← 回文獻卡索引

結合 GEDI LiDAR、衛星資料與機器學習演算法於 Google Earth Engine 平台估算森林地上生物量的多源方法A multi-source approach combining GEDI LiDAR, satellite data, and machine learning algorithms for estimating forest aboveground biomass on Google Earth Engine platform

Zurqani, H. A.｜Ecological Informatics 86: 103052｜DOI: 10.1016/j.ecoinf.2025.103052

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 fusioncarbon stockGEDI LiDARSentinel-1Sentinel-2Google Earth EngineRandom ForestCARTGradient Tree BoostingSVMUSAArkansastemperate bottomland forest

專討核心文獻定位

[46] Ch6 · 多源融合 新增
Zurqani et al. · 2025
在 GEE 平台融合 GEDI、Sentinel-1/2 與地形與冠層高度，GTB 模型 AGB 估算達 R²=0.77

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 recent, fully open-access exemplar of multi-source remote-sensing fusion for AGB, demonstrating how GEDI LiDAR is used as calibration reference together with Sentinel optical and SAR data, topographic derivatives, and canopy height to compare four machine learning algorithms on a cloud-based Google Earth Engine workflow.

結構式摘要｜中英文對照

研究問題
如何在 Google Earth Engine 平台上結合光學與微波衛星觀測與機器學習演算法，提升森林地上生物量的估算與製圖精度？
How can optical and microwave satellite observations be combined with machine learning algorithms on the Google Earth Engine platform to improve forest aboveground biomass estimation and mapping?
資料來源
研究區為美國阿肯色州下密西西比沖積河谷的 Green Tree Reservoir，面積 2184.3 公頃，森林占 68.86%。資料以 GEDI Level 4A 地上生物量密度作為校正參考（圖 1 標示 n=308 參考點），並整合 Sentinel-2 光學、Sentinel-1 SAR、3DEP 地形衍生變數（高程與坡度）與全球冠層高度圖。
The study area is the Green Tree Reservoir in the Lower Mississippi Alluvial Valley, Arkansas, USA, covering 2184.3 ha with forest occupying 68.86%. GEDI Level 4A aboveground biomass density served as calibration reference (Fig. 1 marks n=308 reference points), integrated with Sentinel-2 optical, Sentinel-1 SAR, 3DEP topographic derivatives (elevation and slope), and a global canopy height map.
方法
全程在 Google Earth Engine 平台處理，以 GEDI L4A 取代傳統地面真值作為校正資料，70% 訓練、30% 測試，採重複十摺交叉驗證。比較四種機器學習迴歸演算法 Random Forest、CART、Gradient Tree Boosting 與 SVM，以 R²、RMSE、MAE 評估，並以雙因子 ANOVA 分析模型與變數組合的主效應與交互作用。
All processing was done on the Google Earth Engine platform, using GEDI L4A as calibration data instead of conventional ground truth, with a 70% training and 30% testing split and repeated 10-fold cross-validation. Four machine learning regression algorithms (Random Forest, CART, Gradient Tree Boosting, and SVM) were compared using R², RMSE, and MAE, and a two-way ANOVA assessed the main and interaction effects of model type and variable combinations.
主要結果
原文確認：GTB 模型表現最佳，以 Sentinel-2 波段加地形衍生變數達 R²=0.77、MAE=22.27 Mg/ha、RMSE=37.78 Mg/ha；RF 次之，達 R²=0.74、MAE=31.34 Mg/ha、RMSE=39.93 Mg/ha；CART 達 R²=0.68；SVM 表現最差，最高 R² 僅 0.37。高程與冠層高度是最重要的變數（變數重要性皆達 10.00 與 6.12）。整合 Sentinel-2 波段、植生指數、地形資料與冠層高度可顯著提升模型表現，而納入 Sentinel-1 雷達資料對多數模型未帶來明顯改善。
The original text confirms that the GTB model performed best, achieving R²=0.77, MAE=22.27 Mg/ha, and RMSE=37.78 Mg/ha using Sentinel-2 bands and topographic derivatives. RF followed with R²=0.74, MAE=31.34 Mg/ha, and RMSE=39.93 Mg/ha; CART reached R²=0.68; SVM performed worst with a maximum R² of only 0.37. Elevation and canopy height were the most important variables (importance scores of 10.00 and 6.12 in RF). Integrating Sentinel-2 bands, vegetation indices, topographic data, and canopy height significantly improved performance, while adding Sentinel-1 radar did not consistently improve most models.
限制
原文 Section 4.4 指出：SVM 在所有變數組合下皆表現不佳；高品質地面真值有限，主要倚賴遙測估計值，限制了模型參數的充分優化；衛星衍生資料受空間解析度、大氣干擾與感測器校正影響；訓練樣本數有限。作者建議擴充參考資料集、納入 LiDAR 並嘗試 ANN/深度學習等進階方法。
Section 4.4 notes that SVM underperformed across all variable combinations; high-quality ground truth was limited and relied mainly on remotely sensed estimates, constraining full parameter optimization; satellite-derived data is affected by spatial resolution, atmospheric interference, and sensor calibration; and training samples were limited. The author recommends expanding reference datasets, incorporating LiDAR, and exploring advanced methods such as ANN and deep learning.

Key Findings

發現證據確定性
Among four ML algorithms, Gradient Tree Boosting (GTB) achieved the best AGB estimation accuracy.Original abstract and Section 3.1: GTB reached R²=0.77, MAE=22.27 Mg/ha, RMSE=37.78 Mg/ha using Sentinel-2 bands and topographic derivatives; SVM performed worst at R²=0.37 max.checked_against_original_txt
Integrating Sentinel-2 bands, vegetation indices, topographic derivatives, and canopy height drove accuracy, while Sentinel-1 radar added little.Original Section 3.3 and 4.2: the S2+S2VI+TD+CH combination yielded the best results; including Sentinel-1 did not consistently improve CART and SVM models.checked_against_original_txt
GEDI Level 4A data can serve as a viable calibration reference in place of conventional ground truth for large-area AGB mapping.Original Section 2.3: GEDI L4A used as calibration dataset rather than ground truth, with 70/30 split and 10-fold cross-validation; n=308 reference points (Fig. 1).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, Fig. 8, Fig. 9Table 1 summarizes all remotely sensed input variables (Sentinel-2 bands and indices, Sentinel-1 indices, topographic derivatives, canopy height). Fig. 4 plots RMSE vs R² with MAE as bubble size for the four models. Fig. 8 shows variable importance for RF, CART, and GTB. Fig. 9 shows the AGB maps produced by each model with its best variable combination.GTB R²=0.77, MAE=22.27 Mg/ha, RMSE=37.78 Mg/ha; RF R²=0.74, MAE=31.34 Mg/ha, RMSE=39.93 Mg/ha; CART R²=0.68; SVM R²=0.37 max; GEDI L4A reference range 28.14-399.91 Mg/ha; GTB estimate range 26.52-357.0 Mg/ha; elevation importance 10.00, canopy height 6.12 in RF.Use as a methodological template for a fully open-data, cloud-based AGB pipeline; the headline is that GEDI plus Sentinel plus topography/canopy on GEE can reach R²=0.77 without field plots, which is attractive for Taiwan where field access is hard.Redraw a simplified workflow and a model-comparison chart; confirm exact figure/table numbers against the PDF and the CC BY-NC attribution before any reuse.

Extracted Evidence Table

可支撐主張指標或結果原文位置可引用備註
A fully open-data, GEE-based multi-source pipeline using GEDI as calibration can map forest AGB at moderate-to-high accuracy.GTB R²=0.77 (MAE 22.27, RMSE 37.78 Mg/ha); RF R²=0.74; CART R²=0.68; SVM R²=0.37 max; reference AGB 28.14-399.91 Mg/ha over a 2184.3 ha Arkansas study site.Original abstract, Section 2.1-2.3, Section 3.1-3.5, Table 1, Fig. 4, Fig. 9.TrueNumbers checked against full text. Results are from a single temperate bottomland forest in Arkansas; do not over-extrapolate to other forest types or to Taiwan without local validation.

Critical Appraisal

Strengths

Weaknesses

Validation qualitymoderate; cross-validated against held-out GEDI test data but no independent field validation
Transferability to Taiwanmoderate to high as a method; the open-data GEE pipeline transfers well, but accuracy must be re-validated on Taiwan forest types and terrain
Risk of overclaimingDo not present R²=0.77 as a general or Taiwan-applicable accuracy; it is a single-site Arkansas result calibrated on GEDI rather than field measurements.

與 Jacky 博論 / Review 的用途

博士論文Supports the dissertation's multi-source-fusion and open-database modules, showing a concrete GEDI-plus-Sentinel-plus-GEE recipe and the relative value of optical, SAR, topographic, and canopy-height inputs.
TJFS ReviewSupports the TJFS review's Ch6 multi-source-fusion section as a 2025 open-access exemplar with quantified algorithm comparison and a clear statement on the limited marginal value of Sentinel-1 in this setting.
可引用句候選2025 年，Zurqani 等人發表的文獻中指出，在 Google Earth Engine 平台上以 GEDI Level 4A 作為校正參考，結合 Sentinel-2 波段、植生指數、地形衍生變數與冠層高度，Gradient Tree Boosting 模型可將森林地上生物量估算的 R² 提升至 0.77。
不可用來主張Do not use this paper as evidence of a validated, field-calibrated, Taiwan-wide AGB system, nor to claim that SAR data is generally unhelpful for AGB.

授權與圖表重用

Article licenseCC-BY-NC
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文首頁標示為 Elsevier 開放取用文章，採 CC BY-NC 4.0 授權（http://creativecommons.org/licenses/by-nc/4.0/）。CC BY-NC 允許非商業重製並註明出處，但圖表公開重用前仍建議查核原始授權標示。

待查核清單