Structure-aware unsupervised transmission corridor structure extraction from 3DEP lidar using graph attention embeddings and reinforcement-learned DBSCAN

Abstract

Transmission corridors are critical components of power delivery infrastructure, and their large-area monitoring increasingly relies on national-scale airborne lidar programs. This study presents an unsupervised, scalable framework for extracting transmission corridor structures (including conductors and supporting towers) from U.S. Geological Survey (USGS) 3D Elevation Program (3DEP) lidar point clouds. The workflow begins with height-above-ground filtering and statistical outlier removal to isolate elevated infrastructure candidates. Local Principal Component Analysis (PCA) is then used to compute fine-scale geometric descriptors. To incorporate spatial context and improve structural separability in feature space, we learn point embeddings using a Graph Attention Network (GAT) trained with a structure-consistency triplet objective, yielding representations tailored for downstream density-based grouping rather than semantic classification. To address the spatial non-stationarity of point density and noise across large scenes, we formulate Density-Based Spatial Clustering of Applications with Noise (DBSCAN) parameter selection as a reinforcement learning problem and employ Proximal Policy Optimization (PPO) to adapt 𝜀 and minPts at the block level based on embedding statistics. The final extraction is produced via block-wise DBSCAN and spatial merging to support efficient processing of multi-million-point clouds. Experiments on USGS 3DEP data over Houston, Texas, show that the proposed method produces spatially coherent candidate corridor structures from unstructured lidar. Quantitative assessment is conducted using OpenStreetMap (OSM)-referenced spatial-agreement metrics under a consistent vector-reference protocol, alongside comparative baselines and ablation studies, to examine the relative contributions of graph embedding learning and RL-based adaptive clustering.