邁向城市森林數位孿生，以近距離雷射掃描對大規模都市樹木進行三維建模與參數化Towards the digital twin of urban forest: 3D modeling and parameterization of large-scale urban trees from close-range laser scanning
Chen Chen, Haodong Wang, Duanchu Wang, Di Wang｜International Journal of Applied Earth Observation and Geoinformation 127: 103695｜DOI: 10.1016/j.jag.2024.103695
狀態：AI_DRAFT_FROM_REVIEW｜分級：A｜閱讀深度：FULL_TEXT_CHECKED｜Jacky 審核：False
森林數位孿生產業應用
urban forestdigital twin3D tree modelingpoint cloudMLSULSspace colonization algorithmL-systemLiDARChinaXi'an
專討核心文獻定位
[32]
Ch7 · 數位孿生 ★
新增
以消費級 MLS+ULS 點雲、改良空間殖民演算法與 L-system 壓縮，直接重建大規模都市樹木並走向城市森林數位孿生
使用警示
本頁是文獻知識庫卡片，不等於可直接引用的最終查核稿。只有狀態升級為 CITABLE 後，才可直接進入論文引用候選。
為什麼納入這篇
This paper is a concrete application-layer source showing how a forest digital twin can be built bottom-up from individual tree 3D reconstruction at city scale, using low-cost mobile and UAV laser scanning rather than expensive TLS. It complements the conceptual FDT framework papers by demonstrating an actual large-scale pipeline with quantitative accuracy.
結構式摘要｜中英文對照
| 研究問題 | 如何用低成本、可快速大範圍取得的點雲，自動重建大規模都市樹木的精細三維模型並輕量化儲存，以支援城市森林數位孿生的建構？ How can low-cost, rapidly acquired point clouds be used to automatically reconstruct fine-grained 3D models of large-scale urban trees and store them in a lightweight form, in order to support the construction of an urban forest digital twin? |
|---|---|
| 資料來源 | 原文 Section 3 確認：研究場域為西安電子科技大學校園，範圍 260 m × 140 m，主要樹種為梧桐。MLS 使用手持式 FJD Trion S1 掃描器（掃描速度 320,000 點/秒、相對精度 2 cm，僅記錄座標無 RGB）；ULS 使用 DJI L1 無人機載 LiDAR（測距精度 3 cm、240,000 點/秒、飛行高度 120 m），可自動融合 RGB 形成彩色點雲。兩種資料於 2023 年 6 月採集，先手動粗對位再以 ICP 精對位，對位精度約 5.61 cm。 Section 3 confirms the study site is the Xidian University campus, 260 m by 140 m, with Chinese parasol as the dominant species. MLS used the handheld FJD Trion S1 (320,000 points/s, 2 cm relative accuracy, coordinates only, no RGB); ULS used the DJI L1 UAV LiDAR (3 cm range accuracy, 240,000 points/s, 120 m flight altitude) with automatic RGB fusion to a colored point cloud. Both datasets were collected in June 2023, coarse-aligned manually then refined by ICP to about 5.61 cm accuracy. |
| 方法 | 原文 Section 4 確認三大步驟管線：(1) 樹點萃取，比較 RGB 閾值法（搭配 GMM、Watershed、LeWoS 葉木分類）與深度學習 RNConv 網路（用 SensatUrban 預訓練模型）兩種策略，再人工去除非樹點；(2) 改良空間殖民演算法直接建模，提出 killing distance 的自適應錐削策略、以群聚法定位多棵樹根節點，無需單木分割即可同時建模多棵樹；(3) 以參數化 L-system 把樹骨架壓縮為抽象表達式。最後用通用圓柱與 allometric 半徑關係（West et al. 1999）構成幾何模型並補上葉片（僅供視覺化）。 Section 4 confirms a three-stage pipeline: (1) tree point extraction comparing an RGB-thresholding strategy (with GMM, Watershed, and LeWoS leaf-wood classification) against a deep-learning RNConv network (SensatUrban pre-trained model), followed by manual cleanup; (2) an improved space colonization algorithm for direct modeling, introducing an adaptive tapering of the killing distance and a clustering-based root-locating method so multiple trees are modeled simultaneously without single-tree segmentation; (3) parametric L-system compression of the tree skeleton into abstract expressions. Skeletons are turned into geometry with generalized cylinders and allometric radius relations (West et al. 1999), with leaves added for visualization only. |
| 主要結果 | 原文 Section 5 確認：在 36,400 平方公尺的場域中成功建模 490 棵樹，人工計數共 512 棵，成功率 95.7%（摘要寫成 96%）。樹點萃取 RGB 法整體精度 80.0%、偵測率 95.9%，深度學習法整體精度 86.8%、偵測率 82.5%。重建模型估算 DBH 的 RMSE 為 5.5 cm（13.0%）、樹高 RMSE 為 2.7 m（13.2%）。L-system 抽象表達式檔案僅為原始骨架的約 12%；模型平均檔案從 321 KB 降至 23 KB，壓縮率 92.8%。資料採集僅花 2-3 小時，設備為消費級、約 1 萬至 2 萬美元。 Section 5 confirms 490 trees were successfully modeled over the 36,400 m2 site against 512 manually counted, a 95.7% success rate (the abstract rounds to 96%). RGB extraction reached 80.0% overall accuracy and 95.9% detection; deep learning reached 86.8% overall accuracy and 82.5% detection. DBH was estimated with an RMSE of 5.5 cm (13.0%) and tree height with 2.7 m (13.2%). The abstract L-system expression file was only about 12% of the original skeleton; mean model size dropped from 321 KB to 23 KB, a 92.8% compression rate. Acquisition took only 2-3 hours with consumer-grade devices costing roughly ten to twenty thousand US dollars. |
| 限制 | 原文 Section 6.5 確認：樹點萃取精度仍有改善空間，且並非所有 LiDAR 都能取得 RGB；本研究僅建模落葉樹，針葉樹密集針葉與冠內結構對雷射掃描較難捕捉；葉片重建僅供視覺化、非真實重建；未來應結合輻射傳輸模型支援都市樹木生理理解。 Section 6.5 confirms tree-extraction accuracy still has room to improve, and not all LiDAR captures RGB. Only deciduous trees were modeled; coniferous trees with clumped needles and intra-crown structure are harder to capture by laser scanning. Leaf reconstruction is for visualization only, not real reconstruction. Future work should couple radiative transfer models to support physiological understanding of urban trees. |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| Large-scale urban trees can be reconstructed directly from consumer-grade MLS+ULS point clouds without single-tree segmentation, reaching a 95.7% modeling success rate over 490 of 512 trees in a 36,400 m2 area. | Section 5.2 and Conclusion: 490 trees modeled vs 512 manually counted = 95.7%; the improved space colonization algorithm with root-locating avoids single-tree segmentation. | checked_against_full_text |
| Reconstructed models retain forestry-grade geometric accuracy: DBH RMSE 5.5 cm (13.0%) and tree height RMSE 2.7 m (13.2%). | Section 5.2: fifty randomly selected field-measured trees; DBH RMSE 5.5 cm (13.0%), height RMSE 2.7 m (13.2%), Fig. 16. | checked_against_full_text |
| Parametric L-system compression cuts storage drastically while preserving geometry, reducing mean file size from 321 KB to 23 KB (92.8% compression). | Section 5.4 and Fig. 20: mean file size 321 KB to 23 KB, compression rate 92.8%; recovered height/crown radius RMSE 1.6 m (10.2%) and 1.7 m (16.8%) over 490 trees. | checked_against_full_text |
Key Figures and Tables
公開網站原則：未確認授權前，不直接複製原文圖表；優先使用自製圖表導讀或重繪圖。
| 項目 | 內容 | 關鍵數字 | Jacky 判讀 | 重用策略 |
|---|---|---|---|---|
| Table 3 | Compares the RGB-based and deep-learning extraction strategies by overall accuracy and detection percentage. | RGB-based: accuracy 80.0%, detection 95.9%. Deep learning: accuracy 86.8%, detection 82.5%. | 兩種萃取策略各有取捨，RGB 法偵測率高但精度低，深度學習精度高但偵測率低，可作為談論前端資料準備瓶頸的實證。 | 以自繪表格重述數字，不直接複製原圖表（CC BY-NC-ND 禁改作）。 |
| Table 4 | Point-to-model residual accuracy (Acc) and standard deviation (SD) of the proposed method versus TreeQSM and AdQSM on three sample trees. | 本法 Acc 0.184/0.148/0.263 m；TreeQSM Acc 0.316/0.549/0.363 m；AdQSM Acc 0.051/0.056/0.056 m（三棵樹）。 | AdQSM 殘差最低但原文指出低殘差不等於好模型；本法在不需單木分割下取得可用精度，論點是規模化效率與幾何品質的折衷。 | 自繪比較圖或文字描述，不複製原表。 |
| Fig. 16, Fig. 20 | Fig. 16 為 DBH 與樹高回歸評估；Fig. 20 為 L-system 壓縮前後檔案大小箱型圖。 | DBH RMSE 5.5 cm (13.0%)、樹高 RMSE 2.7 m (13.2%)；檔案 321 KB 降至 23 KB，壓縮率 92.8%。 | 這兩張圖支撐「精度可用 + 輕量可管理」雙重論點，是城市森林數位孿生落地的關鍵實證。 | 以自繪示意圖呈現核心數字趨勢，不複製原圖。 |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| A bottom-up, individual-tree 3D reconstruction pipeline can scale to a whole urban scene and serve as a step towards an urban forest digital twin. | 490/512 trees modeled (95.7%) over 36,400 m2; data acquired in 2-3 h with consumer-grade MLS+ULS. | Abstract; Section 3; Section 5.2; Section 6.1; Conclusion. | True | 可引為大規模、低成本都市樹木建模管線的實證；勿過度宣稱為已完整運作的端到端數位孿生系統。 |
| Reconstructed models are both visually realistic and geometrically usable for forestry parameters. | DBH RMSE 5.5 cm (13.0%), tree height RMSE 2.7 m (13.2%); comparable in magnitude to prior studies. | Section 5.2 (Fig. 16); Section 6.2. | True | 強調精度與既有研究同量級，支援 allometric 應用。 |
| L-system parameterization gives strong, geometry-preserving compression. | Mean file size 321 KB to 23 KB, 92.8% compression; expression about 12% of original skeleton size. | Section 5.4 (Fig. 20). | True | 輕量化是把單木模型放進城市尺度數位孿生可管理的關鍵。 |
Critical Appraisal
Strengths
- 以消費級、低成本設備在 2-3 小時內快速取得大範圍點雲，務實可複製。
- 改良空間殖民演算法直接多樹建模，繞過誤差最大的單木分割步驟。
- L-system 壓縮達 92.8%，解決城市尺度模型儲存與管理瓶頸。
- 同時報告視覺真實度與 DBH/樹高幾何精度，論證較完整。
Weaknesses
- 僅在單一校園場域、以落葉梧桐為主驗證，跨樹種與跨場景泛化未證。
- 葉片重建僅供視覺化，非真實葉片重建，限制生理/輻射傳輸應用。
- 樹點萃取精度（RGB 80.0% / DL 86.8%）仍需人工清理半天，自動化程度有限。
- 標題與摘要使用 digital twin 定位，但本文實為前端 3D 重建與輕量化管線，尚非含即時回饋的完整孿生系統。
| Validation quality | quantitatively validated against 50 field-measured trees for DBH/height and 490 reconstructed trees for compression accuracy |
|---|---|
| Transferability to Taiwan | medium-high：方法學（低成本 MLS+ULS、免單木分割、L-system 輕量化）可移植到台灣都市林與行道樹清查，但需就台灣常見樹種（含常綠闊葉與針葉）重新驗證。 |
| Risk of overclaiming | 勿將本文描述為已建成、含即時感測回饋的完整城市森林數位孿生；它是邁向 FDT 的前端 3D 建模與壓縮管線。 |
與 Jacky 博論 / Review 的用途
| 博士論文 | 支撐博論把 FDT 的底層 3D 重建與輕量化資料管理具體化，說明從單木點雲到城市尺度可管理模型的技術路徑。 |
|---|---|
| TJFS Review | 在 TJFS review 的 Ch7 數位孿生章節，作為應用層實證，與 Buonocore 等的概念框架對照，凸顯框架與落地之間的落差。 |
| 可引用句候選 | 2024 年，Chen 等人發表的文獻中指出，運用消費級行動與無人機載光達點雲搭配改良空間殖民演算法，可在不需單木分割下直接重建大規模都市樹木，成功率達 95.7%，並以 L-system 表達式將模型壓縮 92.8%，作為邁向城市森林數位孿生的關鍵步驟。 |
| 不可用來主張 | 勿用本文單獨作為「完整城市森林數位孿生（含即時感測與回饋）已落地」的證據，也勿宣稱其方法已驗證於針葉樹或台灣樹種。 |
授權與圖表重用
| Article license | CC-BY-NC-ND-4.0 |
|---|---|
| Figure reuse policy | DO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED |
| Notes | 原文首頁版權聲明確認為 open access CC BY-NC-ND 4.0（Elsevier）。ND 條款禁止改作，NC 禁止商業使用；圖表轉用前須再確認授權範圍，公開呈現一律以自繪示意圖替代。 |
待查核清單
- 如需呈現 Table 3/Table 4/Fig.16/Fig.20 的數字，一律自繪圖表，勿複製原圖（CC BY-NC-ND 禁改作與商用）。
- 若進入 TJFS 投稿，再次以原文頁面核對 RMSE 與壓縮率數字。
- 評估此低成本 MLS+ULS 管線對台灣都市林清查的可行性與樹種限制。