Smartphone-based Circular Plot Sampling for Forest Inventory

Su Sun, Yingjie Chen2025 · DOI

TL;DR

This paper introduces a novel smartphone-based system for efficient and accurate forest inventory. By leveraging a circular video capture protocol and integrating advanced computer vision techniques like monocular depth estimation and dense visual SLAM, the system can measure tree diameter at breast height (DBH) and spatial locations within a plot, offering a cost-effective and accessible alternative to traditional methods.

What Problem Does This Paper Address?

Accurate forest inventory is crucial for timber management, carbon assessment, and climate change mitigation. A common method is circular plot sampling, which involves measuring individual tree attributes like DBH and location. However, current methods are either expensive (e.g., terrestrial LiDAR) or labor-intensive (e.g., manual calipers and compasses), limiting their scalability and accessibility, especially for smaller operations or resource-limited regions. Existing smartphone-based approaches often require tree-by-tree scanning or multi-view reconstruction, which can be inefficient for full plot sampling. This paper aims to bridge this gap by providing a complete, smartphone-only solution for efficient and accurate circular plot forest inventory.

Methods

The core of the system is a pipeline that integrates several computer vision components:

  • Circular Video Acquisition: Users mount a smartphone on a simple stand and walk in a circle around a plot center, capturing a 360° video. This protocol is designed to provide natural loop-closure opportunities for SLAM and a known physical radius for metric scale recovery.
  • Dense Monocular Depth Prior: A pre-trained video depth estimation model (Video Depth Anything) provides dense depth maps for each frame. This acts as a "soft prior" to stabilize SLAM in low-texture forest environments, improving tracking robustness and reconstruction quality without being a hard constraint that could introduce scale or offset errors.
  • Dense Visual SLAM: A DROID-SLAM-based framework is used to estimate camera trajectories and reconstruct a globally consistent 3D point cloud. It incorporates the monocular depth prior and a learned place recognition module for loop closure to suppress drift.
  • Segmentation-Guided Measurement: Semantic segmentation identifies tree stems in the reconstructed point cloud. A projection-based cross-sectional analysis at breast height is then used to estimate DBH, which is robust to depth noise and partial occlusion.
  • Metric Scale Recovery: The system recovers real-world scale by using the known physical radius of the smartphone stand, eliminating the need for external sensors or calibration targets.

Datasets

The system was evaluated on two types of forest environments at Purdue University's Martell Forest:

  • Managed Plantation Plots: Four plots with regular tree arrangements and available ground-truth DBH measurements.
  • Natural Forest Plot: One plot with irregular tree distribution, dense underbrush, and uneven ground, where ground-truth DBH measurements were available but precise tree locations were not.

The paper does not introduce new benchmark datasets but utilizes existing forest environments for evaluation.

Key Insights and Ideas

  • Soft Depth Priors for SLAM: Using a pre-trained depth model as a "soft prior" in SLAM, rather than a hard constraint, is crucial for handling the scale and offset ambiguities inherent in monocular depth estimation, especially in challenging forest scenes. This allows the system to leverage depth information without being overly constrained by its potential inaccuracies.
  • Circular Acquisition for Efficiency and Accuracy: The simple circular video capture protocol is a clever design choice that naturally facilitates loop closure for SLAM, provides a known scale reference, and ensures consistent coverage of tree stems, all while requiring minimal user effort.
  • Smartphone as a Complete Inventory Tool: The demonstration that a standard smartphone, combined with clever software, can achieve accuracy comparable to established methods for DBH estimation and tree localization is a significant step towards democratizing forest inventory tools.

Key Results

  • DBH Estimation Accuracy:
    • Managed forest plots: Mean Absolute Error (MAE) of 1.51 cm, Mean Absolute Relative Error (MARE) of 3.98%.
  • Natural forest plot: MAE of 2.30 cm, MARE of 5.69%.
  • Repeatability: Cross-video consistency analysis showed stable and reproducible tree localization and DBH measurements, even when acquisition started from different directions or positions.
  • Processing Time: End-to-end processing for a typical one-minute video takes approximately 5 minutes on an NVIDIA RTX 4090 GPU, significantly faster than traditional SfM-based photogrammetry.
  • Accessibility: The system requires no specialized hardware beyond a consumer smartphone and a simple stand, making it highly accessible.

Limitations and Future Work

  • Limitations:
    • The system's accuracy can be slightly affected by partial occlusion from understory vegetation, particularly in natural forest settings.
  • While robust, the system's performance on very dense stands or in tropical/leaf-on conditions requires further investigation.
  • The current method estimates DBH but not tree height, which limits direct calculation of volume and biomass.
  • Future Work:
    • Estimating tree height to derive volume and biomass.
  • Integrating species classification directly from the captured frames.
  • Further domain adaptation of the depth prior for forest-specific data to potentially reduce per-stem variance.
  • Testing in more diverse forest types and conditions.

Summary updated

Abstract

Circular sample plots are key to forest inventory, but current methods are often costly or labor-intensive. We propose a simple, smartphone-based approach to measure tree DBH and locations within a circular plot. It offers accurate results and is accessible to both researchers and local forest managers.
BibTeX
@misc{sun2025smartphone,
  author = {Su Sun and Yingjie Chen},
  title = {{Smartphone-based Circular Plot Sampling for Forest Inventory}},
  year = {2025},
  doi = {10.5703/1288284317874},
}
Figure