← 回文獻卡索引

在亞熱帶常綠闊葉林中以遙測影像萃取之物候資訊製圖森林地上生物量Mapping Forest Aboveground Biomass with Phenological Information Extracted from Remote Sensing Images in Subtropical Evergreen Broadleaf Forests

Yang, P., Long, J., Lin, H., Zhang, T., Ye, Z. & Liu, Z.｜Remote Sensing 17(9): 1599｜DOI: 10.3390/rs17091599

狀態：AI_DRAFT_FROM_REVIEW｜分級：A｜閱讀深度：FULL_TEXT_CHECKED｜Jacky 審核：False

森林數位孿生底層致能

讀學習筆記 →

aboveground biomassphenological featuresspectral saturationSentinel-2 time seriesvegetation indicesrandom forestTIMESATChinasubtropicalevergreen broadleaf forest

專討核心文獻定位

[86] Ch8 · 研究缺口 新增
Yang et al. · 2025
以 Sentinel-2 時序萃取物候特徵，可緩解常綠闊葉林光譜飽和問題並提升地上生物量估算精度

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 belongs in the research-gap chapter because it targets a persistent and unsolved problem in optical AGB mapping, namely spectral signal saturation in structurally complex subtropical evergreen broadleaf forests, and proposes phenological features from multi-season Sentinel-2 time series as a way to break past the saturation threshold, which frames a concrete methodological gap that a forest digital twin and richer temporal data layers could close.

結構式摘要｜中英文對照

研究問題
在常綠闊葉林因高葉綠素含量造成光譜飽和的情況下，能否利用生長季與非生長季之間的光譜反射變化所萃取的物候特徵，提升地上生物量估算精度？
Under spectral saturation caused by high leaf chlorophyll content in evergreen broadleaf forests, can phenological features extracted from spectral reflectance changes between the growing and non-growing seasons improve aboveground biomass estimation accuracy?
資料來源
研究區位於中國湖南省懷化市沅陵縣的常綠闊葉林，屬濕潤亞熱帶季風氣候，年均溫 16.7 度，森林覆蓋率 76.19%、常綠闊葉林占 55.74%。2021 年 12 月以隨機分層抽樣採集 100 個地面樣區，樣區大小 25.82 公尺乘 25.82 公尺，以 RTK-GNSS 定位並量測胸徑 5 公分以上立木的樹高、胸徑與冠幅，材積以中國二元立木材積表公式計算，地上生物量則以生物量擴展因子（BEF）連續函數法換算。樣區生物量平均為 139.97 t/hm²。遙測資料為 Sentinel-2 L2A 時序，以 2020 與 2022 年影像合成以填補雲與降雨造成的資料缺口，選用九個關鍵波段並重採樣至 10 公尺。
The study area is evergreen broadleaf forest in Yuanling County, Huaihua City, Hunan Province, China, in a humid subtropical monsoon climate with a mean annual temperature of 16.7 degrees, forest cover of 76.19% and evergreen broadleaf forest at 55.74%. In December 2021, 100 ground sample plots were collected by random stratified sampling, each 25.82 by 25.82 meters, positioned with RTK-GNSS, measuring height, DBH, and crown width of trees with DBH at least 5 cm. Tree volume was computed with China's binary volume formula and AGB via a biomass expansion factor (BEF) continuous-function method. Mean plot AGB is 139.97 t/hm2. Remote sensing input is a Sentinel-2 L2A time series, compositing 2020 and 2022 images to fill cloud and rainfall gaps, using nine key bands resampled to 10 meters.
方法
以非生長季（1至3月、11至12月）與生長季（4至10月）的光譜反射差異建立物候特徵方法。採四個時序植生指數 NDVI、EVI2、NDPI、IRECI 萃取物候特徵（PFs），以 Jönsson 與 Eklundh 的非對稱高斯（AG）模型擬合時序曲線，並用 TIMESAT 3.3 軟體以動態閾值法決定季節起訖（SOS/EOS），每條時序萃取 13 個物候特徵。設計六組變數集：A 為基線（光譜波段、植生指數、紋理特徵共 311 個變數），B 至 E 各加入單一植生指數的物候特徵（324 個變數），F 整合四個植生指數全部物候特徵（363 個變數）。以距離相關（DC）係數評估敏感度並做前向特徵選擇，再用多元線性回歸（MLR）、K 近鄰（KNN）、支援向量機（SVM）與隨機森林（RF）四種模型估算，以留一交叉驗證（LOOCV）評估。
The method builds phenological features from spectral reflectance differences between the non-growing season (Jan to Mar, Nov to Dec) and the growing season (Apr to Oct). Four time-series vegetation indices (NDVI, EVI2, NDPI, IRECI) are used to extract phenological features. Time-series curves are fitted with the Asymmetric Gaussian (AG) model of Jönsson and Eklundh, and TIMESAT 3.3 determines start and end of season (SOS/EOS) by a dynamic threshold method, extracting 13 phenological features per series. Six variable sets are designed: A is the baseline (spectral bands, vegetation indices, texture features, 311 variables), B to E each add one index's phenological features (324 variables), and F integrates all four indices' phenological features (363 variables). Distance correlation (DC) coefficients evaluate sensitivity with forward feature selection, and four models (MLR, KNN, SVM, RF) estimate AGB, validated by leave-one-out cross-validation.
主要結果
九個波段在生長季與非生長季之間於 740 至 1610 奈米範圍呈現顯著光譜差異。物候特徵的最高距離相關係數為 0.57，明顯高於基線特徵集的 0.44。NDVI 與 NDPI 的季節變化比 EVI2 與 IRECI 更能反映生物量累積。整合四個植生指數全部物候特徵的變數集 F，rRMSE 介於 21.01% 至 25.06%、R² 介於 0.40 至 0.58。最佳結果為 SVM 配變數集 F（R²=0.58、rRMSE=21.01%）。NDVI 與 NDPI 萃取的物候特徵（R²=0.52、0.51）明顯優於 EVI2 與 IRECI（皆 R²=0.33）。隨特徵集品質提升，模型選擇對精度的影響逐漸縮小。
The nine bands show significant spectral differences between growing and non-growing seasons within 740 to 1610 nm. The maximum DC coefficient for phenological features is 0.57, clearly higher than the baseline feature set's 0.44. Seasonal NDVI and NDPI changes reflect biomass accumulation better than EVI2 and IRECI. Variable set F, integrating all four indices' phenological features, yields rRMSE of 21.01% to 25.06% and R-squared of 0.40 to 0.58. The best result is SVM with set F (R-squared 0.58, rRMSE 21.01%). Phenological features from NDVI and NDPI (R-squared 0.52 and 0.51) clearly outperform EVI2 and IRECI (both R-squared 0.33). As feature-set quality improves, the influence of model selection on accuracy gradually diminishes.
限制
研究僅以單一研究區（湖南沅陵）100 個樣區驗證，BEF 連續函數與二元材積公式為中國本土係數，遷移到其他林型或地區需重新校正。整體 R² 最高僅 0.58，仍屬中等，且地面生物量資料因隱私不公開。光學遙測在高生物量區仍有飽和殘留，物候特徵可緩解但未完全消除。作者於結論指出未來將深入分析不同光譜波段的反射變化，以從反射時序型態中找出更敏感的物候指標。
The study validates on a single area (Yuanling, Hunan) with 100 plots; the BEF continuous function and binary volume formula use China-specific coefficients that would need recalibration for other forest types or regions. The best overall R-squared is only 0.58, still moderate, and ground biomass data are not publicly released due to privacy. Optical sensing still leaves residual saturation at high biomass; phenological features mitigate but do not fully eliminate it. The authors note future work will analyze reflectance variation across spectral bands to find more sensitive phenological indicators from reflectance time-series patterns.

Key Findings

發現證據確定性
Phenological features from multi-season Sentinel-2 time series are more sensitive to AGB than conventional spectral, vegetation-index, and texture features in evergreen broadleaf forests.Abstract and Section 4.2: the maximum DC coefficient of phenological features is 0.57 versus 0.44 for the Bs + VI + TFs baseline set.checked_against_original_txt
NDVI- and NDPI-derived phenological features outperform EVI2- and IRECI-derived ones for AGB mapping.Results and Conclusion: NDVI R-squared 0.52 and NDPI R-squared 0.51 versus EVI2 R-squared 0.33 and IRECI R-squared 0.33; NDVI and NDPI track phenological change more directly.checked_against_original_txt
Integrating phenological features from all four indices (variable set F) gives the best accuracy and mitigates spectral saturation, especially the underestimation of high values and overestimation of low values.Abstract and Table 4: set F yields rRMSE 21.01% to 25.06% and R-squared 0.40 to 0.58, with SVM plus set F reaching R-squared 0.58 and rRMSE 21.01%.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 4R-squared, RMSE, and rRMSE for MLR, SVM, KNN, and RF across the six variable sets A to F.Best: SVM set F R-squared 0.58, RMSE 29.41 t/hm2, rRMSE 21.01%; baseline MLR set A R-squared 0.34, rRMSE 26.21%; KNN consistently lowest.This is the cleanest evidence that adding temporal phenological layers on top of single-date optical features systematically pushes optical AGB estimation past saturation, which supports a temporally enriched data layer in a forest digital twin.Even under CC BY, redraw a self-made bar chart comparing R-squared across variable sets and models rather than reproducing the original table, and cite the source.
Table 3Extracted phenological metrics (SOS, EOS, LOS, BV, MAXMUN, AP) for the EVI2, IRECI, NDPI, and NDVI time series.NDVI MAXMUN highest at 0.86 and BV 0.58; EVI2 MAXMUN lowest at 0.47; IRECI highest amplitude AP 0.55 and lowest BV 0.17; SOS day-of-year 56 to 68 across indices.Shows why NDVI and NDPI carry stronger AGB signal: their seasonal amplitude and baseline contrast track canopy dynamics more clearly than EVI2 and IRECI.Cite numbers in text or redraw a simplified comparison; confirm figure/table labelling before any public reuse.
Figure 8Distance correlation coefficients between candidate features and AGB, ranked, comparing the baseline Bs + VI + TFs set against phenological-feature sets.Phenological-feature DC up to 0.57 (range 0.45 to 0.57) versus baseline maximum 0.44; top three features all phenology-related.The sensitivity ranking is the core justification for prioritizing temporal phenological features over static spectral features when designing the optical layer of a carbon-monitoring system.Redraw a simplified DC comparison chart with citation rather than reproducing the original plot.

Extracted Evidence Table

可支撐主張指標或結果原文位置可引用備註
Spectral signal saturation in subtropical evergreen broadleaf forests is a persistent barrier for optical AGB mapping.Stated as the central challenge driven by complex canopy structure, stand heterogeneity, and high leaf chlorophyll content; gap between growing and non-growing season reflectance shrinks as AGB increases (mean difference 0.18 at low AGB versus 0.11 above 200 t/hm2).Abstract, Introduction, and Section 4.1 (Figure 6).TrueStrong saturation gap statement for the research-gap chapter; frame the residual saturation honestly rather than as fully solved.
Adding phenological features improves AGB accuracy and reduces high-value underestimation and low-value overestimation.Phenological DC 0.57 versus baseline 0.44; set F SVM R-squared 0.58 and rRMSE 21.01% versus baseline MLR R-squared 0.34 and rRMSE 26.21%; scatter plots show sets D, E, F reduce under- and over-estimation.Abstract, Section 4.3, Table 4, Figure 10.TrueDescribe the gain as a clear but moderate improvement; best R-squared is still only 0.58 in a single study area.

Critical Appraisal

Strengths

Weaknesses

Validation qualityLeave-one-out cross-validation with standard regression metrics (R-squared, RMSE, rRMSE) on 100 plots; no independent external validation region beyond the study area.
Transferability to TaiwanHigh in concept and medium in practice; Taiwan's subtropical evergreen broadleaf forests share the saturation problem and seasonal phenology, but the BEF, binary volume formula, and species-specific coefficients would need Taiwan-specific recalibration.
Risk of overclaimingDo not claim phenological features fully solve spectral saturation; they reduce it and improve accuracy to a moderate R-squared of 0.58 in one subtropical area, not across all forest types.

與 Jacky 博論 / Review 的用途

博士論文Supports the dissertation argument that temporal and phenological information is a necessary data layer for breaking optical saturation in dense subtropical canopies, which a forest digital twin should ingest as continuously updated time-series rather than single-date imagery.
TJFS ReviewProvides a concrete research-gap and methodological citation for the TJFS review, showing that multi-season phenological features from free Sentinel-2 data measurably improve subtropical evergreen broadleaf AGB estimation and that NDVI and NDPI outperform EVI2 and IRECI.
可引用句候選2025 年，Yang 等人發表的文獻中指出，在亞熱帶常綠闊葉林中，從 Sentinel-2 多季時序萃取的物候特徵對地上生物量的敏感度（距離相關係數 0.57）明顯高於傳統光譜與紋理特徵（0.44），並可緩解光譜飽和、把整合四個植生指數物候特徵的最佳估算精度提升至 R² 0.58、rRMSE 21.01%。
不可用來主張Do not use this paper as evidence for tropical or coniferous forests, for LiDAR or SAR-based methods, or as proof that optical saturation is fully eliminated; it tests subtropical evergreen broadleaf forest with optical time series only and reaches moderate accuracy.

授權與圖表重用

Article licenseCC BY 4.0
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文版權頁明列 Creative Commons Attribution (CC BY) 4.0 授權（MDPI 開放取用）。CC BY 允許在註明出處下重用，但仍先以自繪詮釋替代，正式公開前再確認圖表標註與授權連結。

待查核清單