於結構與地形複雜之密生針葉林利用 Sentinel-1 與 -2 資料估算總莖材生物量Total Stem Biomass Estimation Using Sentinel-1 and -2 Data in a Dense Coniferous Forest of Complex Structure and Terrain
Georgopoulos, N.; Sotiropoulos, C.; Stefanidou, A.; Gitas, I.Z.｜Forests 13(12): 2157, pp.1-18｜DOI: 10.3390/f13122157
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
AGBtotal stem biomassmultisource fusionSentinel-1Sentinel-2SARGLCM texturepolarization indicesrandom forestk-fold cross-validationGreeceMediterranean
專討核心文獻定位
[91]
Ch6 · 多源融合
新增
在結構與地形複雜的密生針葉林，僅用 Sentinel-1 SAR 的 RF 模型最準，S1+S2 融合僅些微落後
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 in the multisource-fusion chapter as a case where, in a dense uneven-aged coniferous forest with complex terrain, individual Sentinel-1 SAR outperformed both Sentinel-2 optical and the S1+S2 combination for plot-level total stem biomass, with fusion only marginally behind SAR alone. It usefully bounds claims that fusion always wins and that optical always beats SAR.
結構式摘要｜中英文對照
| 研究問題 | 在不均齡結構且地形複雜的密生針葉林中，Sentinel 光學與 SAR 資料能否可靠估算樣區尺度的總莖材生物量，且個別使用與融合使用何者較佳？ In a dense uneven-aged coniferous forest with complex terrain, can Sentinel optical and SAR data reliably estimate plot-level total stem biomass, and which performs better, individual sensors or their fusion? |
|---|---|
| 資料來源 | 研究區為希臘中部 Pindos 山的 Pertouli 大學實驗林，面積 33 平方公里，海拔 1100 至 2073 公尺，平均坡度 35 度，優勢種為 Abies borisii-regis（約佔 90%），屬天然更新多層結構林。地面資料為 50 個樣區的 DBH 與樹高，每個樣區面積 1000 平方公尺，再以異速生長方程式由 DBH 換算各樣區 TSB，TSB 介於 3.3 至 10.3 Mg/1000 m²。衛星資料為 1 景 Sentinel-1B C 波段 GRD 影像（2018 年 7 月 22 日）與 1 景 Sentinel-2A L2A 影像（2018 年 7 月 23 日）。 The study area is the Pertouli University Forest on Pindos Mountain in central Greece, 33 km2, altitude 1100 to 2073 m, mean slope 35 degrees, dominated by Abies borisii-regis (about 90%), a naturally regenerated multi-layered stand. Ground data are DBH and tree height from 50 plots, each 1000 m2, with plot TSB derived from DBH via an allometric equation, ranging 3.3 to 10.3 Mg/1000 m2. Satellite inputs are one Sentinel-1B C-band GRD scene (22 July 2018) and one Sentinel-2A L2A scene (23 July 2018). |
| 方法 | 原文以三步驟流程進行：Sentinel-1 處理與特徵萃取（SNAP 中含定標、地形平坦化、地形校正、Frost 斑點濾波，並計算 6 個極化指數與每極化 10 個 GLCM 紋理）、Sentinel-2 處理與特徵萃取（GEE 取 L2A，計算 8 個植生指數、1 個水分指數與 LAI/FAPAR/FCOVER 三個生物物理參數），以及隨機森林建模與精度評估。先以 Pearson 相關（findCorrelation，門檻 0.9）去除高度共線變數，再建立 SAR、optical、SAR-optical 三個 RF 模型，以重複三次的 10 折交叉驗證評估 R²、RMSE、MAE（原文第 2.4 節、第 3 節）。 The workflow has three parts. Sentinel-1 processing and feature extraction in SNAP includes calibration, terrain flattening, terrain correction, and Frost speckle filtering, then computes 6 polarization indices and 10 GLCM textures per polarization. Sentinel-2 processing in GEE uses L2A products and computes 8 vegetation indices, 1 water index, and three biophysical parameters (LAI, FAPAR, FCOVER). Random forest modeling follows, with Pearson correlation (findCorrelation, threshold 0.9) removing collinear variables. Three RF models (SAR, optical, SAR-optical) are built and assessed by 10-fold cross-validation repeated three times using R2, RMSE, and MAE (Section 2.4, Section 3). |
| 主要結果 | 原文確認：個別使用 Sentinel-1 的 SAR 模型最準（R²=0.74、RMSE=1.76 Mg/1000 m²、MAE=1.48 Mg/1000 m²），優於 optical 模型（R²=0.63、RMSE=2.19、MAE=1.89）與 SAR-optical 融合模型（R²=0.73、RMSE=1.92、MAE=1.63）。融合僅些微落後 SAR 單獨使用，差異不顯著，故就精度與計算時間而言單用 SAR 更有效率。變數重要性方面，Sentinel-1 的 VH_GLCMVariance 紋理與 Sentinel-2 的 SWIR 波段 B11 為最重要變數。三個模型在 TSB 超過 7.5 Mg/1000 m² 處皆出現飽和效應。 The original text confirms: the individual Sentinel-1 SAR model was most accurate (R2=0.74, RMSE=1.76 Mg/1000 m2, MAE=1.48 Mg/1000 m2), beating the optical model (R2=0.63, RMSE=2.19, MAE=1.89) and the SAR-optical fusion model (R2=0.73, RMSE=1.92, MAE=1.63). Fusion only marginally underperformed SAR alone and the difference was not significant, so SAR alone is more efficient in both accuracy and computation time. For variable importance, the Sentinel-1 VH_GLCMVariance texture and the Sentinel-2 SWIR band B11 were the most important variables. All three models showed a saturation effect above 7.5 Mg/1000 m2 of TSB. |
| 限制 | 原文自承：樣本數偏少（50 個樣區），故採 k 折交叉驗證以避免過擬合；三個模型在 TSB 超過 7.5 Mg/1000 m² 處皆飽和，光學感測器無法直接量測樹幹生物量、密林易低估、疏林易高估，C 波段 SAR 穿透力低也導致飽和；將模型套用至植被組成與地形不同的生態系會降低精度；建議未來收集更多樣區並嘗試 ANN 與 XGBoost（原文第 4 節、第 5 節）。 The authors note: the sample size is rather limited (50 plots), so k-fold cross-validation was used to avoid overfitting; all three models saturate above 7.5 Mg/1000 m2, since optical sensors cannot directly measure bole biomass (underestimating in dense stands, overestimating in sparse ones) and C-band SAR has low canopy penetration; applying the models to ecosystems with different vegetation composition and terrain would reduce accuracy; they suggest collecting more plots and testing ANN and XGBoost approaches (Sections 4, 5). |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| Individual Sentinel-1 SAR produced the most accurate plot-level TSB estimates, outperforming Sentinel-2 optical and the S1+S2 fusion in this dense complex-terrain coniferous forest. | Table 5 and Conclusions: SAR R2=0.74, RMSE=1.76, MAE=1.48; optical R2=0.63; SAR-optical R2=0.73. | checked_against_original_txt |
| SAR-optical fusion only marginally underperformed SAR alone, so SAR alone is more efficient in accuracy and computation. | Conclusions: the difference between SAR and SAR-optical models cannot be considered significant; SAR alone is more efficient in prediction accuracy and computational time. | checked_against_original_txt |
| The Sentinel-1 VH_GLCMVariance texture and the Sentinel-2 SWIR band B11 were the most important predictors for TSB. | Section 3.2 and Conclusions: VH_GLCMVariance highest in SAR and SAR-optical models; B11 (SWIR, correlated with B12 at r=0.99) highest in optical model. | checked_against_original_txt |
| All three RF models showed a saturation effect for TSB above 7.5 Mg/1000 m2. | Section 4 and Conclusions: a saturation effect was introduced over 7.5 Mg/1000 m2 for the SAR, optical, and combined RF models. | 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 5, Figure 5, Figure 6 | Table 5 lists 10-fold CV metrics (R2, RMSE, MAE) for the SAR, optical, and SAR-optical models; Figure 5 shows predicted vs measured TSB scatterplots; Figure 6 shows predictor importance for each model. | SAR R²=0.74、RMSE=1.76、MAE=1.48；optical R²=0.63、RMSE=2.19、MAE=1.89；SAR-optical R²=0.73、RMSE=1.92、MAE=1.63（單位 Mg/1000 m²）；樣區 50；飽和門檻 7.5 Mg/1000 m²。 | 可作為多源融合章節的關鍵案例，提醒在地形複雜的密生針葉林，C 波段 SAR 與其紋理可能勝過光學與融合，台灣山區針葉林設計資料源時值得參考。 | Redraw a simplified CV-metric bar chart comparing the three models; confirm exact figure/table numbers against the PDF before publication. |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| In a dense uneven-aged coniferous forest with complex terrain, Sentinel-1 SAR alone outperforms Sentinel-2 optical and S1+S2 fusion for plot-level total stem biomass. | SAR R2=0.74 (RMSE 1.76, MAE 1.48); optical R2=0.63; SAR-optical R2=0.73; saturation above 7.5 Mg/1000 m2; key predictors VH_GLCMVariance and SWIR B11. | Abstract, Section 3.2, Section 4, Section 5; Table 5, Figures 5-6. | True | Numbers traced to the original full text. Use as a fusion-and-sensor case in complex terrain, not as a universal claim that SAR always beats optical. |
Critical Appraisal
Strengths
- Targets total stem biomass specifically, a component under-studied by SAR.
- Addresses a challenging multi-layered uneven-aged forest with complex terrain.
- Systematic comparison of SAR, optical, and fusion under repeated k-fold cross-validation.
Weaknesses
- Small sample size of 50 plots limits generalization.
- TSB derived from a DBH-only allometric equation, not direct stem measurement.
- Saturation above 7.5 Mg/1000 m2 and single-date single-scene imagery.
| Validation quality | 10-fold cross-validation repeated three times; accuracy moderate to good (best R2=0.74) but constrained by small sample and saturation |
|---|---|
| Transferability to Taiwan | moderate to high; Taiwan mountainous coniferous forests share complex terrain and dense canopy, and Sentinel data are freely available, but Mediterranean fir results should not be transferred as Taiwan numbers |
| Risk of overclaiming | Do not claim SAR universally beats optical; this is a complex-terrain dense-coniferous case, and fusion underperformed only marginally. |
與 Jacky 博論 / Review 的用途
| 博士論文 | Supports the dissertation's multisource-fusion module by showing that in complex-terrain dense coniferous forests SAR and its texture can outperform optical and fusion, informing realistic sensor-selection choices for a Taiwan FDT pipeline. |
|---|---|
| TJFS Review | Supports the TJFS review's multisource-fusion chapter with a case where fusion does not beat the best single sensor, balancing fusion-success exemplars and tempering overclaiming about fusion and about optical superiority. |
| 可引用句候選 | 2022 年，Georgopoulos 等人發表的文獻中指出，在結構與地形複雜的密生針葉林中，僅使用 Sentinel-1 SAR 的隨機森林模型對樣區尺度總莖材生物量的估算最準（R²=0.74），優於僅用 Sentinel-2 光學與 S1 加 S2 融合，且融合僅些微落後，顯示多源融合對此類森林的增益有限。 |
| 不可用來主張 | Do not use this paper as evidence that SAR always outperforms optical or that fusion is generally useless; its result is specific to a dense uneven-aged Mediterranean fir forest with complex terrain. |
授權與圖表重用
| Article license | CC-BY |
|---|---|
| Figure reuse policy | DO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED |
| Notes | 原文版權頁標示 Creative Commons Attribution CC BY 4.0 開放取用授權（MDPI Forests）；圖表重用前仍建議查核原始授權標示。 |
待查核清單
- Fill txt and translation_md asset paths once produced.
- Verify exact R2/RMSE/MAE values and figure/table numbers against the PDF before journal citation.
- Cross-check this SAR-favorable complex-terrain result against fusion-success exemplars when writing the Ch6 narrative.
- Prepare a self-made three-model CV-metric comparison diagram before any figure reuse.