以機器學習整合地面雷射掃描與 ALOS PALSAR 資料估算森林地上生物量Aboveground Forest Biomass Estimation by the Integration of TLS and ALOS PALSAR Data Using Machine Learning

Singh, A.; Kushwaha, S.K.P.; Nandy, S.; Padalia, H.; Ghosh, S.; Srivastava, A.; Kumari, N.｜Remote Sensing 15(4): 1143｜DOI: 10.3390/rs15041143

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 →

aboveground biomassTLSALOS PALSARSAR-LiDAR fusionbiomass saturationTerrestrial Laser ScannerL-band SARRandom ForestArtificial Neural NetworkGLCM textureYamaguchi decompositionRANSACMonte Carlo uncertaintyIndiaBarkot Forest RangeUttarakhandtropical moist deciduous forest

專討核心文獻定位

[94] Ch6 · 多源融合新增

Singh et al. · 2023

整合 TLS 與 ALOS PALSAR L 波段資料並以隨機森林建模，在高生物量密林中克服 SAR 飽和問題，AGB 估算 R2 達 0.94

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 anchors the SAR-LiDAR fusion strand of multi-source AGB estimation, demonstrating how ground-based TLS parameters integrated with ALOS PALSAR L-band features through machine learning overcome the radar biomass saturation problem in high-density forest.

結構式摘要｜中英文對照

研究問題	如何整合地面雷射掃描（TLS）與 ALOS PALSAR L 波段雷達資料，並以機器學習克服 SAR 在高密度森林的生物量飽和問題來估算地上生物量？ How can terrestrial laser scanning (TLS) and ALOS PALSAR L-band SAR data be integrated with machine learning to overcome the biomass saturation problem of SAR in high-density forest and estimate aboveground biomass?
資料來源	原文確認，研究區位於印度北阿坎德邦 Dehradun 林區的 Barkot 林場，屬熱帶濕潤落葉林，總面積 84.96 平方公里，以娑羅雙（Shorea robusta）為主。地面以 13 個 31.5 公尺乘 31.5 公尺樣區、TLS（Riegl VZ-400）量測胸高直徑與樹高。雷達採用 2018 年 4 月取得的 ALOS PALSAR 四極化影像（HH、HV、VH、VV），多視後像元解析度 18.42 公尺。共擷取 23 個 TLS 與 SAR 參數進行整合。 The original text confirms the study area is the Barkot Forest Range of Dehradun Forest division, Uttarakhand, India, a tropical moist deciduous forest of 84.96 km2 dominated by Shorea robusta. Ground data come from 13 plots of 31.5 m by 31.5 m with TLS (Riegl VZ-400) measuring dbh and tree height. The radar input is an April 2018 ALOS PALSAR quad-polarization image (HH, HV, VH, VV), multilooked to 18.42 m pixel resolution. A total of 23 TLS and SAR parameters were retrieved for integration.
方法	原文方法為：以 TLS 點雲透過 RANSAC 圓柱擬合取得胸高直徑與樹高，並用全國性樹種材積方程式與 1.59 生物量擴展因子計算生物量；ALOS PALSAR 經前處理後擷取 GLCM 紋理、Yamaguchi 分解成分、極化參數（BMI、CSI、VSI、RVI 等）與 HH、HV 後向散射值；再以隨機森林（RF）與類神經網路（ANN）兩種機器學習演算法整合 TLS 與 SAR 參數估算 AGB，並以拔靴法與蒙地卡羅法量化不確定性。 The original method retrieves dbh and tree height from TLS point clouds via RANSAC cylinder fitting and computes biomass using national species-specific volumetric equations with a biomass expansion factor of 1.59. ALOS PALSAR is preprocessed to derive GLCM texture, Yamaguchi decomposition components, polarimetric parameters (BMI, CSI, VSI, RVI, etc.) and HH/HV backscatter. Random Forest (RF) and Artificial Neural Network (ANN) then integrate the TLS and SAR parameters to estimate AGB, with uncertainty quantified by bootstrap resampling and a Monte Carlo approach.
主要結果	原文確認，隨機森林表現優於類神經網路：RF 的 R2 為 0.94、RMSE 為 59.72 ton ha-1、RMSE% 為 15.97、RMSECV 為 0.15；ANN 的 R2 為 0.77、RMSE 為 98.46 ton ha-1、RMSE% 為 26.32、RMSECV 為 0.23。RF 預測生物量範圍為 122.46 至 581.89 ton ha-1，不確定性範圍為 15.75 至 85.14 ton ha-1，不確定性百分比為 20.54%。整合 SAR 與 LiDAR 資料顯示可有效克服 SAR 的生物量飽和。 The original text confirms Random Forest outperformed ANN: RF achieved R2 of 0.94, RMSE of 59.72 ton ha-1, RMSE% of 15.97, and RMSECV of 0.15; ANN achieved R2 of 0.77, RMSE of 98.46 ton ha-1, RMSE% of 26.32, and RMSECV of 0.23. RF-predicted biomass ranged from 122.46 to 581.89 ton ha-1, with uncertainty ranging from 15.75 to 85.14 ton ha-1 and an uncertainty percentage of 20.54%. Integrating SAR and LiDAR data effectively overcame the SAR biomass saturation.
限制	原文討論指出，本法主要適用於溫帶森林帶但也可用於其他林型，雷達影像於 2018 年 4 月（落葉季）取得，後向散射多來自木質部而非冠層；研究僅以 13 個樣區校正，樣本數有限；不確定性在低密度區較高，因 ALOS 衍生樹屬性與生物量相關性較低。整體仍屬單一熱帶落葉林試驗區，外推性受限。 The discussion notes the method applies mainly to temperate forest zones but can extend to other forest types; the radar image was acquired in April 2018 (leaf-off season), so backscatter came mostly from woody parts rather than canopy. Calibration relied on only 13 plots, a limited sample size; uncertainty was higher in low-density areas because ALOS-derived tree attributes correlated weakly with biomass there. Overall it remains a single tropical deciduous study area, limiting transferability.

Key Findings

發現	證據	確定性
Integrating TLS-derived parameters with ALOS PALSAR L-band features via Random Forest overcomes the SAR biomass saturation problem in high-density forest.	Original Abstract and Conclusion: RF achieved R2 of 0.94 with RMSE 59.72 ton ha-1; integration shows great potential to overcome AGB saturation of SAR data.	checked_against_original_txt
Random Forest clearly outperforms Artificial Neural Network for this integrated AGB estimation.	Original Table 1: RF R2 0.94 / RMSE 59.72 ton ha-1 versus ANN R2 0.77 / RMSE 98.46 ton ha-1.	checked_against_original_txt
Among SAR features, Canopy Structure Index and log-transformed HV backscatter correlate most strongly with biomass; GLCM texture (variance, entropy) also contributes.	Original Section 3.3: CSI versus biomass R2 0.85; HV intensity log-transformed R2 0.77; variance R2 0.52, entropy R2 0.21.	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	Side-by-side accuracy of RF and ANN models for AGB prediction using R2, RMSE, RMSE%, and RMSECV.	RF: R2 0.94, RMSE 59.72 ton ha-1, RMSE% 15.97, RMSECV 0.15. ANN: R2 0.77, RMSE 98.46 ton ha-1, RMSE% 26.32, RMSECV 0.23.	Direct evidence that RF beats ANN for SAR-LiDAR integrated AGB, useful for the fusion-plus-ML argument in the review.	Article is CC BY 4.0, so Table 1 may be reused with attribution; can also redraw as a simple two-row comparison.
Figure 12	Wall-to-wall map of predicted AGB (t/ha) over Barkot Forest Range and the corresponding per-pixel uncertainty map.	Predicted biomass 122.46 to 581.89 ton ha-1; uncertainty 15.75 to 85.14 ton ha-1; overall uncertainty 20.54%.	Shows fusion can produce a spatial AGB product with explicit uncertainty, supporting the MRV-readiness narrative.	CC BY 4.0 allows reuse with attribution; consider a simplified redraw highlighting the uncertainty pattern.

Extracted Evidence Table

可支撐主張	指標或結果	原文位置	可引用	備註
SAR-LiDAR integration with machine learning overcomes L-band biomass saturation in high-density forest.	RF integrated model R2 0.94, RMSE 59.72 ton ha-1, predicting biomass up to 581.89 ton ha-1.	Original Abstract, Table 1, Section 3.5, Conclusion.	True	Results come from a single 13-plot tropical deciduous site; cite as illustrative, not as a universal benchmark.
Random Forest outperforms ANN for integrated AGB estimation.	RF R2 0.94 / RMSE% 15.97 versus ANN R2 0.77 / RMSE% 26.32.	Original Table 1, Section 3.5, Conclusion.	True	Comparison limited to two algorithms on one dataset with 13 plots.

Critical Appraisal

Strengths

Clear demonstration that SAR-LiDAR fusion plus machine learning mitigates L-band biomass saturation.
Quantifies prediction uncertainty with bootstrap and Monte Carlo, producing an explicit uncertainty map.
Open access CC BY 4.0, so figures and tables can be reused with attribution.

Weaknesses

Calibration and validation rest on only 13 plots in a single tropical deciduous forest, limiting generalizability.
Radar image acquired in leaf-off season, so canopy backscatter contribution is reduced and seasonally specific.

Validation quality	moderate; RF validated with cross-validation (RMSECV 0.15) but on a small 13-plot single-site sample
Transferability to Taiwan	moderate; the TLS-plus-L-band-SAR fusion concept transfers, but Taiwan's steep terrain, different forest types, and ALOS PALSAR availability are practical constraints
Risk of overclaiming	Do not present R2 0.94 as a universal SAR-LiDAR fusion benchmark; it derives from one small-sample tropical site.

與 Jacky 博論 / Review 的用途

博士論文	Provides a concrete SAR-LiDAR fusion case for the dissertation's multi-source AGB estimation, linking ground-based TLS inventory with spaceborne L-band radar.
TJFS Review	Supports the TJFS review's multi-source fusion chapter by showing how integrating SAR backscatter, texture, and polarimetric decomposition with TLS overcomes radar saturation, and frames RF as the stronger fusion learner.
可引用句候選	2023 年，Singh 等人發表的文獻中指出，整合地面雷射掃描與 ALOS PALSAR L 波段雷達資料並以隨機森林建模，可在高生物量密林中克服合成孔徑雷達的生物量飽和問題，其地上生物量估算 R2 達 0.94。
不可用來主張	Do not use this paper as evidence that SAR-LiDAR fusion gives validated wall-to-wall biomass maps across forest types; it is a single small-sample tropical site.

授權與圖表重用

Article license	CC BY 4.0
Figure reuse policy	REUSE_ALLOWED_WITH_ATTRIBUTION_UNDER_CC_BY
Notes	MDPI open access article distributed under Creative Commons Attribution (CC BY) 4.0 license; figures may be reused with proper attribution.

待查核清單

Confirm Table 1 layout and Figure 12 uncertainty map in the PDF before publication.
Cross-link with other SAR (PolInSAR) and GEDI source cards for the multi-source fusion narrative.
Consider a self-drawn diagram of the TLS-plus-SAR-to-RF integration workflow (Figures 4 and 5).