← 回文獻卡索引

基於不同前處理之 Sentinel-1 與 Sentinel-2 協同運用於森林地上生物量預測Synergistic Use of Sentinel-1 and Sentinel-2 Based on Different Preprocessing for Predicting Forest Aboveground Biomass

Fang, G.; Yu, H.; Fang, L.; Zheng, X.｜Forests 14(8): 1615, pp.1-19｜DOI: 10.3390/f14081615

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 →

AGBmultisource fusionpreprocessingSentinel-1Sentinel-2SAR textureGLCMatmospheric correctionspeckle filteringMLRXGBoostChinasubtropical

專討核心文獻定位

[87] Ch6 · 多源融合 新增
Fang et al. · 2023
前處理與 S1/S2 融合對低 AGB 森林貢獻有限，僅含 S2 光譜波段的組合表現最佳

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 a counter-example in the multisource-fusion chapter: it shows that for low-biomass forests, integrating Sentinel-1 SAR with Sentinel-2 optical data, and varying atmospheric correction or speckle filtering, gives little or no improvement over optical-only models, which usefully bounds overclaiming about fusion.

結構式摘要｜中英文對照

研究問題
光學感測器、SAR 及兩者整合在採用不同前處理方法時，是否會明顯改變森林地上生物量的預測結果？
Do optical sensors, SAR, and their integration, under different preprocessing methods, make a meaningful difference in predicting forest aboveground biomass?
資料來源
研究區位於中國浙江杭州臨安區，森林以亞熱帶常綠闊葉林為主；地面樣本為 2017 年浙江省森林資源調查的 20,690 個樣區，分為麻櫟、杉木、馬尾松與全部樣區四類。衛星資料為 1 景 Sentinel-1 SAR 與 2 景無雲 Sentinel-2 影像（原文第 2.3.1 節）。
The study area is Lin'an District, Hangzhou, Zhejiang, China, dominated by subtropical evergreen broadleaf forest. Ground data are 20,690 plots from the 2017 Zhejiang forest inventory, split into oak, Chinese fir, Masson pine, and all-plots groups. Satellite inputs are one Sentinel-1 SAR scene and two cloud-free Sentinel-2 scenes (Section 2.3.1).
方法
原文建立 16 組特徵集，比較 Sentinel-2（含/不含大氣校正）、Sentinel-1 SAR（含/不含斑點濾波）、以及兩種 SAR 紋理（一階占有度與二階共生矩陣 GLCM，三種視窗 5×5、17×17、31×31），分別以多元線性迴歸 MLR 與極限梯度提升 XGBoost 建模，並以 10 折交叉驗證、R、RMSE、rRMSE 評估（原文第 2.4、2.5 節）。
The authors build 16 feature sets comparing Sentinel-2 (with/without atmospheric correction), Sentinel-1 SAR (with/without speckle filtering), and two SAR texture families (first-order occurrence and second-order GLCM, in 5x5, 17x17, 31x31 windows). Each is modeled with multiple linear regression (MLR) and XGBoost, evaluated by 10-fold cross-validation using R, RMSE, and rRMSE (Sections 2.4, 2.5).
主要結果
原文確認：只含 S2 光譜波段的組合在 16 組特徵集中表現最佳，只含 S1 的組合普遍最差；整合 S1 與 S2 並未帶來改善。前處理（大氣校正、斑點濾波）對 S2 波段、SAR 後向散射與 SAR 紋理的能力影響甚微。SAR 紋理（尤其 VH 紋理）優於原始雙極化後向散射。S2+SAR 整合的 rRMSE：馬尾松 37.33-38.58%（最佳）、杉木 47.06-51.50%、全部樣區 64.99-69.24%、麻櫟 68.37-73.85%（最差）。XGBoost 未優於 MLR，部分模型甚至更差。
The original text confirms: groups containing only S2 spectral bands were the best among the 16 feature sets, while groups with only S1-based data were generally the worst; integrating S1 and S2 gave no improvement. Preprocessing (atmospheric correction, speckle filtering) had only minor effects on S2 bands, SAR backscatter, and SAR textures. SAR textures, especially VH-based textures, outperformed raw dual-polarization backscatter. For S2+SAR integration, rRMSE was Masson pine 37.33-38.58% (best), Chinese fir 47.06-51.50%, all plots 64.99-69.24%, oak 68.37-73.85% (worst). XGBoost did not beat MLR and was sometimes worse.
限制
原文自承：研究僅限低 AGB 的細緻地表覆蓋樣區，且僅在無雲影像的最佳大氣條件下測試；C 波段 SAR 穿透力與波長低，預測精度（如麻櫟 rRMSE 達 73.85%）遠不及部分前人研究（如他人達 26% 或 19.69% rRMSE），尚不適合直接實務應用；建議改用 L 波段或更長波長資料、衛星植生指數，並擴大研究區與時間段（原文第 4.4、5 節）。
The authors note: the study covers only low-AGB fine land-cover plots and was tested only under best-case cloud-free atmospheric conditions; C-band SAR has low penetration and wavelength, so accuracy (e.g., oak rRMSE up to 73.85%) is far worse than some prior work (26% or 19.69% rRMSE elsewhere) and not yet fit for direct practical use; they suggest L-band or higher-wavelength data, satellite vegetation indices, and broader study areas and time periods (Sections 4.4, 5).

Key Findings

發現證據確定性
For low-biomass forests, integrating Sentinel-1 SAR with Sentinel-2 optical data did not outperform optical-only models.Conclusion and Section 4.3: groups with only S2 imagery were best-performing among all feature sets; integrating S1 and S2 did not contribute an improvement.checked_against_original_txt
Atmospheric correction and speckle filtering had only marginal influence on AGB prediction capacity.Abstract and Section 5: speckle filtering and atmospheric correction marginally influence S2 bands, SAR backscatter, and SAR-based textures.checked_against_original_txt
SAR-based textures, particularly VH-based textures, outperformed raw SAR dual-polarization backscatter.Section 4.2 and Table 6: the two most important SAR-based textural measures outstripped VV and VH backscatter models; VH textures occupied more top positions.checked_against_original_txt
XGBoost did not outperform MLR for AGB modeling; some XGBoost models were even worse.Section 4.3 and Conclusion: the performance difference between the two algorithms was small, and dozens of XGBoost models were worse than MLR.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 4, Figure 6, Figure 7Table 4 defines the 16 feature sets (A-P) combining S2 TOA/BOA with SAR and texture options; Figures 6 and 7 plot predicted rRMSE for groups A-L and M-P across oak, Chinese fir, Masson pine, and all plots.S2+SAR rRMSE：麻櫟 68.37-73.85%、杉木 47.06-51.50%、馬尾松 37.33-38.58%、全部樣區 64.99-69.24%；樣區數 20,690。可作為多源融合章節的反例證據，提醒台灣低生物量森林若用 C 波段 SAR + S2，融合未必勝過純光學，需審慎評估資料源選擇。Redraw a simplified feature-set comparison or rRMSE bar chart; 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 low-biomass subtropical forests, S1+S2 fusion and varied preprocessing do not improve AGB estimation over S2-only optical models.S2-only groups best of 16 feature sets; S2+SAR rRMSE 37.33-73.85% by species; XGBoost not better than MLR; speckle filtering and atmospheric correction marginal.Abstract, Section 4.3, Section 4.4, Section 5; Table 6, Figures 6-7.TrueNumbers traced to the original full text. Use as a fusion-limitation counter-example, not as a general claim that fusion never helps.

Critical Appraisal

Strengths

Weaknesses

Validation quality10-fold cross-validation reported; absolute accuracy weak, especially for oak and all-plots groups
Transferability to Taiwanhigh; same subtropical climate and forest types and freely available Sentinel data, but mainly as a cautionary fusion-limits reference
Risk of overclaimingDo not present these rRMSE values as Taiwan results; they are a low-biomass subtropical-China case where fusion did not help.

與 Jacky 博論 / Review 的用途

博士論文Supports the dissertation's multisource-fusion module as a cautionary boundary case: shows when SAR-optical fusion and heavier preprocessing add little, helping set realistic expectations for a Taiwan FDT pipeline.
TJFS ReviewSupports the TJFS review's multisource-fusion chapter by providing a credible negative result, balancing fusion-success exemplars and reducing the risk of overstating fusion benefits.
可引用句候選2023 年，Fang 等人發表的文獻中指出，在低生物量的亞熱帶森林中，整合 Sentinel-1 SAR 與 Sentinel-2 光學影像、以及調整大氣校正或斑點濾波等前處理，並未優於僅使用 Sentinel-2 光譜波段的模型，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 multisource fusion is generally ineffective; its negative result is specific to low-AGB forests using C-band SAR.

授權與圖表重用

Article licenseCC-BY
Figure reuse policyDO_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；圖表重用前仍建議查核原始授權標示。

待查核清單