TreeScope: An Agricultural Robotics Dataset for LiDAR-Based Mapping of Trees in Forests and Orchards

Xu Liu, Derek Cheng, Fernando Cladera Ojeda, Ankit Prabhu, Xu Liu 0007, Alan Zhu, Patrick Corey Green, Reza Ehsani, Pratik Chaudhari, Vijay Kumar

TL;DR

This paper introduces TreeScope, the first comprehensive LiDAR dataset specifically designed for agricultural robotics, enabling precise mapping and diameter estimation of trees in forests and orchards. By providing high-resolution data from diverse robotic platforms and meticulously annotated ground truth, TreeScope aims to accelerate research in precision agriculture and forestry, overcoming the limitations of manual methods and existing datasets.

What Problem Does This Paper Address?

Accurate counting and diameter measurement of trees are crucial for forestry and agricultural management, informing decisions about biomass estimation, carbon sequestration, fire risk, and timber yield. Traditional methods, relying on manual tools like calipers, are labor-intensive, time-consuming, and prone to errors. While robotics offers a promising solution for non-destructive and efficient measurements, the field has been hampered by a lack of publicly available, high-quality datasets specifically tailored for agricultural and forestry applications. Existing datasets are often limited to urban environments or lack the detailed ground truth necessary for robust algorithm development and evaluation. This paper addresses this gap by introducing TreeScope, a novel dataset designed to facilitate the development of advanced robotic perception and mapping algorithms for trees.

Methods

The TreeScope dataset was collected using two primary robotic platforms: a Falcon 4 UAV equipped with an Ouster OS1-64 LiDAR for under-canopy Unmanned Aerial Vehicle Laser Scanning (ULS) and a versatile mobile sensor platform that can be mounted on vehicles, carts, or carried by humans, utilizing an Ouster OS0-128 LiDAR and an Intel RealSense D435 RGBD camera for Mobile Laser Scanning (MLS).

For semantic segmentation, the authors evaluated state-of-the-art networks like RangeNet++, SqueezeSegV2, and SqueezeSegV3, comparing their performance on TreeScope data.

For tree diameter estimation, two novel methods were developed and evaluated:

  • Density-based clustering ring extraction (DBCRE): This unsupervised method uses DBSCAN clustering on accumulated point clouds to segment tree trunks. It samples control points on individual tree clusters, builds a k-d tree for efficient nearest neighbor searches, and identifies ring points corresponding to the tree trunk's curvature to calculate diameter.
  • Semantic lidar odometry and mapping (SLOAM): This real-time system models trees and ground points as cylinder and plane landmarks, respectively. It estimates pose transformations by associating landmarks across LiDAR sweeps to register point clouds and uses geometric least-squares optimization to extract cylinder parameters, from which tree diameter is derived.

Datasets

The core contribution of this paper is the TreeScope v1.0 dataset. It comprises:

  • High-resolution LiDAR data (ULS and MLS) collected from various forest and orchard environments in Virginia, New Jersey, and California.
  • Manually annotated semantic labels for tree stems, ground points, and miscellaneous objects.
  • Ground-truth field measurements of tree diameters at breast height (DBH), and in some cases, tree heights and full diameter profiles.
  • ROS bags containing raw sensor data (LiDAR, IMU, GPS, RGBD) and processed data (LiDAR-inertial odometry, velocity-corrected point clouds).

TreeScope is unique in providing this data from robotic platforms in agricultural and forestry settings, with detailed annotations and measurements, making it a valuable benchmark for semantic segmentation and diameter estimation tasks.

Key Insights and Ideas

  • Robotics is key to unlocking precision agriculture and forestry: The paper strongly advocates for robotics as a transformative technology for data collection in these fields, offering efficiency and accuracy beyond manual methods.
  • Under-canopy LiDAR is crucial but challenging: Collecting high-quality LiDAR data beneath dense forest canopies presents significant technical hurdles, which TreeScope aims to address.
  • The need for specialized datasets: Existing datasets are often insufficient for agricultural robotics due to their focus on urban environments or lack of relevant ground truth. TreeScope fills this critical gap.
  • Diameter estimation is more than just DBH: The inclusion of full diameter profiles for some trees highlights the potential for more detailed biomass estimation and analysis.
  • Trade-offs in segmentation algorithms: The evaluation shows that while more complex models like RangeNet++ offer higher accuracy, simpler models like SqueezeSegV2 provide faster inference, which is critical for real-time applications.

Key Results

  • Semantic Segmentation: RangeNet++ achieved better overall accuracy (OA) and mean Intersection over Union (IoU) metrics compared to SqueezeSegV3 on TreeScope datasets. However, SqueezeSegV2 demonstrated significantly faster inference speeds, making it a viable option for real-time applications.
  • Diameter Estimation: The novel DBCRE method consistently outperformed the SLOAM method across all datasets, achieving lower RMSE values for diameter estimation. For instance, on the VAT-0723U dataset, DBCRE had an RMSE of 3.8 cm (12.3%), while SLOAM had an RMSE of 9.8 cm (31.8%).
  • Ground Truth Reliability: The paper reports an RMSE of 0.31 cm (1.5% of DBH) for measurements between two human annotators, indicating high reliability in the ground-truth data.

Limitations and Future Work

Limitations: * The current release (v1.0) focuses on tree stems and ground points for semantic segmentation. * The diameter estimation methods have different operational characteristics: DBCRE is an offline method requiring accumulated data, while SLOAM is an online method suitable for real-time applications but requires denser point clouds for optimal accuracy.

Future Work: * Augmenting semantic labels: Future iterations will include canopy and branch semantic labels to enable more robust timber biomass volume estimation. * Incorporating other sensor data: The authors plan to release data from other sensors on their platforms (e.g., RGBD cameras) in future releases. * Expanding dataset coverage: The dataset is intended to grow, potentially including more diverse environments and tree species.

Summary updated

Abstract

Data collection for forestry, timber, and agriculture currently relies on manual techniques which are labor-intensive and time-consuming. We seek to demonstrate that robotics offers improvements over these techniques and accelerate agricultural research, beginning with semantic segmentation and diameter estimation of trees in forests and orchards. We present TreeScope v1.0, the first robotics dataset for precision agriculture and forestry addressing the counting and mapping of trees in forestry and orchards. TreeScope provides LiDAR data from agricultural environments collected with robotics platforms, such as UAV and mobile robot platforms carried by vehicles and human operators. In the first release of this dataset, we provide ground-truth data with over 1,800 manually annotated semantic labels for tree stems and field-measured tree diameters. We share benchmark scripts for these tasks that researchers may use to evaluate the accuracy of their algorithms. Finally, we run our open-source diameter estimation and off-the-shelf semantic segmentation algorithms and share our baseline results.
BibTeX
@inproceedings{liu2024treescope,
  author = {Xu Liu and Derek Cheng and Fernando Cladera Ojeda and Ankit Prabhu and Xu Liu and Alan Zhu and Patrick Corey Green and Reza Ehsani and Pratik Chaudhari and Vijay Kumar},
  title = {{TreeScope: An Agricultural Robotics Dataset for LiDAR-Based Mapping of Trees in Forests and Orchards}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2024},
  doi = {10.1109/icra57147.2024.10611103},
}
Figure