A multitude of industries and scientific disciplines require highly accurate digital terrain models, making Lidar a logical choice of survey technology. However, the unstructured nature and sheer size of most data corpuses is too large for manual noise filtering, classification, and segmentation.
A new automated classification pipeline was needed that ingests raw point clouds with no known source or sensor characteristics that ultimately creates accurate digital terrain models (DTM’s), classifies individual points (common labels are vegetation, man-made structure, ground, among others), and provides reconstructed LOD2 building models. The resulting ensemble model used a combination of highly predictive local spatial features (e.g. principal components of point groups) and regression values from a U-net model built to identify building footprints. A random forest classifier terminated the machine learning portion where a sequence of postprocessing heuristics was then employed to create all derived products.
The pipeline was successfully implemented, performing with an average F1 score on target classes of over 0.95. The methodology is now the state of the art for this problem set and is currently deployed on a Department of Defense software platform.