imagery and attributes but acquired differently (one manually and the other with AI-extracted attributes), showing that the model with AI-inferred attributes is very close to geo-specific standards, but does not require labor-intensive manual attribution or collection of expensive DSM data.