← 回文獻卡索引

結合遷移學習的深度神經網路於北方林以 Sentinel-2 影像估算森林變數Deep Neural Networks with Transfer Learning for Forest Variable Estimation Using Sentinel-2 Imagery in Boreal Forest

Astola, Seitsonen, Halme, Molinier & Lönnqvist｜Remote Sensing 13(12): 2392｜DOI: 10.3390/rs13122392

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 →

forest variable estimationgrowing stock volumeboreal forestdeep neural networktransfer learningSentinel-2random forestFinlandboreal

專討核心文獻定位

[72] Ch4 · LiDAR 新增
Astola et al. · 2021
Sentinel-2 加影像取樣窗與地形特徵的 DNN 可估算北方林材積，遷移學習在野外樣區少於 250 個時最有幫助

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 source for using deep neural networks and transfer learning with open optical Sentinel-2 imagery to estimate forest structural variables, complementing the LiDAR chapter by showing how to cut field-data cost when LiDAR or CHM are unavailable.

結構式摘要｜中英文對照

研究問題
在北方林中，深度神經網路結合 Sentinel-2 影像、影像取樣窗、地形與冠層高度等特徵，能否準確估算材積等森林結構變數，遷移學習又能否減少所需的野外調查資料？
In boreal forest, can deep neural networks combining Sentinel-2 imagery, image sampling windows, topography and canopy height features accurately estimate growing stock volume and other forest structural variables, and can transfer learning reduce the amount of field data required?
資料來源
研究區為芬蘭中部四個林業區，總面積 68,315 平方公里。野外參考資料由芬蘭森林中心提供，於 2016 至 2017 年以三種樣區半徑取得；資料含材積、胸高直徑、樹高、斷面積、年齡與株數等。光學影像為 2017 年夏季 22 幅 Sentinel-2 Level-1C 產品，並加入影像與太陽角度、地形與冠層高度模型等輔助特徵。
The study area covers four forestry districts in Central Finland with a total area of 68,315 km2. Field reference data were provided by the Finnish Forest Centre, acquired during 2016 to 2017 with three plot radii, and included stem volume, diameter, height, basal area, age and stem number. Optical inputs were 22 Sentinel-2 Level-1C products from summer 2017, plus auxiliary features such as imaging and sun angles, topography and a canopy height model.
方法
以全連接深度神經網路（DNN）對材積等四個森林變數及其樹種別分量進行回歸；輸入特徵組合 Sentinel-2 波段、影像與太陽角度、地形與冠層高度模型，並採用 3 乘 3 像素的影像取樣窗。系統性比較不同輸入特徵、DNN 層數、訓練樣本數與取樣窗大小，並與隨機森林（RF）對照；遷移學習以在 Central Ostrobothnia 預訓練的模型，再用其他三區資料微調。
Fully connected deep neural networks were used to regress four forest variables and their species-specific components. Input features combined Sentinel-2 bands, imaging and sun angles, topography and a canopy height model, with a 3 by 3 pixel image sampling window. The study systematically compared input features, DNN depth, training sample size and sampling window size, and benchmarked against random forest. Transfer learning fine-tuned a model pre-trained on Central Ostrobothnia using data from the three other districts.
主要結果
原文確認：不含冠層高度模型時，使用 RGB 與近紅外波段加上影像與太陽角度與地形特徵的模型在樣區層級最佳，RMSE% 為 42.6%、偏差絕對值 0.8%。納入冠層高度模型特徵後相對 RMSE% 降至 28.6 至 30.7%，但相對偏差絕對值升至 0.9 至 4.0%。最佳取樣窗為 3 乘 3 像素；兩到三個隱藏層即達最佳，再加深僅微幅改善並增加變異。遷移學習主要在訓練樣區少於 250 個時有益。DNN 與 RF 的表現差異很小。
The original text confirms that, leaving out canopy height model features, the model using RGB and NIR bands plus imaging and sun angles and topography achieved the best plot-level accuracy with RMSE% = 42.6% and |BIAS%| = 0.8%. Including canopy height model features lowered relative RMSE% to 28.6 to 30.7% but raised the absolute relative bias to 0.9 to 4.0%. A 3 by 3 pixel sampling window was optimal, and two to three hidden layers gave the best results with only marginal gains from deeper networks. Transfer learning was beneficial mainly with fewer than 250 field plots, and the performance difference between DNN and random forest was marginal.
限制
原文指出，野外樣區的實用最少數量約落在 200 到 330 個之間；材積估算在材積高於約 300 立方公尺每公頃時出現飽和效應；訓練在最深的網路（16 或 20 層）時失敗率上升並出現梯度消失。遷移學習在訓練資料充足時相對從頭訓練並無優勢。
The text notes that the practical minimum number of field plots is roughly between 200 and 330; stem volume estimation shows a saturation effect above about 300 m3/ha; training failed more frequently with the deepest networks (16 or 20 layers) due to vanishing gradients. Transfer learning offered no advantage over training from scratch when sufficient training data were available.

Key Findings

發現證據確定性
Adding image sampling windows, sun and imaging angles, and topography to Sentinel-2 reflectance significantly improves DNN-based growing stock volume prediction.Conclusions: the proposed image sampling and feature concept gained about 8 percentage points improvement in RMSE% for total stem volume; best non-CHM plot-level result RMSE% = 42.6%, |BIAS%| = 0.8% (Abstract).checked_against_original_txt
Transfer learning helps mainly when field reference data are scarce, and DNN performance is close to random forest.Section 4.5 and Conclusions: transfer learning is beneficial mainly with fewer than 250 field plots; DNN and RF performance differences were marginal (Section 4.4, 5.3).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, Table 8, Figure 9Table 1 lists per-district means, standard deviations and plot counts of forest variables; Table 8 reports RMSE%, BIAS% and R2 for four forest variables; Figure 9 compares transfer-learning versus from-scratch RMSE% across training set sizes.Total area 68,315 km2; field plot counts 374 / 387 / 362 / 712 across the four districts; transfer learning and from-scratch models practically equivalent at 250 or more field plots.Use this as evidence that open optical data plus DNN can substitute for costly LiDAR/field campaigns at the boreal scale, while transfer learning frames the field-data minimization argument relevant to Taiwan.Figures are CC BY 4.0 and may be reused with attribution; still prefer a redrawn summary chart for the seminar.

Extracted Evidence Table

可支撐主張指標或結果原文位置可引用備註
DNN with engineered Sentinel-2 features can estimate boreal forest stem volume at useful accuracy without LiDAR.Best non-CHM plot-level RMSE% = 42.6%, |BIAS%| = 0.8%; with CHM features RMSE% = 28.6 to 30.7%.Abstract; Section 4.1 to 4.2; Conclusions (Section 6).TrueCite as optical-only and optical-plus-CHM accuracy benchmark in boreal forest; do not generalize uncritically to tropical or subtropical biomes.
Transfer learning reduces the field-data requirement for new geographic regions.Beneficial mainly with fewer than 250 field plots; equivalent to from-scratch models at 250 or more plots.Section 4.5; Section 5.4; Conclusions.TrueUse to support arguments about cutting field campaign cost; note the benefit vanishes with abundant data.

Critical Appraisal

Strengths

Weaknesses

Validation qualitystrong: test-set evaluation, multiple repeated runs, and RF benchmark
Transferability to Taiwanmoderate as a method blueprint; accuracy and saturation must be re-validated for subtropical Taiwan forests
Risk of overclaimingDo not claim optical DNN replaces LiDAR for high-biomass or carbon-grade accuracy; the optical-only RMSE% is still large and saturates.

與 Jacky 博論 / Review 的用途

博士論文Supports the dissertation's middle layer on machine-learning estimation from open satellite data and on minimizing field campaigns through transfer learning.
TJFS ReviewProvides a quantified boreal benchmark for optical DNN forest-variable estimation and a concrete transfer-learning threshold for the TJFS review.
可引用句候選2021 年，Astola 等人發表的文獻中指出，結合 Sentinel-2 影像取樣窗與地形特徵的深度神經網路可在北方林估算材積，且遷移學習在野外樣區少於 250 個時最能減少所需的實地調查。
不可用來主張Do not use this paper to claim that optical satellite DNN achieves carbon-grade accuracy or that it generally outperforms random forest.

授權與圖表重用

Article licenseCC BY 4.0
Figure reuse policyREUSE_ALLOWED_WITH_ATTRIBUTION_CC_BY
NotesMDPI open access, Creative Commons Attribution (CC BY) 4.0 license stated on first page; figures reusable with attribution.

待查核清單