以樣區層級而非個體層級的預測表現來選擇估算森林生物量的異速生長方程式Selecting allometric equations to estimate forest biomass from plot- rather than individual-level predictive performance
Picard, Fonton, Boyemba Bosela, Fayolle, Loumeto, Ngua Ayecaba, Sonké, Yongo Bombo, Maïdou, Ngomanda｜Biogeosciences 22: 1413-1426｜DOI: 10.5194/bg-22-1413-2025
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
allometric equationsaboveground biomassmodel selectionMonte Carlo simulationnull forest modelerror partitioningCongo Basincentral Africa
專討核心文獻定位
[62]
Ch2 · 地面量測
新增
以剛果盆地844棵樹資料與蒙地卡羅模擬證明異速方程式的樣區層級預測表現會隨樣區大小改變，MSS可拆成偏差加樹級殘差加係數不確定三項，小樣區由樹級殘差主導故與樹級選模一致，大樣區樹級殘差消失故應改以樣區層級表現選模
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
這篇把異速生長方程式的選模問題從個體層級拉到樣區層級，量化證明同一批樹級資料在不同樣區尺度下的最佳模型會不同，正好支撐地面量測章節對異速方程式不確定性與選模準則的論述。
結構式摘要｜中英文對照
| 研究問題 | 用來把森林清查資料換算成生物量的異速生長方程式，到底該以個體樹層級還是樣區層級的預測表現來選擇，而樣區大小又如何影響這個選擇？ Should the allometric equations used to convert forest inventory data into biomass be selected based on tree-level or plot-level predictive performance, and how does plot size affect this choice? |
|---|---|
| 資料來源 | 原文確認：使用 Fayolle et al. 2018 的剛果盆地個體樹生物量資料集 X，含 844 棵樹、52 個物種、49 個屬，採自六個剛果盆地國家（喀麥隆、中非共和國、剛果、剛果民主共和國、赤道幾內亞、加彭），胸徑範圍 10.3 至 208.0 公分；另用 Chave et al. 2014 泛熱帶資料集中剛果盆地子集 X'（177 棵樹）來擬合樹高對胸徑模型。 The text confirms use of the Congo Basin individual-tree biomass dataset X from Fayolle et al. 2018, with 844 trees from 52 species and 49 genera collected in six Congo Basin countries (Cameroon, Central African Republic, Congo, Democratic Republic of the Congo, Equatorial Guinea, Gabon), with DBH ranging 10.3 to 208.0 cm; a Congo Basin subset X' (177 trees) of the Chave et al. 2014 pantropical dataset was used to fit a height-diameter model. |
| 方法 | 原文方法確認：建立一個以反 J 形指數分布描述森林直徑結構的虛無森林模型（zero/null forest model），由中非樣區平均的林分密度 N = 467 ha-1 與斷面積 G = 29.8 m2 ha-1 推得指數參數 λ = 0.0689 cm-1。以蒙地卡羅方法重抽樣資料集 X 產生 0.1 到 50 公頃的模擬樣區，計算觀測與預測樣區生物量差異的均方和 MSS，並用變異數分析把 MSS 拆成平方偏差、樣區變異與係數變異三項，分別可由森林層級統計量 (NbF)2、(N/A)MSEF 與 N2MEF 近似。比較五個異速方程式（式 3 至 7）與一個兩步模型鏈（先由胸徑預測樹高、再代入生物量式），所有模型以對數轉換線性迴歸擬合，並以 F 檢定比較巢狀模型。單模型蒙地卡羅用 K = 1000、J = 1000，模型鏈用 K = 800、J = L = 50，全部以 R 計算。 The text confirms a null forest stand model representing forest diameter structure as a reverse-J-shaped exponential distribution, with the exponential parameter λ = 0.0689 cm-1 derived from central African mean stand density N = 467 ha-1 and basal area G = 29.8 m2 ha-1. A Monte Carlo approach resampled dataset X to generate simulated plots of 0.1 to 50 ha, computing the mean sum of squared errors (MSS) between observed and predicted plot biomass, and partitioning MSS via analysis of variance into squared bias, plot variability and coefficient variability, approximated by (NbF)2, (N/A)MSEF and N2MEF respectively. Five allometric equations (Eqs. 3-7) and one two-step model chain (height predicted from diameter, then fed into the biomass equation) were compared; all models were fitted by linear regression on log-transformed data, with F-tests comparing nested models. Monte Carlo used K = 1000, J = 1000 for single models and K = 800, J = L = 50 for the chain, all computed in R. |
| 主要結果 | 原文確認：樣區層級的預測表現 MSS 可被三項公式 (NbF)2 + (N/A)MSEF + N2MEF 良好近似，第一項為偏差、第二項為樹級殘差、第三項為模型係數不確定。對小樣區（0.1 ha），MSS 由樹級殘差項主導，最佳模型是式 4（其 (N/A)MSEF 最小），與樹級選模一致。隨樣區增大樹級殘差項以樣區面積反比消失，大樣區（10 ha）時偏差最小的式 7 反而勝出，因為偏差大的式 4 不再被殘差優勢補償。模型鏈在 1 ha 時不如僅用胸徑的模型，但樣區夠大時模型鏈表現較佳。新增變數是否有益會同時取決於層級與樣區大小，樹級 F 檢定與樣區層級 MSS 並非總是一致。 The text confirms that plot-level predictive performance MSS can be well approximated by the three-term formula (NbF)2 + (N/A)MSEF + N2MEF, the first term being bias, the second the tree residual error, and the third the model coefficient uncertainty. For small plots (0.1 ha), MSS is dominated by the tree residual error term, and the best model is Eq. 4 (smallest (N/A)MSEF), consistent with tree-level selection. As plots grow, this term vanishes proportionally to the inverse of plot area, and for large plots (10 ha) the least-biased Eq. 7 wins because the high-bias Eq. 4 is no longer compensated by its residual advantage. The model chain underperformed the diameter-only model at 1 ha but outperformed it for large enough plots. Whether adding a predictor is beneficial depends jointly on the level and plot size, and the tree-level F-test and plot-level MSS do not always agree. |
| 限制 | 原文確認：MSS 未納入量測誤差（作者引 Chave et al. 2004 指其在樣區層級貢獻通常很小）；以模擬森林而非真實森林清查資料產生樣區，作者預期真實資料下偏差對 MSS 的貢獻會隨樣區增大而上升，若如此則本研究結果對偏差角色的估計偏保守；通用對在地公式之爭與 1 ha 以下偏好通用方程式的結論仍需以其他資料集確認。 The text confirms that MSS excluded measurement errors (the authors cite Chave et al. 2004 that these usually contribute little at the plot level); plots were generated from simulated rather than real forest inventory data, and the authors expect the bias contribution to MSS to increase with plot size for real data, which would make their results conservative regarding the role of bias; the general-versus-local equation debate and the preference for general equations below 1 ha still need confirmation with other datasets. |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| The plot-level predictive performance of an allometric equation depends on plot size and can be partitioned into a bias term, a tree residual error term, and a coefficient uncertainty term. | Original Results and Conclusions: MSS is well approximated by (NbF)2 + (N/A)MSEF + N2MEF, where the three terms correspond to bias, tree residual error, and uncertainty on model coefficients. | checked_against_original_txt |
| For small plots model selection by plot-level performance agrees with tree-level selection, but for large plots the tree residual term vanishes so the two selections can diverge and the least-biased model is preferred. | Original Results: at 0.1 ha the lowest-MSS model was Eq. 4 (lowest (N/A)MSEF), matching tree-level ranking; at 10 ha the plot variability term was no longer decisive and the least-biased Eq. 7 outperformed Eq. 4. | checked_against_original_txt |
| For large plots, a chain combining a general biomass equation with a local height-diameter equation provides a good trade-off between bias and coefficient uncertainty. | Original Results and Conclusions: the model chain underperformed the diameter-only model at 1 ha but outperformed it as plot area increased and plot variability vanished. | 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 2, Figure 2, Figure 3 | Table 2 lists tree- and forest-level performance statistics (AIC, σ, R2, biases, MSE, ME) for five fitted equations on the 844-tree dataset; Figure 2 plots coefficient variability, plot variability and squared bias against plot area; Figure 3 partitions MSS into squared bias, plot variability and coefficient variabilities for six models or chains at 0.1, 1 and 10 ha. | 844 trees fitted; tree-level best model Eq. 4 (lowest AIC, smallest residual error); plot variability decreases proportionally to inverse of plot area; squared bias and coefficient variability approximated by (NbF)2 and N2MEF respectively for plots greater than 50 ha; Monte Carlo K = 1000, J = 1000. | 把這組圖表當成異速方程式選模準則隨尺度改變的實證骨架，說明同一批樹級資料在 0.1 ha 與 10 ha 下的最佳模型可以不同，是 FDT 碳儲量層必須顯式建模的尺度依賴誤差。 | 本文為 CC BY 4.0，圖表可註明出處重製；發布時仍偏好自繪一張聚焦三項誤差拆解隨樣區面積變化的簡化圖。 |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| 異速方程式的最佳選擇會隨樣區尺度而改變，小樣區與大樣區可能選到不同模型。 | MSS = (NbF)2 + (N/A)MSEF + N2MEF；0.1 ha 由樹級殘差主導且最佳為式 4，與樹級選模一致；10 ha 殘差項消失，偏差最小的式 7 勝出。 | 原文 Results p.1417-1419；Conclusions p.1422；Table 2、Figure 2、Figure 3。 | True | 可引用作為選模準則尺度依賴的具體論證；結果來自剛果盆地 844 棵樹的模擬森林，非全球或台灣通用值。 |
Critical Appraisal
Strengths
- 把選模準則從個體層級拉到樣區層級，提出可直接計算的三項近似公式 (NbF)2 + (N/A)MSEF + N2MEF。
- 以 844 棵剛果盆地樹的真實資料加蒙地卡羅模擬，系統性比較五個模型與一個模型鏈在 0.1 至 50 ha 的表現。
- 開放取用 CC BY 4.0，程式碼上傳 Zenodo（DOI 10.5281/zenodo.12748213），方法可重現。
Weaknesses
- 以模擬森林而非真實森林清查資料產生樣區，作者自承真實資料下偏差貢獻可能更大。
- MSS 未納入量測誤差，雖引文獻指其貢獻通常較小。
- 1 ha 以下偏好通用方程式等結論仍待以其他資料集確認。
| Validation quality | 理論與蒙地卡羅模擬驗證紮實（三項近似公式對 MSS 拆解吻合），但以模擬森林替代真實清查資料，屬模型內部一致性驗證而非外部田野驗證 |
|---|---|
| Transferability to Taiwan | 中等；以樣區層級而非樹級表現選模的方法學可直接移轉，但 N、G、λ 與物種別異速係數需以台灣本地清查資料重新估計 |
| Risk of overclaiming | 不可把式 4 或式 7 的優劣當成普世結論，原文明言結果取決於樣區大小、森林直徑分布與資料集，且需其他資料集確認。 |
與 Jacky 博論 / Review 的用途
| 博士論文 | 支撐博論主張，異速生長方程式的選擇本身帶有尺度依賴的不確定性，FDT 的碳儲量層在不同空間解析度下不能套用單一最佳公式，而必須顯式建模偏差、樹級殘差與係數不確定三項。 |
|---|---|
| TJFS Review | 強化 TJFS review 地面量測章節，提供異速方程式選模準則隨樣區尺度改變的可引用論證與三項誤差拆解框架。 |
| 可引用句候選 | 2025 年，Picard 等人發表的文獻中指出，異速生長方程式的樣區層級預測表現會隨樣區大小改變，可由偏差、樹級殘差與模型係數不確定三項近似，小樣區的最佳選模與樹級一致，大樣區則應改以樣區層級表現選模並偏好偏差較小的模型。 |
| 不可用來主張 | 不可用本文作為台灣特定碳儲量數字，或作為某一固定異速方程式在所有尺度皆最佳的證據。 |
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 | 首頁版權聲明確認本文 © Author(s) 2025，以 Creative Commons Attribution 4.0 License 散布（開放取用），圖表在註明出處後可重製；發布時仍偏好自繪簡化版聚焦三項誤差拆解訊息。 |
待查核清單
- 發布前可視覺檢查 Table 2、Figure 2 與 Figure 3 的數字與模型排序。
- 自繪一張聚焦三項誤差拆解隨樣區面積變化的簡化圖。
- 比對台灣森林清查的 N、G 與物種別異速係數以建立本地角度。
- 若需要可從 Zenodo 取得作者程式碼重現蒙地卡羅模擬。