可規模化的年度地上生物量產品用於監測生態系復育專案的碳影響A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects
Atzberger, C., Immitzer, M., Hemes, K.S., Kästenbauer, M., López, J., Terra, T., Rajadel-Lambistos, C., de Souza, S.F., Trabaquini, K., Wolff, N.｜Remote Sensing of Environment 327: 114774｜DOI: 10.1016/j.rse.2025.114774
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
AGBfoundation modelcarbon creditsMRVecosystem restorationself-supervised learningBarlow Twinsrepresentation learningGEDISentinel-2Landsatneural network regressionBrazilParatropical
專討核心文獻定位
[30]
Ch5 · 機器學習
新增
用自監督 Barlow Twins 把年度多光譜時序壓成 32 維表徵，結合 GEDI 相對高度差，產出 10 m 年度 AGB 圖，RMSE 約 24.6 Mg/ha
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 core Ch5 source showing how a self-supervised foundation model (Barlow Twins on multi-spectral EO time series) combined with sparse GEDI measurements can produce scalable annual AGB maps, representing the shift from classical regression toward foundation-model-based AGB estimation.
結構式摘要｜中英文對照
| 研究問題 | 能否把被雲嚴重污染、觀測不規則的 Landsat 與 Sentinel-2 光學時序，壓縮成與 AGB 高度相關的年度表徵，並用單一年份校準的模型產出可跨 10 年以上、10 至 30 m 解析度的年度 AGB 圖？ Can heavily cloud-corrupted and irregular Landsat and Sentinel-2 optical time series be condensed into annual representations highly correlated with AGB, and can a model calibrated on a single year produce 10–30 m annual AGB maps spanning more than ten years? |
|---|---|
| 資料來源 | 研究區為巴西帕拉州（Pará，面積大於 125 萬平方公里）。預測變數來自 Sentinel-2（2017-2024）與 Landsat 7/8（2013-2019）時序，每年壓成 32 個表徵。目標變數來自 GEDI 相對高度差 ΔRH95-10，僅用 2019 年資料校準，過濾後保留 979,637 個 GEDI 樣點。驗證用 38 個原位農林系統樣站（共 100 個 30×30 m 樣方、6994 株 DBH>5 cm 木本植物），外加 GEDI L4B、ESA CCI 等共五組獨立資料集，皆未用於校準。 The study area is Pará State, Brazil (>1.25 million km2). Predictors come from Sentinel-2 (2017-2024) and Landsat 7/8 (2013-2019) time series, each condensed into 32 representations per year. The target variable is the GEDI relative-height difference ΔRH95-10, calibrated using only 2019 data, with 979,637 filtered GEDI footprints retained. Validation uses 38 in-situ agroforestry sites (100 plots of 30×30 m, 6994 woody plants with DBH>5 cm) plus five independent datasets including GEDI L4B and ESA CCI, none used for calibration. |
| 方法 | 核心方法是先用自監督的光譜-時序 Barlow Twins 把年度多光譜時序壓成 32 維表徵（僅用 2019 年資料訓練後凍結），再把 GEDI ΔRH95-10 透過縮放因子 SF≈9 轉成 AGB*RH 代理目標（Eq. 1），最後以一個 6 層全連接前饋神經網路（NNregressor，神經元 128-64-32-16-8-4）回歸出 AGB。模型只用 2019 年校準後即凍結，套用到所有年份；另做一個改用 GEDI L4A AGB 為目標的消融研究比較。 The core method first uses a self-supervised spectral-temporal Barlow Twins model to condense annual multi-spectral time series into 32 representations (trained only on 2019 data, then frozen). GEDI ΔRH95-10 is converted into an AGB*RH proxy target via a scaling factor SF≈9 (Eq. 1). A 6-layer fully connected feedforward neural network (NNregressor; neurons 128-64-32-16-8-4) then regresses AGB. The model is calibrated only on 2019 and then frozen for all years. An ablation study regressing directly against the GEDI L4A AGB target is used for comparison. |
| 主要結果 | 對 38 個原位農林樣站驗證，模型解釋了超過 65% 變異（R2 = 0.67），RMSE 約 24.6 Mg/ha，偏差幾乎為零（bias = 3.49 Mg/ha），無明顯飽和。與五個公開產品（GLAD CHM、ETH CHM、ESA CCI、Meta CHM、GTDX AGB）相比，RMSE 降低 15-55%。Landsat 與 Sentinel-2 在 2017-2019 重疊期的 AGB 產品一致性高（R2 約 0.9、RMSE 約 25 Mg/ha）。成功產出帕拉州 2013-2024 的年度 AGB 圖，證明可規模化。 Validated against 38 in-situ agroforestry sites, the model explains over 65% of the variance (R2 = 0.67) with an RMSE of about 24.6 Mg/ha and nearly zero bias (bias = 3.49 Mg/ha), without observable saturation. Compared with five openly available products (GLAD CHM, ETH CHM, ESA CCI, Meta CHM, GTDX AGB), RMSE is reduced by 15–55%. AGB products from Landsat and Sentinel-2 in the 2017-2019 overlap period are highly consistent (R2 ≈ 0.9, RMSE ≈ 25 Mg/ha). Annual AGB maps for Pará State from 2013 to 2024 were produced, demonstrating scalability. |
| 限制 | GEDI 參考資料本身可重現性偏低（Holcomb 等人報告重複觀測 R2<0.4、RMSE 約 56 Mg/ha），低 AGB 區有系統性高估、高碳儲量區有低估；25 m 的 GEDI 足跡被直接指派到中心像元，並未校正足跡與像元的位置錯位；目前僅在帕拉州一個生態區驗證，跨生態區的縮放因子仍待評估。 The GEDI reference data itself has low reproducibility (Holcomb et al. report R2<0.4 between repeat observations, RMSE ~56 Mg/ha). There is systematic overestimation at low AGB and underestimation at high carbon stocks. The 25 m GEDI footprint is simply assigned to the centroid pixel without correcting footprint-pixel mismatch. Validation is currently limited to one ecoregion in Pará, and transferring the scaling factor to other ecoregions remains to be evaluated. |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| A self-supervised foundation model (Barlow Twins) calibrated on a single year of GEDI data can produce consistent annual deca-metric AGB maps spanning more than a decade. | Original Section 2.4 and 4: a single Barlow Twins model and a single NNregressor per sensor were trained only on 2019, frozen, and applied to 2013-2024; Fig. 8 and Fig. 13 show only slight deterioration in untrained years. | checked_against_original_txt |
| Using the GEDI relative-height difference ΔRH95-10 (scaled by SF≈9 into AGB*RH) as the target outperforms calibrating directly against the GEDI L4A AGB product. | Original Section 3.1, Eq. 1, Fig. 9: the AGB*RH approach reaches R2=0.67 and RMSE≈24.6 Mg/ha with near-zero bias, clearly superior to the L4A-based ablation. | checked_against_original_txt |
| The approach reduces RMSE by 15–55% relative to five widely used openly available AGB/CHM products while being the only one available at annual time steps from 2013-2024. | Original Abstract, Section 3.3, Fig. 11, Table 7: comparison against GLAD CHM, ETH CHM, ESA CCI, Meta CHM, GTDX AGB on the same 38-site in-situ data. | 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 |
|---|---|---|---|---|
| Fig. 9, Fig. 11, Table 7 | Fig. 9 (left) compares the AGB*RH product against 38 in-situ agroforestry sites; Fig. 11 compares six AGB products against the chronosequence data (n=38) by RMSE and R2; Table 7 tabulates spatial resolution, temporal coverage, annual-update status and predictors of all six products. | R2 = 0.67, RMSE ≈ 24.6 Mg/ha, bias = 3.49 Mg/ha for the in-situ validation; RMSE reduced 15–55% vs five openly available products; product at 10 m (2017-2024) and 30 m (2013-2016). | 可用 Fig. 11 / Table 7 當作 foundation model 勝過傳統與既有開放產品的直接證據，帶出 Ch5 從 RF/CNN 走向自監督基礎模型的論述。 | 原文為 CC-BY，可標註出處後直接重用 Fig. 11 與 Table 7；或自繪精簡版六產品 RMSE 比較圖。 |
| Eq. 1, Table 4, Table 5, Table 6 | Eq. 1 定義 AGB*RH = SF × ΔRH95-10；Table 4 列出 GEDI 校準細節（SF=9、2019、979,637 樣點、95/5 訓練測試切分）；Table 5 列 Barlow Twins 超參數（編碼器 128/64/32、輸出 32 維、稀疏時序樣本 17 Landsat / 26 Sentinel-2）；Table 6 列 NNregressor 結構（6 層 128-64-32-16-8-4）。 | SF≈9；32 維表徵；NNregressor 6 層；GEDI 樣點 979,637；Barlow Twins 訓練樣本 900,000。 | 說明方法可重現性與「單年校準即凍結」設計的關鍵參數來源。 | CC-BY 可重用；建議自繪流程示意取代直接貼表。 |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| Foundation-model representations from optical time series, plus sparse GEDI, can estimate forest AGB with RMSE <25 Mg/ha. | RMSE ≈ 24.6 Mg/ha, R2 = 0.67, bias = 3.49 Mg/ha against 38 in-situ agroforestry sites (year 2022). | Abstract (p.1); Section 3.1 (p.9); Fig. 9. | True | RMSE<25 Mg/ha is the headline number; cite together with the 38-site validation context. |
| The product reduces RMSE by 15–55% compared with five openly available AGB/CHM datasets. | Comparison against GLAD CHM, ETH CHM, ESA CCI, Meta CHM, GTDX AGB on n=38 chronosequence sites. | Abstract (p.1); Section 3.3 (p.12); Fig. 11; Table 7 (p.14). | True | Most competing products are not available at annual time steps; emphasise the annual-update advantage. |
Critical Appraisal
Strengths
- Fully self-supervised foundation model, non-contrastive, computationally efficient and scalable to global coverage even under high cloudiness.
- Single-year (2019) calibration then frozen, demonstrating transferability across 12 years (2013-2024).
- Rigorous comparison against five openly available products on the same independent in-situ data.
- Open access (CC-BY) with sample data and AGB products released on Zenodo (10.5281/zenodo.15113318).
Weaknesses
- GEDI reference data has low reproducibility (R2<0.4, RMSE ~56 Mg/ha for repeat footprints), limiting the quality ceiling of the target.
- Systematic overestimation at low AGB and underestimation at high carbon stocks.
- Footprint-to-pixel geolocation mismatch not corrected.
- Validation limited to one ecoregion (Pará); cross-biome transfer of the scaling factor unverified.
| Validation quality | strong empirical validation against 38 independent in-situ sites plus four other datasets |
|---|---|
| Transferability to Taiwan | medium-high as a methodology; scaling factor and reference assumptions would need re-derivation for Taiwan subtropical/temperate forests |
| Risk of overclaiming | Do not claim the product is globally validated; authors state global processing is underway and validation is concentrated in Pará. |
與 Jacky 博論 / Review 的用途
| 博士論文 | 支撐博論把 AI 從傳統機器學習推進到自監督基礎模型，並示範 GEDI 太空 LiDAR 與光學時序融合的可規模化年度 AGB 製圖路線。 |
|---|---|
| TJFS Review | 在 TJFS review 的 Ch5 機器學習段，作為 foundation model + 自監督學習的代表性實證，銜接 Ch4 LiDAR 與 Ch6 多源融合。 |
| 可引用句候選 | 2025 年，Atzberger 等人發表的文獻中指出，以自監督的 Barlow Twins 將年度多光譜時序壓縮成低維表徵，再結合 GEDI 相對高度差作為代理目標，可在帕拉州產出 2013 至 2024 的年度 AGB 圖，對原位資料的 RMSE 約 24.6 Mg/ha，並較五個既有公開產品降低 15 至 55%。 |
| 不可用來主張 | 不可用本文宣稱基礎模型已在台灣或全球完成驗證；其驗證集中於巴西帕拉州單一生態區。 |
授權與圖表重用
| Article license | CC-BY |
|---|---|
| Figure reuse policy | REUSE_ALLOWED_WITH_ATTRIBUTION_CC_BY |
| Notes | 原文 Page 1 版權頁明確標示 0034-4257/© 2025 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)。圖表可於標註出處後重用。 |
待查核清單
- 如需公開重用 Fig. 11 / Table 7，標註 CC-BY 出處即可，無版權障礙。
- 若用於台灣情境，需重新評估縮放因子 SF 與 OGF 參考 AGB 假設是否適用亞熱帶森林。
- 可從 Zenodo (10.5281/zenodo.15113318) 取得樣本資料與 AGB 產品作進一步比對。