Automated structural parameter estimation in planted and natural forests using unmanned aerial vehicles and vision foundation models

Abstract

Effective monitoring of planted and natural forests is critical for assessing stand development and ensuring long-term ecological and economic success. However, traditional field-based inventories are labor-intensive and costly, limiting their applicability across large or inaccessible areas. Although unmanned aerial vehicles (UAVs) photogrammetry provides a scalable alternative, accurately delineating individual tree crowns in diverse and complex stand structures remains a significant challenge. We introduce and validate a cost-effective framework for automated individual tree inventory by integrating high-resolution imagery from a consumer-grade UAV with a two-stage deep learning pipeline. The framework employs a YOLO-based object detection model to localize individual trees, subsequently using these detections to prompt the Segment Anything Model 2 for precise, zero-shot tree crown segmentation. The framework was validated across diverse subtropical forests, including orchards, plantations, and natural forests. The deep learning models achieved high accuracy in detection (mAP50 = 0.881) and segmentation (mIoU = 0.854). The framework demonstrated robust performance in estimating horizontal structural parameters, especially in managed stands (R2=0.83 for orchards; R2>0.75 for plantations), and robust accuracy for tree height (R2>0.59). This fusion of consumer UAVs and foundation models offers a powerful, scalable tool for individual-tree-level inventory, with significant implications for precision silviculture and monitoring in subtropical forests.

Plant phenotyping relevance

UAV画像と深層学習による個体樹冠 segmentation・樹木位置検出・樹高および構造パラメータ推定が研究の中心で、森林植物の形態形質を技術的に開発・検証している。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.