基於機器學習之新疆不同森林類型地上生物量遙感估算Remote sensing estimation of aboveground biomass of different forest types in Xinjiang based on machine learning
Zhou J, Zan M, Zhai L, Yang S, Xue C, Li R, Wang X｜Scientific Reports 15: 6187｜DOI: 10.1038/s41598-025-90906-3
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
AGBmachine learningforest type stratificationprovincial-scale mappingrandom forestXGBoostsupport vector machineBoruta feature selectionLandsat MODISChinaXinjiangarid zone
專討核心文獻定位
[45]
Ch5 · 機器學習
新增
以 Boruta 篩選變數加 SVM/XGBoost/RF 比較，分森林類型建模使新疆省級 AGB 估算 RF 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 recent empirical machine-learning study that directly compares SVM, XGBoost, and random forest for provincial-scale forest AGB estimation and shows that stratifying by forest type substantially improves accuracy, which supports the Ch5 machine-learning narrative.
結構式摘要｜中英文對照
| 研究問題 | 在地形複雜、地面樣本不足的新疆乾旱區，如何結合遙感與森林調查資料，並透過機器學習與森林類型分層建模，做出省級尺度的森林地上生物量估算與製圖。 In the arid and topographically complex Xinjiang region with limited ground samples, how can remote sensing and forest inventory data be combined with machine learning and forest-type stratification to estimate and map forest aboveground biomass at the provincial scale. |
|---|---|
| 資料來源 | 以 2011 年新疆森林資源清查資料（含 1335 個樣點）為地面基礎，配合 Landsat 5 影像、MODIS MOD09A1 產品、氣象與地形資料。原文針對四大森林類型建模，分別為常綠針葉林 ENF、落葉針葉林 DNF、落葉闊葉林 DBF 與混交林 MF。 The study uses the 2011 Xinjiang forest resource inventory with 1,335 sample points as the ground basis, together with Landsat 5 imagery, MODIS MOD09A1 products, and meteorological and topographic data. It models four major forest types, namely evergreen needleleaf forest (ENF), deciduous needleleaf forest (DNF), deciduous broadleaf forest (DBF), and mixed forest (MF). |
| 方法 | 先以連續生物量擴展係數法 CBEF 將樣地材積換算為 AGB，再用 Boruta 演算法從 70 個候選因子中篩選變數，各森林類型取重要性前 12 名特徵建模。資料以 7:3 切分訓練與驗證集，建立 SVM、XGBoost、RF 三種模型，以 R2、RMSE、MAE 評估，全程在 R 語言實作。 The study first converts plot volume to AGB using the continuous biomass expansion factor (CBEF) method, then applies the Boruta algorithm to screen variables from 70 candidate factors, keeping the top 12 important features per forest type. Data are split 7:3 into training and validation sets to build SVM, XGBoost, and RF models, evaluated by R2, RMSE, and MAE, with all work implemented in the R language. |
| 主要結果 | RF 模型在三種演算法中表現最佳，四種森林類型的 R2 皆大於 0.65，RMSE 介於 30.59 至 60.46 Mg/hm2、MAE 介於 24.42 至 41.75 Mg/hm2。不分森林類型的全樣本模型精度最低（R2 = 0.57）。新疆森林平均 AGB 為 152.01 Mg/hm2，空間上呈高山高、平原低的明顯異質性，高值集中於天山、阿爾泰與崑崙山三大山系。溫度與 DEM 在所有模型皆被選為特徵變數。 Among the three algorithms, RF performed best, with R2 greater than 0.65 for all four forest types, RMSE between 30.59 and 60.46 Mg/hm2, and MAE between 24.42 and 41.75 Mg/hm2. The undifferentiated all-plots model had the lowest accuracy (R2 = 0.57). The average forest AGB in Xinjiang was 152.01 Mg/hm2, with clear spatial heterogeneity that was higher in mountains and lower in plains, and high values concentrated in the Tianshan, Altai, and Kunlun mountain systems. Temperature and DEM were selected as feature variables in all models. |
| 限制 | 原文以 2011 年清查資料建模，受資料時效限制，難以套用最新遙感技術；CBEF 法基於全國尺度研究資料，對乾旱脅迫下的新疆特定區域可能引入偏差；高 AGB 區域有遙感影像飽和問題；Boruta 變數選擇雖有優勢但可能引入隨機誤差。原文建議未來整合最新調查、引入高精度多源資料與深度學習模型。 The model relies on 2011 inventory data, so its timeliness limits the use of the latest remote-sensing technology. The CBEF method is based on national-scale research data and may introduce bias for specific arid Xinjiang regions under drought stress. High-AGB areas suffer from remote-sensing signal saturation, and Boruta variable selection, while advantageous, may introduce random errors. The authors suggest integrating up-to-date surveys, high-precision multi-source data, and deep-learning models in the future. |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| Random forest outperformed SVM and XGBoost, achieving R2 greater than 0.65 for all four forest types in Xinjiang. | Original Conclusion (2) and Comparison of model results: RF had the lowest RMSE and MAE and highest R2; R2 > 0.65, RMSE 30.59-60.46 Mg/hm2, MAE 24.42-41.75 Mg/hm2 (p.10, p.12). | checked_against_original_txt |
| Stratifying models by forest type substantially improved AGB estimation accuracy relative to a single undifferentiated all-plots model. | Original 'Comparison of model results' and 'Effect of forest types' sections: the all-plots model had the lowest accuracy (R2 = 0.57, RMSE = 67.77, MAE = 44.28 Mg/hm2), and fitting accuracy ranked MF > ENF > DBF > DNF (p.10-11). | checked_against_original_txt |
| Beyond visible bands and vegetation indices, climate and topographic factors, especially temperature and DEM, were key feature variables in all models. | Original Conclusion (1) and Discussion: temperature and DEM were selected as characteristic factors in all models; texture factors and tasseled cap indices contributed to ENF, DNF, and MF (p.11-12). | 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 | Per-forest-type plot counts, mean DBH, mean tree height, and mean/max/min biomass for ENF, DNF, DBF, MF and all types. | Plots: ENF 292, DNF 425, DBF 388, MF 230, all 1335; mean biomass ENF 192.74, DNF 120.27, DBF 62.04, MF 131.8, all 128.36 Mg/hm2. | 可作為四森林類型樣本量與生物量基線的對照表，說明分型建模的資料基礎。 | Redraw a simplified summary table; do not reuse the original figures or tables publicly because the article is CC-BY-NC-ND (no derivatives). |
| Table 2 | 本研究與多篇既有研究的新疆平均 AGB 對照與相對誤差。 | 本研究省級平均 152.01 Mg/hm2；多數對照相對誤差小於 20%，僅 DBF（Populus euphratica）相對誤差約 51.85% 偏高。 | 顯示分型估算與既有研究多數一致，DBF 因受水資源影響誤差較大，可作為討論不確定性的切入點。 | Redraw a simplified comparison after confirming reuse terms; original is no-derivatives licensed. |
| Figure 6 | 四森林類型 RF 模型的觀測對預測散點驗證圖。 | RF: R2 0.65-0.75; ENF R2 0.68, MF R2 0.73, DBF R2 0.66, DNF R2 0.65; all plots R2 0.57, RMSE 67.77, MAE 44.28 Mg/hm2. | RF 仍有低值高估、高值低估現象，可作為討論機器學習估算偏差的具體例證。 | Self-draw a schematic; do not reproduce the original figure publicly under the no-derivatives licence. |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| Among SVM, XGBoost, and RF, random forest gave the most accurate and stable provincial-scale forest AGB estimates in Xinjiang. | RF R2 > 0.65 for all four forest types; RMSE 30.59-60.46 Mg/hm2; MAE 24.42-41.75 Mg/hm2; lowest errors among the three models. | Comparison of model results (p.10); Conclusion (2) (p.12). | True | Empirical full-text result; figures verified against original text. |
| Stratifying by forest type improves AGB estimation accuracy over a single undifferentiated model. | All-plots model lowest at R2 = 0.57 (RMSE 67.77, MAE 44.28 Mg/hm2); per-type models all higher; accuracy order MF > ENF > DBF > DNF. | Comparison of model results and Effect of forest types (p.10-11). | True | Confirmed against original results and conclusion sections. |
| Average forest AGB in Xinjiang is about 152 Mg/hm2 with strong spatial heterogeneity. | Average 152.01 Mg/hm2; mountainous areas generally > 160 Mg/hm2, plains typically < 80 Mg/hm2; per-type means ENF 176.57, DNF 128.69, DBF 63.74, MF 144.45 Mg/hm2. | Spatial mapping of biomass (p.9-10); Conclusion (3) (p.12). | True | Note minor in-text inconsistency in original: DBF mean stated as both 83.74 and 63.74 Mg/hm2 in different passages; flag as 待查 before quoting an exact DBF mean. |
Critical Appraisal
Strengths
- 實證比較三種機器學習演算法，並以分森林類型建模驗證精度提升，結論清楚。
- 結合 Boruta 變數選擇與多源遙感、氣象、地形資料，方法流程完整可複現。
- 達成新疆省級尺度 AGB 製圖，並與多篇既有研究做相對誤差對照。
Weaknesses
- 以 2011 年清查資料建模，時效性受限，未能套用最新遙感技術。
- CBEF 法基於全國資料，對乾旱脅迫下特定區域可能偏差；高 AGB 區有影像飽和。
- 原文文中 DBF 平均 AGB 同時出現 83.74 與 63.74 Mg/hm2 兩個數值，存在內部不一致，引用須謹慎。
| Validation quality | good; 7:3 train-validation split with R2/RMSE/MAE and 10-fold cross-validation for hyperparameter tuning |
|---|---|
| Transferability to Taiwan | moderate to high as a method reference; the forest-type stratification and Boruta plus RF workflow transfers well, though arid-zone tree species differ from Taiwan subtropical forests |
| Risk of overclaiming | Do not present the DBF mean AGB as a single exact value given the in-text inconsistency, and do not generalize the Xinjiang arid-zone accuracy directly to humid subtropical forests. |
與 Jacky 博論 / Review 的用途
| 博士論文 | 支撐博論在機器學習估算 AGB 的方法選擇與森林類型分層策略，並提供 RF 在省級尺度製圖的實證依據。 |
|---|---|
| TJFS Review | 在 TJFS review 的 Ch5 機器學習章節可作為 SVM/XGBoost/RF 三模型實證比較與分型建模提升精度的具體案例錨點。 |
| 可引用句候選 | 2025 年，Zhou 等人發表的文獻中指出，於新疆乾旱區結合 Boruta 變數選擇與三種機器學習模型，隨機森林表現最佳且分森林類型建模能顯著提升地上生物量估算精度。 |
| 不可用來主張 | 不要以本篇單獨宣稱機器學習在所有森林環境皆優於傳統方法，亦不要直接把乾旱區精度結果套用到台灣亞熱帶森林。 |
授權與圖表重用
| Article license | CC-BY-NC-ND-4.0 |
|---|---|
| Figure reuse policy | DO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED |
| Notes | 原文授權頁明示 Creative Commons Attribution-NonCommercial-NoDerivatives 4.0；可非商業分享但禁止改作，圖件不可改繪後公開分享，引用須標來源。 |
待查核清單
- 回原文釐清 DBF 平均 AGB 究竟為 83.74 或 63.74 Mg/hm2（文中兩處不一致），確認後再做精確引用。
- 確認 CC-BY-NC-ND 無改作條款下圖表呈現方式，需自繪示意而非改繪原圖。
- 如需詳列 70 個候選變數與最終 12 個特徵，回 Table 3 萃取完整清單。