Improving point cloud classification of USGS 3DEP lidar data using specific deep learning models

The U.S. Geological Survey (USGS) 3D Elevation Program (3DEP) provides lidar data essential for generating Digital Terrain Models (DTMs). However, current classification standards often lack fine-grained details for features such as buildings, trees, and roads. This study evaluates deep learning (DL) models to refine and enrich 3DEP point cloud classification, moving beyond minimum USGS standards. We assessed four DL architectures, utilizing the 2019 Data Fusion Contest data set for initial training and benchmarking. Performance was determined by comparing model-derived classifications against manually annotated 3DEP tiles. Experimental results are promising, indicating that automated DL approaches can effectively augment 3DEP data sets with higher thematic resolution. Future work can include rigorous benchmarking to evaluate practical scalability and model robustness. Once validated, these DL-based methods can provide high-fidelity, automated re-classification to significantly enhance the utility of The National Map for diverse geospatial applications.