Bonus: Point Cloud Classification in ArcGIS Pro with PointCNN#
This workflow uses deep learning to classify LiDAR point clouds using PointCNN in the ArcGIS Pro arcgis.learn module.
Update ArcGIS Pro
Install Deep Learning Libraries for the correct version of ArcGIS Pro: Esri/deep-learning-frameworks
Prepare point cloud data:
Merge all polygon annotations into a single part polygon (using the editing tool Merge after selecting all rows in the attribute table or Select Layer by Attribute).
Use las2las (open source) to filter out ground classified points.
Use lasclip (licensed) with -classify 20, -interior, and -poly with the polygon annotation or use Change LAS Class Codes in ArcGIS Pro.
Create 2 polygons: one to define training (about 60-70%) and validation (30-40%) areas. The polygons should extend outside of the las file area and not overlap each other.
Use lasclip to generate separate input point cloud (training) and validation point cloud.
Run the Prepare Point Cloud Training Data tool (block size = 250 m).
Run the Train Point Cloud Classification Model tool. This step may take a long time depending on the selected maximum number of epochs (default = 25).