TLDR: This research introduces a new method for accurately identifying tree points in complex urban environments using multispectral LiDAR data and advanced deep learning models. It highlights the efficiency of the Superpoint Transformer (SPT) model and the significant improvement in accuracy when incorporating spectral information like pseudo Normalized Difference Vegetation Index (pNDVI), which reduced tree detection error rates by over 10%. The study also provides a new public dataset for urban tree research.
Urban areas worldwide are facing increasing environmental challenges, from rising temperatures to air pollution. Trees, as a vital part of urban green infrastructure, play a crucial role in mitigating these issues and enhancing the well-being of city residents. Therefore, accurately monitoring and managing urban trees is more important than ever for effective city planning and supporting greening policies.
Traditional methods for mapping trees often struggle with the complex and varied environments found in cities. However, advancements in Light Detection and Ranging (LiDAR) technology are offering new solutions. Specifically, multispectral (MS) LiDAR systems are proving to be a game-changer. Unlike older single-channel LiDAR, MS-LiDAR captures not only the 3D spatial information of objects but also their spectral properties, providing a richer dataset for analysis.
A New Approach to Tree Point Extraction
A recent study, titled Multispectral LiDAR data for extracting tree points in urban and suburban areas, by Narges Takhtkeshha, Gabriele Mazzacca, Fabio Remondino, Juha Hyyppä, and Gottfried Mandlburger, explores the potential of MS-LiDAR combined with deep learning (DL) models to precisely identify tree points in urban and suburban settings. The core idea is to use binary semantic segmentation – essentially, teaching a computer to distinguish between ‘tree’ and ‘non-tree’ points within a complex point cloud.
The researchers evaluated three state-of-the-art transformer-based deep learning models: Superpoint Transformer (SPT), Point Transformer V3 (PTv3), and Point Transformer V1 (PTv1). These models are designed to process large 3D datasets efficiently and accurately. SPT, for instance, is particularly noted for its time efficiency, achieving high accuracy by processing ‘superpoints’ rather than individual points, while PTv3 focuses on simplicity and scalability through various serialization patterns.
The Loosdorf-tree Dataset
To conduct their research, the team utilized a unique dataset collected over the cities of Loosdorf and Melk in Lower Austria using a RIEGL VQ-1560i-DW multispectral airborne laser scanner. This system operates at green (532 nm) and near-infrared (1064 nm) wavelengths. A significant contribution of this study is the creation and public availability of the Loosdorf-tree dataset, which includes nearly 50 million manually annotated points classified as either tree or non-tree.
Key Findings and Performance
The study’s results highlight the remarkable effectiveness of this integrated approach. The Superpoint Transformer (SPT) model emerged as the top performer, achieving a mean intersection over union (mIoU) of 85.28% for tree detection. This indicates a high level of accuracy in correctly identifying tree points.
A crucial finding was the significant impact of incorporating spectral information, particularly the pseudo Normalized Difference Vegetation Index (pNDVI), which is derived from the green and near-infrared reflectance values. When pNDVI was used alongside spatial data, the error rate for detecting trees with normalized heights above 2 meters was reduced by an impressive 10.61 percentage points compared to using spatial information alone. This underscores the power of MS-LiDAR’s spectral capabilities in differentiating trees from other urban structures like facades, cables, or electric towers, which often have similar geometries.
Furthermore, the research also looked at the computational efficiency of the models. SPT not only delivered the highest accuracy but also proved to be the most time-efficient, requiring significantly less processing time compared to PTv1 and PTv3, making it highly practical for large-scale applications.
Also Read:
- RoofSeg: Advancing 3D Building Reconstruction with Precise Roof Plane Segmentation
- Automated Landfill Detection: A Deep Learning Study Leverages Aerial Imagery for Environmental Monitoring
Implications for Urban Management
This research offers a robust, single-data-source solution for accurately mapping urban trees. The ability to automatically and precisely extract tree points is a critical first step for detailed urban tree inventories, which in turn supports better city management and environmental planning. The findings suggest that leveraging multispectral LiDAR data, especially with advanced deep learning models like SPT and the inclusion of spectral indices like pNDVI, can greatly enhance our capacity to monitor and protect urban green spaces. Future improvements could involve using MS-LiDAR systems with more spectral channels and incorporating incidence angle corrections for even greater accuracy.
