以近距離遙測為基礎的整合分析評估森林地上生物量與碳儲量估算之精度Assessing the accuracy of forest above-ground biomass and carbon storage estimation by meta-analysis based close-range remote sensing

Liu, Chen & Zhao｜Forestry Research 5: e017｜DOI: 10.48130/forres-0025-0017

狀態：AI_DRAFT_FROM_REVIEW｜分級：A｜閱讀深度：FULL_TEXT_CHECKED｜Jacky 審核：False

森林數位孿生底層致能

AGBcarbon storageclose-range remote sensingmeta-analysismeta-analysisground LiDARUAV LiDARallometric equationmulti-source fusionglobal

專討核心文獻定位

[70] Ch4 · LiDAR 新增

Liu et al. · 2025

整合分析 187 篇研究指出地面光達在單木與樣區尺度精度最高，UAV 光達適合林分尺度，多源融合可緩解誤差累積

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 quantifies, through a large meta-analysis, how AGB estimation accuracy of close-range remote sensing varies with scale, forest type, method, and allometric variables, providing the empirical accuracy baseline for the LiDAR chapter.

結構式摘要｜中英文對照

研究問題	近距離遙測在不同尺度、林型、方法與自變數下估算森林地上生物量的精度有多高，又該如何透過多源資料整合提升精度。 How accurate is close-range remote sensing for estimating forest aboveground biomass across scales, forest types, methods, and independent variables, and how can multi-source integration improve that accuracy?
資料來源	蒐集自 CNKI、萬方、Google Scholar、VIP 與 Web of Science，檢索期間 2010 年 1 月 1 日至 2025 年 5 月 15 日，初篩超過 2,000 篇英文與 400 多篇中文文獻，最終納入 187 篇文章與 233 筆研究資料，資料庫含逾 1,000 個變數。 Literature was retrieved from CNKI, Wanfang, Google Scholar, VIP, and Web of Science covering 1 January 2010 to 15 May 2025. After screening over 2,000 English and more than 400 Chinese studies, 187 articles and 233 research datasets were included, with a database of more than 1,000 variables.
方法	採單組率整合分析，以決定係數 R2 作為效應量，先以 Freeman-Tukey 雙反正弦轉換穩定變異數，再以逆變異數加權合併效應量並推估 95% 信賴區間，以卡方檢定與 I2 評估異質性，依 p 與 I2 決定固定或隨機效應模型，分析軟體為 Stata 15SE、Excel 2021 與 PyCharm 2023。感測器分為 RGB、光譜、地面光達與 UAV 光達。 A single-group rate meta-analysis was used with R2 as the effect size. A Freeman-Tukey double arcsine transformation stabilized variance, pooled effect sizes were computed by inverse-variance weighting with 95% confidence intervals, and heterogeneity was assessed with the chi-square test and I2 to select fixed- or random-effects models. Analysis used Stata 15SE, Excel 2021, and PyCharm 2023. Sensors were grouped into RGB, spectra, ground LiDAR, and UAV LiDAR.
主要結果	依研究尺度，地面光達在單木尺度精度最高 ES 0.93（95% CI 0.92 至 0.95），顯著優於 UAV 光達 ES 0.81 與 RGB ES 0.8；樣區尺度地面光達 ES 0.87、UAV 光達偏低 ES 0.76；林分尺度除光譜 ES 0.74 外各方法皆高。三大方法精度相近：異速生長方程式 ES 0.83、非參數法 ES 0.82、參數法 ES 0.80。自變數方面 D ES 0.91、D2H ES 0.90 最高，僅以 H 為自變數最低 ES 0.72。多源資料 ES 0.8 並未明顯高於地面光達 ES 0.93，原因是多源多用於結構複雜或大尺度的林地。 By scale, ground LiDAR had the highest single-tree accuracy at ES 0.93 (95% CI 0.92 to 0.95), significantly above UAV LiDAR (ES 0.81) and RGB (ES 0.8). At the plot scale ground LiDAR reached ES 0.87 while UAV LiDAR was lower at ES 0.76; at the stand scale all methods except spectra (ES 0.74) were high. The three method categories were comparable: allometric growth equation ES 0.83, non-parametric ES 0.82, parametric ES 0.80. Among variables, D (ES 0.91) and D2H (ES 0.90) were highest while H alone was lowest at ES 0.72. Multi-source data (ES 0.8) did not exceed ground LiDAR (ES 0.93), because multi-source was mostly applied to structurally complex or large-scale stands.
限制	多數子群組異質性高（I2 多在 70% 至 99% 之間），R2 作為單一效應量無法反映絕對誤差；驗證真值仍依賴樣區調查或森林資源清查，各省區資料品質不一。地面光達雖精度最高卻受耗時與空間覆蓋有限所限，深度學習在近距離遙測情境仍研究不足。 Most subgroups showed high heterogeneity (I2 often between 70% and 99%), and R2 as a single effect size does not capture absolute error. Ground-truth validation still relies on plot surveys or forest inventories whose quality varies by province. Ground LiDAR, though most accurate, is constrained by time cost and limited spatial coverage, and deep learning remains underexplored in close-range remote sensing contexts.

Key Findings

發現	證據	確定性
Ground LiDAR achieves the highest AGB estimation accuracy at single-tree and plot scales, while UAV LiDAR is preferred for stand-scale assessment.	Results: single-tree ground LiDAR ES 0.93 (95% CI 0.92-0.95) vs UAV LiDAR ES 0.81; plot ground LiDAR ES 0.87; conclusion recommends UAV LiDAR for stand scale.	checked_against_original_txt
Allometric, non-parametric, and parametric methods give broadly comparable accuracy, and including D or D2H as the allometric variable yields the highest accuracy while H alone is weakest.	Results: allometric ES 0.83, non-parametric ES 0.82, parametric ES 0.80; D ES 0.91, D2H ES 0.90, H ES 0.72.	checked_against_original_txt
No single sensor is optimal; multi-source fusion mitigates signal saturation and error accumulation in structurally complex forests.	Multi-source vs single-sensor section and Conclusions: multi-source ES 0.8 reflects deployment in complex large-scale stands rather than inferiority of fusion.	checked_against_original_txt

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 1, Fig. 4	Sample distribution across single-tree, plot, and stand scales by sensor, plus pooled effect sizes per scale.	Single-tree 56.55% of samples, plot 32.98%, stand 10.48%; ground LiDAR single-tree ES 0.93 (95% CI 0.92-0.95).	Use as the empirical baseline that ground LiDAR dominates at fine scales while accuracy and sample size both fall as scale broadens.	CC BY 4.0 allows reuse with attribution; may cite numbers directly or redraw a simplified accuracy-by-scale chart.
Tables 3-4, Fig. 6a-b	Pooled effect sizes for parametric, non-parametric, and allometric methods, and for H, D, D2H, and other independent variables.	Allometric ES 0.83, non-parametric ES 0.82, parametric ES 0.80; D ES 0.91, D2H ES 0.90, H ES 0.72.	Supports the argument that variable choice (DBH-based) matters more than method category for single-tree AGB accuracy.	CC BY 4.0 allows reuse with attribution.

Extracted Evidence Table

可支撐主張	指標或結果	原文位置	可引用	備註
Ground LiDAR is the most accurate close-range sensor for single-tree AGB, but accuracy and sample size decline as scale broadens.	Single-tree ground LiDAR ES 0.93 (95% CI 0.92-0.95); plot ES 0.87; stand-scale sample only 10.48% with recommendation to use UAV LiDAR.	Results, Accuracy at different research scales (Page 5); Table 1; Fig. 4.	True	R2-based effect size; do not present as absolute biomass error.
DBH-based allometric variables outperform height-only models for AGB estimation accuracy.	D ES 0.91 (95% CI 0.85-0.97), D2H ES 0.90 (95% CI 0.88-0.93), H ES 0.72 (95% CI 0.56-0.89).	Results, Accuracy with different independent variables (Page 6); Table 4; Fig. 6b.	True	D2H has narrower CI (more stable) though D has slightly higher point estimate.

Critical Appraisal

Strengths

Large evidence base of 187 articles and 233 datasets.
Quantitative, transformation-corrected pooled accuracy estimates across scale, forest type, method, and variable.
Open access under CC BY 4.0, so figures and tables may be reused with attribution.

Weaknesses

High heterogeneity across most subgroups limits generalization of pooled estimates.
R2 effect size measures explanatory power, not absolute biomass error or bias.
Validation truth still depends on plot surveys and inventories of uneven quality.

Validation quality	Pooled R2 with Freeman-Tukey correction and inverse-variance weighting; statistically rigorous but constrained by source-study heterogeneity.
Transferability to Taiwan	high as a sensor-selection and accuracy-baseline reference for Taiwan forest AGB and carbon work
Risk of overclaiming	Do not claim that multi-source fusion is inherently less accurate than ground LiDAR; the lower pooled ES reflects deployment in more complex large-scale stands.

與 Jacky 博論 / Review 的用途

博士論文	Provides the empirical accuracy baseline for the LiDAR layer of the thesis, supporting the case that ground LiDAR anchors single-tree truth while multi-source fusion scales to landscapes.
TJFS Review	Supports the TJFS review's LiDAR and multi-source sections with quantitative ES figures and a clear scale-vs-accuracy trade-off narrative.
可引用句候選	2025 年，Liu 等人發表的文獻中指出，整合 187 篇研究的分析顯示地面光達在單木尺度的估算精度最高，但隨研究尺度擴大精度與樣本數同步下降，因此林分尺度宜改用 UAV 光達並結合多源資料。
不可用來主張	Do not use this paper as evidence of a specific absolute AGB error rate or as proof that any single sensor works best at all scales.

授權與圖表重用

Article license	CC-BY-4.0
Figure reuse policy	REUSE_ALLOWED_WITH_ATTRIBUTION_CC_BY_4_0
Notes	Copyright 2025 by the author(s), published by Maximum Academic Press; open access under CC BY 4.0. Figures and tables may be reused with attribution.

待查核清單

Optionally inspect Fig. 4 to Fig. 6 and Tables 1 to 5 in the PDF for any additional subgroup numbers worth citing.
Decide whether to redraw an accuracy-by-scale chart under CC BY 4.0 attribution for the review.
Cross-check against other Ch4 LiDAR cards (Lin 2025) to avoid duplicated claims.