← 回文獻卡索引

以階層式方法結合機器學習與衛星影像推進森林碳儲量製圖Advancing forest carbon stocks' mapping using a hierarchical approach with machine learning and satellite imagery

Illarionova, S., Tregubova, P., Shukhratov, I., Shadrin, D., Efimov, A. & Burnaev, E.｜Scientific Reports 14: 21032｜DOI: 10.1038/s41598-024-71133-8

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

森林數位孿生方法平台

讀學習筆記 →

carbon stockgrowing stock volumeforest structure mappingXGBoostSentinel-2hierarchical approachBCEF conversionRussiaboreal

專討核心文獻定位

[39] Ch8 · 研究缺口 新增
Illarionova et al. · 2024
以 Sentinel-2 加 XGBoost 建立全自動碳儲量製圖流程，階層式法的碳儲量估算略優於直接法

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 sits at the research-gap chapter because it explicitly addresses the absence of a standardized, fully automated pipeline for estimating forest carbon stock from freely available satellite data, and it contrasts a direct against a hierarchical machine-learning approach, which frames a concrete methodological gap that a forest digital twin should close.

結構式摘要｜中英文對照

研究問題
能否僅用免費且定期更新的衛星觀測資料建立全自動的機器學習流程，準確估算森林結構特徵與莖部碳儲量？
Can an effective fully automated machine-learning pipeline estimate forest structure characteristics and stem carbon stock using solely freely available and regularly updated satellite observations?
資料來源
研究區為俄羅斯阿爾漢格爾斯克州歐洲泰加林，主要樹種為歐洲雲杉與歐洲赤松。使用 2018 年清查所得的林分級調查資料，調查面積 126,641 公頃，共 12,617 個林分，平均林分面積 10 公頃。遙測資料為 Sentinel-2 L2A，採用 10 個光譜波段並加入 NDVI、EVI、GNDVI 三個植生指數，挑選 2018 至 2020 年雲量低於 10% 的夏季影像共八個日期，逐像素標註後以 80/20 切分訓練與測試集。
The study area is the European taiga of the Arkhangelsk region, Russia, dominated by Norway spruce and Scots pine. It uses stand-level forest inventory data from a 2018 survey covering 126,641 hectares with 12,617 stands and an average stand area of 10 hectares. Remote sensing input is Sentinel-2 L2A using 10 spectral bands plus three vegetation indices (NDVI, EVI, GNDVI), with eight summer dates between 2018 and 2020 having less than 10% cloud cover, labeled pixel-wise and split 80/20 into training and test sets.
方法
以 Extreme Gradient Boosting（XGBoost）逐像素進行分類與回歸，分類用於優勢樹種，回歸用於林齡、樹高、斷面積等。比較兩種取得材積（生長蓄積量）與莖部碳儲量的方法：直接法是從衛星影像直接預測目標圖，階層式法則先預測中間參數再透過公式換算，材積由斷面積乘樹高（公式1）導出，碳儲量由生長蓄積量乘 BCEF 與含碳係數（公式2）導出，BCEF 與含碳係數依樹種與年齡組取自既有北方歐亞地區研究。
Extreme Gradient Boosting (XGBoost) is applied pixel-wise for classification and regression: classification for dominant species and regression for age, height, and basal area. Two approaches for growing stock volume and stem carbon stock are compared. The direct approach predicts target maps straight from imagery, while the hierarchical approach first predicts intermediate parameters and then applies conversion formulas; timber volume is derived as basal area times height (Formula 1), and carbon mass as growing stock volume times BCEF times a carbon content coefficient (Formula 2), with BCEF and carbon coefficients by species and age group taken from a prior northern-Eurasia study.
主要結果
優勢樹種分類測試集整體 F1-score 為 0.75，松最高 0.85、樺木 0.81、雲杉 0.72、山楊 0.62。回歸方面林齡平均 R² 為 0.75（MAPE 0.195）、樹高 R² 0.58、斷面積 R² 0.56。材積以階層式法（R² 0.57、MAPE 0.391）優於直接法（R² 0.52、MAPE 0.402）。碳儲量同樣以階層式法（R² 0.53、MAPE 0.361）略優於直接法（R² 0.51、MAPE 0.380）。秋季影像（2018-09-11）表現明顯較差，建議排除或校正。
On the test subset the overall species classification F1-score is 0.75, with pine highest at 0.85, birch 0.81, spruce 0.72, and aspen 0.62. For regression, age reaches an average R-squared of 0.75 (MAPE 0.195), height 0.58, and basal area 0.56. For timber stock the hierarchical approach (R-squared 0.57, MAPE 0.391) outperforms the direct approach (R-squared 0.52, MAPE 0.402). For carbon stock the hierarchical approach (R-squared 0.53, MAPE 0.361) again slightly outperforms the direct approach (R-squared 0.51, MAPE 0.380). The autumn image (2018-09-11) performed clearly worse and is recommended for exclusion or correction.
限制
全文只測試莖部碳庫，未涵蓋其他生物量碳庫；採用單一機器學習演算法 XGBoost 以驗證流程可行性；採逐像素方法而非整林分聚合，且林分級資料的不確定性難以在林分內部控制。資料集僅就請求提供，不公開。作者指出階層式法雖較慢但較穩健，未來將嘗試其他 ML 或 DL 演算法及多任務架構。
The study tests only the stem carbon pool, not other biomass pools; it uses a single algorithm, XGBoost, to demonstrate pipeline feasibility; it adopts a pixel-wise rather than stand-aggregated approach, and uncertainties within stand-level data are hard to control inside a stand. The dataset is available only on request, not openly. The authors note the hierarchical approach is slower but more robust, and future work will employ other ML or DL algorithms and multitask architectures.

Key Findings

發現證據確定性
A fully automated pipeline using only free Sentinel-2 imagery plus XGBoost can map forest structure and stem carbon stock without field measurements or state forest registers.Abstract and Methods state the goal of estimating growing stock volume and stem carbon stock in a fully automated manner using solely freely available satellite observations, integrating the BCEF equation into the computation.checked_against_original_txt
The hierarchical approach (predict intermediate parameters then convert) outperforms the direct approach for both timber stock and carbon stock.Results and Table 14: carbon stock hierarchical R-squared 0.53 vs direct 0.51; timber stock hierarchical R-squared 0.57 vs direct 0.52.checked_against_original_txt
The hierarchical setup is more interpretable and flexible because each intermediate parameter is trained independently and can be substituted with better data such as LiDAR-based height.Discussion states intermediate parameters can be substituted (for example height from LiDAR or high-resolution RGB) to improve carbon stock estimation, and that hierarchical methods, though slower, offer robustness and interpretability.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 14Summary table comparing MAPE, MAE, RMSE, and R-squared for age, height, basal area, timber stock (direct and hierarchical), and carbon stock (direct and hierarchical).Age R-squared 0.75 (MAPE 0.195); height R-squared 0.58; basal area R-squared 0.56; timber direct R-squared 0.52 vs hierarchical 0.57; carbon direct R-squared 0.51 vs hierarchical 0.53 (MAPE 0.361).This single table is the cleanest evidence that adding a physically meaningful conversion layer on top of ML predictions improves carbon estimation, which supports a digital-twin style layered architecture.License is CC BY-NC-ND, so do not reproduce or adapt the original table publicly; redraw a self-made comparison chart of direct vs hierarchical R-squared with citation.
Table 5, Table 6Macro-averaged precision, recall, and F1-score for species over eight dates (Table 5) and per-species metrics (Table 6) for spruce, birch, pine, aspen.Overall F1-score 0.75; pine 0.85, birch 0.81, spruce 0.70-0.72, aspen 0.62.Useful to show species-level reliability varies widely, which matters when species-specific conversion factors feed the carbon calculation.Do not reproduce original table publicly under CC BY-NC-ND; cite numbers in text only or redraw.
Figure 1Overall workflow from Sentinel-2 imagery and forest inventory data through the supervised ML pipeline to conversion recalculations.13 input features per date (10 bands plus three vegetation indices); eight observation dates; 80/20 train-test split.The pipeline diagram is the conceptual anchor showing where the conversion layer sits, which maps onto a digital-twin data-to-decision flow.NoDerivatives license: do not adapt; create an original simplified pipeline diagram with citation if needed.

Extracted Evidence Table

可支撐主張指標或結果原文位置可引用備註
There is no standardized fully automated pipeline for estimating forest characteristics through remote sensing and ML.Stated explicitly as a primary challenge in forestry studies, motivating an end-to-end solution using only freely available satellite data.Introduction (the passage on the absence of a standardized pipeline) and Methods.TrueStrong gap statement for the research-gap chapter; frame as the authors' own positioning, not as a universal proven fact.
The hierarchical approach yields better and more interpretable carbon stock estimates than the direct approach.Carbon stock hierarchical R-squared 0.53 (MAPE 0.361) vs direct R-squared 0.51 (MAPE 0.380); timber hierarchical R-squared 0.57 vs direct 0.52.Results section and Table 14; Discussion.TrueThe improvement is modest; describe it as a slight but consistent gain with added interpretability, not a large performance jump.

Critical Appraisal

Strengths

Weaknesses

Validation qualityCross-date test-subset evaluation with standard regression and classification metrics; no independent external validation region beyond the held-out forest.
Transferability to TaiwanMedium; the pipeline concept transfers well but BCEF and carbon coefficients are calibrated for boreal northern-Eurasia species and would need Taiwan-specific allometric or conversion factors.
Risk of overclaimingDo not claim the hierarchical approach is dramatically more accurate; the carbon R-squared gain is only about 0.02, and the result is for the stem pool in a boreal region.

與 Jacky 博論 / Review 的用途

博士論文Supports the dissertation argument that physically meaningful conversion layers should be combined with ML predictions, and that a forest digital twin can substitute intermediate parameters (such as LiDAR height) to improve carbon estimation.
TJFS ReviewProvides a concrete research-gap citation for the TJFS review, showing the lack of a standardized automated carbon pipeline and the value of hierarchical over direct estimation.
可引用句候選2024 年，Illarionova 等人發表的文獻中指出，僅用免費的 Sentinel-2 影像搭配 XGBoost 即可建立全自動的森林碳儲量製圖流程，且先預測中間參數再以換算公式求碳儲量的階層式法，在莖部碳儲量上略優於直接從影像預測的作法。
不可用來主張Do not use this paper as evidence for total forest carbon or for tropical and subtropical forests; it tests only the stem pool in boreal taiga and uses region-specific conversion factors.

授權與圖表重用

Article licenseCC BY-NC-ND 4.0
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文授權頁明列 Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International（CC BY-NC-ND 4.0）。NoDerivatives 條款表示不得公開分享改作素材，圖表一律以自繪詮釋替代，引用須註明出處與授權連結。

待查核清單