Advanced point cloud completion for urban trees : a novel approach using enhanced SnowflakeNet

Abstract

Accurate assessment of urban tree metrics using Light Detection and Ranging (LiDAR) technology is crucial for sustainable urban forest management. However, point cloud incompleteness poses a significant challenge, primarily due to occlusions inherent in ALS (Airborne Laser Scanning) and MLS (Mobile Laser Scanning) acquisition modes, and secondarily from complex urban features such as buildings, vehicles, and infrastructure that obstruct scanner views. These data gaps can compromise the accuracy of tree parameter estimation. While deep learning methods have shown promise in point cloud completion, existing approaches struggle with the complex characteristics of trees, particularly in dense urban settings. This study aims to address the incomplete tree point cloud challenge by developing an enhanced deep learning framework specifically optimized for tree structure reconstruction from partial LiDAR data. Our method extends the SnowflakeNet architecture by integrating a Global Context Module to capture overall tree morphology and Channel Attention mechanisms to emphasize critical structural features. Comprehensive evaluations were conducted on both ALS and MLS data modalities using consistent metrics across controlled simulation scenarios with varying occlusion patterns. The results demonstrate substantial improvements over the original SnowflakeNet and other benchmark methods. For simulated ALS datasets, our model reduces CD_L2 completion errors by 67.07 %-75.28 % and RMSE by 7.69 %-8.51 % across various masking ratios. Similar performance gains are observed in MLS simulations, with 25.32 %-29.52 % reductions in CD_L2 and 7.69 %-9.33 % reductions in RMSE. The model also exhibits promising cross-regional transferability, completing European tree species despite training primarily on subtropical Hong Kong specimens. These advancements provide urban forest managers with more reliable tree monitoring and assessment tools, contributing to more effective urban forest management practices.