← 回文獻卡索引

新視角合成與攝影測量於三維林分重建及單木參數萃取之比較分析Comparative Analysis of Novel View Synthesis and Photogrammetry for 3D Forest Stand Reconstruction and Extraction of Individual Tree Parameters

Guoji Tian, Chongcheng Chen, Hongyu Huang｜Remote Sensing 17(9): 1520｜DOI: 10.3390/rs17091520

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

讀學習筆記 →

3D reconstructionnovel view synthesisindividual tree parametersurban forestpoint cloudNeRF3DGSphotogrammetrySfM-MVSTLSChinaFuzhou

專討核心文獻定位

[49] Ch7 · 數位孿生 ★ 新增
Tian et al. · 2025
以兩塊都市林分比較 NeRF、3DGS 與攝影測量，NVS 重建效率遠勝且 NeRF 在複雜林分品質最佳，但 DBH 仍以攝影測量較準

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 an application-layer benchmark that directly compares the two leading novel-view-synthesis methods (NeRF and 3DGS) against classic photogrammetry and TLS reference on real forest stands. It quantifies where image-based neural reconstruction can and cannot replace photogrammetry for extracting forestry parameters, which is the front-end data layer that any forest digital twin must rely on.

結構式摘要｜中英文對照

研究問題
在比單木更複雜的都市林分中，NeRF 與 3DGS 這類新視角合成技術能否在重建效率、點雲品質與單木參數精度上替代或補足傳統攝影測量？
In urban forest stands more complex than single trees, can novel view synthesis methods such as NeRF and 3DGS replace or supplement classic photogrammetry in terms of reconstruction efficiency, point cloud quality, and individual tree parameter accuracy?
資料來源
原文 Section 2.1 與 2.3 確認，研究場域為福州大學旗山校區兩塊對比林分。Plot_1 為開闊地、45 棵欒樹，冬季落葉、樹幹與樹冠清晰，樹高 4 至 11 公尺（平均 7.5）、胸徑 15 至 23 公分（平均 16.7）；Plot_2 一側密林，33 棵秋楓（常綠闊葉），樹高 7 至 12 公尺（平均 9.4）、胸徑 16 至 27 公分（平均 21.2）。影像以 iPhone 11（1200 萬畫素，地面拍攝）與 DJI Phantom 4 無人機（2000 萬畫素，手持與空拍）取得，重疊度皆超過 70%；參考點雲以 RIEGL VZ-400 地面雷射掃描儀於每塊地六站取得，配準精度約 4 mm。三組影像為 Plot_1_Phone 279 張、Plot_1_UAV 268 張、Plot_2_UAV 322 張（Table 1）。
Sections 2.1 and 2.3 confirm two contrasting plots on the Qishan campus of Fuzhou University. Plot_1 is open with 45 golden rain trees, leafless in winter with clearly visible trunks and canopies, tree height 4-11 m (mean 7.5) and DBH 15-23 cm (mean 16.7). Plot_2 has one densely vegetated side with 33 autumn maple trees (evergreen broadleaf), tree height 7-12 m (mean 9.4) and DBH 16-27 cm (mean 21.2). Images were captured with an iPhone 11 (12-megapixel, ground) and a DJI Phantom 4 UAV (20-megapixel, handheld and aerial) with over 70% overlap; the reference point cloud was acquired with a RIEGL VZ-400 TLS using six scan stations per plot at about 4 mm registration accuracy. The three image sets are Plot_1_Phone 279, Plot_1_UAV 268, and Plot_2_UAV 322 images (Table 1).
方法
原文 Section 2.2 與 2.3 確認三條重建管線共用 COLMAP 的 SfM 步驟取得相機位姿與稀疏點雲，再分別走向：攝影測量以 COLMAP 的 MVS 產生稠密點雲；NeRF 採 NeRFStudio 的 Nerfacto 演算法、30,000 次迭代、每步 4096 條光線、學習率 0.01；3DGS 採原始 3DGS、20,000 次迭代、批量 1、學習率 0.01。三組稠密點雲在 CloudCompare 中以五個非樹標記點與 TLS 點雲做 ICP 配準（配準精度 6 至 9 mm），再於 LiDAR360 以高斯濾波去噪、CSF 演算法分離地面與植被並正規化，最後以距離判別群聚法自動分割單木並萃取樹高、胸徑與冠幅。重建實驗硬體為 12 核 CPU、24 GB RAM、NVIDIA RTX 4090（24 GB VRAM）、PyTorch 2.0、CUDA 11.8。
Sections 2.2 and 2.3 confirm that all three pipelines share a COLMAP SfM step to obtain camera poses and a sparse point cloud, then diverge: photogrammetry uses COLMAP MVS for the dense cloud; NeRF uses the Nerfacto algorithm in NeRFStudio with 30,000 iterations, 4096 rays per step, learning rate 0.01; 3DGS uses the original 3DGS with 20,000 iterations, batch size 1, learning rate 0.01. The three dense clouds were ICP-registered to the TLS cloud in CloudCompare using five non-tree markers (6-9 mm accuracy), denoised by Gaussian filtering and ground-vegetation separated by the CSF algorithm in LiDAR360, then automatically segmented into individual trees by distance-discriminant clustering to extract tree height, DBH, and crown diameter. Experiments ran on a 12-core CPU, 24 GB RAM, NVIDIA RTX 4090 (24 GB VRAM), PyTorch 2.0, CUDA 11.8.
主要結果
原文 Section 3 確認三組結論。效率上（Table 2），NeRF 重建只需 12 至 15 分鐘、3DGS 約為 NeRF 的 1.3 倍，COLMAP 則慢 37 至 51 倍（如 Plot_1_UAV COLMAP 724.495 分鐘對 NeRF 14.0 分鐘）。點雲數量上（Table 3），COLMAP 產生最多點、為 NeRF 的 4.4 至 20 倍、3DGS 的 13 至 66 倍，3DGS 最稀疏（多在 80 萬至 160 萬點）。品質上，COLMAP 點最多但樹幹樹冠噪點多、會出現重複樹幹與交錯樹；NeRF 在較複雜的 Plot_2 最接近 TLS、樹高精度最佳（Plot_1_UAV NeRF 樹高 R²=0.9526、RMSE 0.29 m）；惟胸徑萃取仍以攝影測量較準（Plot_1 UAV_COLMAP DBH R²=0.8825、RMSE 0.99 cm，優於 NeRF 的 R²=0.8381、RMSE 1.17 cm）。單木分割上，Plot_2 中 COLMAP 過分割為 38 棵（多切 5 棵），NeRF 與 3DGS 皆正確切出 33 棵。3DGS 因樹幹點過稀，在 Plot_2 無法成功萃取胸徑。無人機影像因解析度高、視角多，萃取精度普遍優於地面手機影像。
Section 3 confirms three sets of results. On efficiency (Table 2), NeRF needs only 12-15 min and 3DGS about 1.3x NeRF, while COLMAP is 37-51x slower (for example Plot_1_UAV COLMAP 724.495 min vs NeRF 14.0 min). On point count (Table 3), COLMAP produces the most points, 4.4-20x NeRF and 13-66x 3DGS, with 3DGS the sparsest (often 0.8-1.6 million points). On quality, COLMAP has the most points but heavy trunk and canopy noise with duplicated trunks and intersecting trees; NeRF is closest to TLS in the more complex Plot_2 and best for tree height (Plot_1_UAV NeRF height R2=0.9526, RMSE 0.29 m); but DBH is still more accurate from photogrammetry (Plot_1 UAV_COLMAP DBH R2=0.8825, RMSE 0.99 cm, better than NeRF R2=0.8381, RMSE 1.17 cm). For segmentation, in Plot_2 COLMAP over-segmented to 38 trees (5 extra) while NeRF and 3DGS both correctly extracted 33. 3DGS failed to extract DBH in Plot_2 because of overly sparse trunk points. UAV images, with higher resolution and more varied perspectives, generally gave more accurate extraction than ground-level smartphone images.
限制
原文 Section 4 與 Section 5 指出多項限制。NVS 與攝影測量的下游品質高度依賴上游 SfM 的相機位姿精度，COLMAP 預設的窮舉匹配在 Plot_1_UAV、Plot_2_UAV 僅解出 173/268 與 268/322 張影像的位姿，須改用含 GPS 的空間匹配才完整。NeRF 在地面手機資料這類視角受限時，地面區域會出現重建錯誤；3DGS 生成稠密點雲能力差，樹幹過稀導致無法估胸徑。胸徑萃取整體仍以攝影測量較準，NVS 尚不能完全取代。Plot_2 因樹高且樹冠密、TLS 垂直掃描角僅 100 度，參考值本身在冠頂有缺失，限制了以 TLS 樹高與冠幅作為基準的可靠度。研究僅在兩塊小型都市林分、特定樹種驗證，跨樹種與跨場景泛化未證。
Sections 4 and 5 note several limitations. Downstream NVS and photogrammetry quality depends heavily on upstream SfM camera-pose accuracy; COLMAP default exhaustive matching solved only 173/268 and 268/322 image poses for Plot_1_UAV and Plot_2_UAV, requiring GPS-based spatial matching for completeness. NeRF produced ground-region reconstruction errors when viewpoints were limited, as with the ground-level phone data; 3DGS generated sparse clouds with too few trunk points to estimate DBH. DBH extraction overall remains more accurate from photogrammetry, so NVS cannot fully replace it. In Plot_2, because trees are tall with dense canopy and the TLS vertical scan angle is only 100 degrees, the reference itself is missing canopy-top points, limiting the reliability of TLS height and crown diameter as ground truth. The study validated only two small urban plots with specific species, so cross-species and cross-scene generalization is unproven.

Key Findings

發現證據確定性
Novel view synthesis (NeRF and 3DGS) reconstructs forest stands far faster than classic photogrammetry, finishing within about 20 minutes versus hundreds of minutes for COLMAP.Section 3.1, Table 2: NeRF 12-15 min, 3DGS about 1.3x NeRF, COLMAP 37-51x slower (Plot_1_UAV COLMAP 724.495 min vs NeRF 14.0 min); training time appears independent of scene size.checked_against_full_text
NeRF gives the best reconstruction quality in complex, dense-foliage stands and the highest tree-height accuracy, while photogrammetry produces duplicated or intersecting trunks.Sections 3.2-3.3 and Conclusion: NeRF closest to TLS in Plot_2; Plot_1_UAV NeRF height R2=0.9526, RMSE 0.29 m; COLMAP shows duplicated trunks and over-segmented Plot_2 to 38 vs correct 33.checked_against_full_text
DBH extraction is still more accurate from photogrammetry than from NeRF, and 3DGS is too sparse to estimate DBH in complex stands.Section 3.3, Fig. 11 and Fig. 14: Plot_1 UAV_COLMAP DBH R2=0.8825, RMSE 0.99 cm vs UAV_NeRF R2=0.8381, RMSE 1.17 cm; in Plot_2_UAV the 3DGS model could not yield DBH due to sparse trunk points.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 2Dense-reconstruction wall-clock time for COLMAP, NeRF, and 3DGS across the three image datasets.COLMAP: Plot_1_Phone 544.292, Plot_1_UAV 724.495, Plot_2_UAV 453.834 min. NeRF: 15.0 / 14.0 / 12.0 min. 3DGS: 18.23 / 17.39 / 17.46 min. COLMAP is 37-51x slower than the two NVS methods.這張表是 NVS 效率優勢最硬的證據，可用來論述把神經式重建放進森林數位孿生前端的可行性，但須提醒 COLMAP 在 Plot_2_UAV 反而比 Plot_1_UAV 快，效率受場景紋理複雜度影響。以自繪長條圖重述時間數字，標註出處 Tian et al. 2025 Table 2，不直接貼原表。
Table 3Total point count per plot and method against the TLS Lidar reference.Plot_1_Lidar 25,617,648；Plot_1_UAV_COLMAP 53,153,623、NeRF 2,573,330、3DGS 806,149；Plot_2_UAV_COLMAP 55,861,268、NeRF 5,465,952、3DGS 831,164。COLMAP 點數為 NeRF 的 4.4-20 倍、3DGS 的 13-66 倍。點數多不等於品質好，COLMAP 點最多卻噪點最多；3DGS 點最少且樹幹過稀，這組數字支撐稠密度與可用性脫鉤的論點。自繪比較圖或文字描述核心倍數關係，不複製原表。
Fig. 10, Fig. 11, Fig. 14Fig. 10 為 Plot_1 樹高線性擬合，Fig. 11 為 Plot_1 胸徑擬合，Fig. 14 為 Plot_2_UAV 胸徑擬合。Plot_1_UAV 樹高 NeRF R2=0.9526、RMSE 0.29 m（最佳）；Plot_1 胸徑 UAV_COLMAP R2=0.8825、RMSE 0.99 cm 優於 UAV_NeRF R2=0.8381、RMSE 1.17 cm；Plot_2_UAV 胸徑 COLMAP R2=0.8747、RMSE 1.60 cm，NeRF R2=0.8648、RMSE 1.63 cm，3DGS 無法萃取。這三張圖是樹高歸 NeRF、胸徑歸攝影測量這個核心結論的量化來源，可用來談 NVS 取代攝影測量的邊界條件。以自繪示意圖呈現 R2 與 RMSE 趨勢，標註出處，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
NVS methods dramatically cut forest-stand reconstruction time compared with photogrammetry.NeRF 12-15 min, 3DGS about 1.3x NeRF, COLMAP 37-51x slower (Table 2).Section 3.1, Table 2.True可引為 NVS 前端效率優勢的實證；勿宣稱其已完整整合成數位孿生系統。
NeRF best handles complex dense-foliage stands and tree height, but photogrammetry remains better for DBH.Plot_1_UAV NeRF height R2=0.9526 RMSE 0.29 m; Plot_1 UAV_COLMAP DBH R2=0.8825 RMSE 0.99 cm vs NeRF R2=0.8381 RMSE 1.17 cm.Sections 3.2-3.3, Fig. 10, Fig. 11, Fig. 14, Conclusion.True強調樹高與胸徑各有適用方法，是談 NVS 取代邊界的關鍵實證。
3DGS cannot generate dense enough trunk points for reliable DBH in complex stands.Plot_2_UAV_3DGS DBH could not be extracted; 3DGS point counts often only 0.8-1.6 million.Section 3.3 (Fig. 14), Table 3, Conclusion item 2.True用以提醒 3DGS 在森林尺度的稠密重建限制。

Critical Appraisal

Strengths

Weaknesses

Validation qualityquantitatively validated against TLS reference using R2 and RMSE for tree height, DBH, and crown diameter across two plots and three reconstruction methods
Transferability to Taiwanmedium-high：以消費級相機與無人機進行 NVS 重建的方法可移植到台灣都市林與行道樹清查，但需就台灣常見常綠闊葉與針葉樹種、以及更大尺度天然林重新驗證。
Risk of overclaiming勿將本文描述為完整森林數位孿生；它是 NVS 與攝影測量在三維重建前端的比較基準，且結論明確指出胸徑仍以攝影測量較準、3DGS 尚不足以估胸徑。

與 Jacky 博論 / Review 的用途

博士論文支撐博論在 FDT 前端三維重建層的技術選型討論，說明 NeRF/3DGS 與攝影測量在效率與不同樹參數精度上的取捨，是後續接上即時感測與碳估算前的基礎環節。
TJFS Review在 TJFS review 的 Ch7 數位孿生章節，作為以新視角合成做林分重建的量化應用層實證，與 Chen 2024 的城市樹木建模、Buonocore 等的概念框架對照，補強框架到落地之間的技術細節。
可引用句候選2025 年，Tian 等人發表的文獻中指出，在都市林分的三維重建中，NeRF 與 3DGS 等新視角合成方法的重建效率遠勝傳統攝影測量，且 NeRF 在密冠複雜林分品質最佳、樹高精度最高（R²=0.9526、RMSE 0.29 m），但胸徑萃取仍以攝影測量較準，3DGS 則因樹幹點過稀而難以估算胸徑。
不可用來主張勿用本文單獨作為「完整森林數位孿生已落地」的證據，也勿宣稱 NVS 已能全面取代攝影測量或其結論已驗證於針葉樹或台灣天然林。

授權與圖表重用

Article licenseCC-BY-4.0
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文首頁版權聲明確認為 open access、CC BY 4.0（MDPI, Basel）。CC BY 允許在標註出處下重製與改作；惟仍依本知識庫保守政策，公開呈現以自繪示意圖替代，正式投稿前再核對授權與出處標註。

待查核清單