以深度學習基礎模型結合 LiDAR 估算城市森林生物量Estimating Urban Forests Biomass with LiDAR by Using Deep Learning Foundation Models
Liu, H.; Mou, C.; Yuan, J.; Chen, Z.; Zhong, L.; Cui, X.｜Remote Sensing 16(9): 1643｜DOI: 10.3390/rs16091643
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
urban forestvegetation biomasscarbon sinkfoundation modelLiDARdeep learning foundation modelSegment Anything ModelCLIPInformerallometric growth equationtime-series predictionUnited Statesurban
專討核心文獻定位
[77]
Ch5 · 機器學習
新增
提出 3D-CiLBE 流程，串接 SAM 分割、CLIP 物種辨識與 Informer 時序預測，從城市 LiDAR 估算植被生物量，分割 mIoU 達 0.94、物種辨識 92.72%，較 2D 方法誤差降低 8.2%
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 Ch5 core source demonstrating how multiple deep learning foundation models (SAM, CLIP, Informer) can be assembled into an end-to-end pipeline to estimate and predict urban forest biomass from LiDAR, representing the shift from single classical models toward foundation-model-based, low-cost, transferable biomass estimation.
結構式摘要｜中英文對照
| 研究問題 | 在植被分散、三維地形複雜且物種多變的城市森林環境中，能否利用深度學習基礎模型結合高解析度 LiDAR 與 OSM 資料，準確且低成本地估算並長期預測城市植被生物量？ In urban forest environments with dispersed vegetation, complex 3D terrain, and high species variability, can deep learning foundation models combined with high-resolution LiDAR and OSM data accurately and cheaply estimate and predict urban vegetation biomass over the long term? |
|---|---|
| 資料來源 | 研究區為美國八個氣候條件不同的城市（Boston、Chicago、Denver、Detroit、Houston、Las Vegas、Los Angeles、Miami），蒐集 2012 至 2020 年的 LiDAR（USGS）與對應 OSM（OpenStreetMap）資料，並加入單株樹木 LiDAR 影像（GRO）。LiDAR 統一重採樣至 1 m 解析度、裁切為 224×224 影像。共標註 1160 張影像（每城 145 張），以 5:1 切成訓練集 967 張與測試集 193 張。 The study area covers eight U.S. cities with differing climates (Boston, Chicago, Denver, Detroit, Houston, Las Vegas, Los Angeles, Miami). LiDAR (USGS) and corresponding OSM (OpenStreetMap) data from 2012 to 2020 were collected, plus single-tree LiDAR imagery (GRO). LiDAR was resampled to a uniform 1 m resolution and cropped to 224×224 images. A total of 1160 labeled images (145 per city) were split 5:1 into 967 training and 193 test tiles. |
| 方法 | 提出 3D-CiLBE（3D City Long-term Biomass Estimation）方法，由三個模組構成。第一是 LiDAR-SAM，將 Segment Anything Model（SAM）改造後用於 LiDAR 影像植被區分割，並以 OSM 道路網遮罩與 Area Prompt 輔助提示，再用 RangeNet++ 做 3D 與 2D 投影還原以擷取單株樹高與胸徑（DBH）。第二是 MLiDAR-CLIP，修改 CLIP 並導入多視角投影與隨機邊界遮罩，做單株樹木物種辨識。第三是 St-Informer，先以異速生長方程式 Bio = a × Height × DBH^b 計算每期生物量，再用 Informer 模型做長期時序預測。訓練與推論可在 CPU 上完成，凍結多數編碼器層以降低成本。 The paper proposes 3D-CiLBE (3D City Long-term Biomass Estimation), comprising three modules. First, LiDAR-SAM adapts the Segment Anything Model (SAM) to segment vegetation areas in LiDAR imagery, aided by an OSM road-network mask and Area Prompt cues, and uses RangeNet++ for 3D-to-2D projection and reconstruction to extract per-tree height and DBH. Second, MLiDAR-CLIP modifies CLIP with multi-viewpoint projection and random boundary masking to recognize single-tree species. Third, St-Informer first computes biomass via the allometric growth equation Bio = a × Height × DBH^b, then uses the Informer model for long-term time-series prediction. Training and inference can run on CPU, and most encoder layers are frozen to lower cost. |
| 主要結果 | 在城市綠化區辨識任務，3D-CiLBE 的 LiDAR-SAM 分割達最佳 mIoU 0.94、OA 0.98，mIoU 較 PointNet++ 高 9.0%、較 SAM 高 3.0%。物種辨識上 MLiDAR-CLIP 平均辨識準確率 92.72%，較 SVM 高 11.5%、較 PointCLIP 高 4.8%。生物量估算的均方誤差 MSE 為 0.045 kg/m^2，較 2D 方法平均降低 8.2%，2017 年八城平均 DR 為 1.8。St-Informer 時序預測 MSE 約 0.05 kg/m^2、R2 0.73、RMSE 0.22 kg/m^2，與線性回歸的 MSE 差距小於 0.06 kg/m^2。Chicago 案例研究 MSE 0.041 kg/m^2、RMSE 0.21 kg/m^2、R2 0.74，且 Informer 較線性模型預測 MSE 降低 28.8%。 For urban greening-area identification, the LiDAR-SAM module of 3D-CiLBE achieves an optimal mIoU of 0.94 and OA of 0.98, with mIoU 9.0% higher than PointNet++ and 3.0% higher than SAM. For species recognition, MLiDAR-CLIP reaches a mean accuracy of 92.72%, 11.5% higher than SVM and 4.8% higher than PointCLIP. Biomass estimation reaches an MSE of 0.045 kg/m^2, 8.2% lower on average than 2D methods, with an average DR of 1.8 across eight cities in 2017. St-Informer time-series prediction yields MSE ≈ 0.05 kg/m^2, R2 0.73, RMSE 0.22 kg/m^2, and the MSE gap with linear regression is below 0.06 kg/m^2. The Chicago case study reports MSE 0.041 kg/m^2, RMSE 0.21 kg/m^2, R2 0.74, with Informer reducing predicted MSE by 28.8% relative to the linear model. |
| 限制 | 原文 3.6 節指出，在氣象變化複雜的城市 LiDAR 影像品質可能偏低、難達高精度估算；在易受天然災害、空間佈局不規則的城市，模型難以一致估算。緩解方式為減少凍結層數（成本上升換取精度）或蒐集更多時序資料。此外因缺乏官方城市植被生物量資料集，真值是透過多源遙測法推得而非直接實地量測，存在參考基準不確定性。 Section 3.6 notes that in cities with complex meteorological variation, LiDAR image quality can be low and high-precision estimation is hard; in cities prone to natural disasters with irregular spatial layout, consistent estimation is difficult. Mitigations include reducing the number of frozen layers (higher cost for higher accuracy) or collecting more time-series data. Moreover, because no official urban vegetation biomass dataset exists, the ground truth is derived through multi-source remote sensing methods rather than direct field measurement, introducing reference-baseline uncertainty. |
Key Findings
| 發現 | 證據 | 確定性 |
|---|---|---|
| Assembling three deep learning foundation models (SAM, CLIP, Informer) into one pipeline enables accurate, low-cost, end-to-end urban forest biomass estimation and long-term prediction from LiDAR and OSM data. | Original Abstract and Section 2.2: 3D-CiLBE chains LiDAR-SAM segmentation, MLiDAR-CLIP species recognition, and St-Informer prediction; training and inference run on CPU. | checked_against_original_txt |
| Adapting SAM for LiDAR (LiDAR-SAM) with OSM-mask and Area-Prompt cues markedly improves urban vegetation segmentation over PointNet++ and vanilla SAM. | Original Section 3.2.1, Table 3: LiDAR-SAM mIoU 0.94, OA 0.98, 9.0% higher mIoU than PointNet++ and 3.0% higher than SAM; ablation adds 2.04% (Area Prompt) and 4.05% (OSM mask). | checked_against_original_txt |
| 3D LiDAR-based estimation outperforms 2D remote-sensing methods for urban biomass, lowering error by up to 8.2% with an average decision ratio (DR) of 1.8. | Original Abstract and Section 3.2.3, Table 4: biomass MSE 0.045 kg/m^2, 8.2% lower than 2D method, average DR 1.8 in 2017 across eight cities. | checked_against_original_txt |
| The whole pipeline trains far fewer parameters than the base foundation models, demonstrating a low-cost, scalable design. | Original Section 3.5, Table 6: LiDAR-SAM trains 14M parameters / 2.51 GFLOPs vs SAM(ViT-H) 636M / 81.34 GFLOPs; MLiDAR-CLIP 12M / 1.67 GFLOPs vs CLIP up to 1600M. | 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 3 | 比較 PointNet++、SAM 與 LiDAR-SAM 在城市綠地 LiDAR 影像分割上的 IoU-P、mIoU、OA、Recall、Kappa 等指標。 | LiDAR-SAM mIoU 0.94、OA 0.98、IoU-P 0.93、Kappa 0.92；mIoU 較 PointNet++ 高 9.0%、較 SAM 高 3.0%。 | 可作為 Ch5 論述「基礎模型微調勝過傳統點雲模型」的直接證據，帶出 SAM 改造用於 LiDAR 的價值。 | 原文 CC-BY，可標註出處後重用 Table 3；或自繪精簡三模型 mIoU/OA 比較長條圖。 |
| Table 4 | 列出 Boston 等八城的 2D 法、3D-CiLBE 估值、真值、與真值差距及決策比 DR。 | 平均 DR 1.8；3D-CiLBE 誤差較 2D 平均低 8.2%；單城 DR 介於 1.4 至 3.4。 | 支撐「三維資訊提升城市生物量估算精度」的論點，可串接 Ch4 LiDAR 與 Ch6 多源融合。 | CC-BY 可重用；建議自繪八城 DR 條形圖避免直接貼整表。 |
| Table 6 | 比較 SAM 各版本、CLIP 各版本與本文 LiDAR-SAM、MLiDAR-CLIP 的參數量（M）與 GFLOPs。 | LiDAR-SAM 14M / 2.51 GFLOPs；MLiDAR-CLIP 12M / 1.67 GFLOPs；SAM(ViT-H) 636M / 81.34 GFLOPs；CLIP 最大 1600M。 | 佐證「低成本可規模化」設計，呼應基礎模型凍結多數層以節省算力的策略。 | CC-BY 可重用；可自繪參數量對數軸對比圖凸顯量級差異。 |
Extracted Evidence Table
| 可支撐主張 | 指標或結果 | 原文位置 | 可引用 | 備註 |
|---|---|---|---|---|
| Foundation models (SAM/CLIP/Informer) can be assembled to estimate urban forest biomass from LiDAR with low error. | Biomass MSE 0.045 kg/m^2, 8.2% lower than 2D methods; segmentation mIoU 0.94; species recognition 92.72%. | Abstract (p.1); Section 3.2 (p.10-11); Table 3; Table 4. | True | Headline numbers; cite together with the eight-US-city, 2012-2020 study context. |
| Informer-based time-series prediction marginally outperforms linear regression for long-term urban biomass forecasting. | St-Informer MSE ≈ 0.05 kg/m^2, R2 0.73, RMSE 0.22 kg/m^2; Chicago case Informer reduces predicted MSE by 28.8% vs linear. | Section 3.2.4 (p.12); Section 3.4.3 (p.17); Fig. 6; Fig. 13. | True | Authors note the MSE gap with linear regression is below 0.06 kg/m^2, so for rough estimates linear regression may suffice. |
Critical Appraisal
Strengths
- End-to-end pipeline assembling three complementary foundation models (SAM, CLIP, Informer) for a hard, dispersed-vegetation urban setting.
- Low-cost, CPU-trainable design with most encoder layers frozen; far fewer trained parameters than base models.
- Multi-city (eight US cities, 2012-2020) and multi-task evaluation with comparative, ablation and case-study experiments.
- Open access (CC-BY) with public code at github.com/ForestryIIP/3DCiLBE.
Weaknesses
- No official urban biomass dataset exists; ground truth is derived from multi-source remote sensing rather than direct field measurement, weakening the reference baseline.
- Performance may degrade in cities with complex meteorology or irregular spatial layout (authors' own Section 3.6).
- Validation confined to US cities; transfer to other climates/species unverified.
- Time-series gain over linear regression is small (MSE gap <0.06 kg/m^2).
| Validation quality | moderate-to-strong: multi-city, multi-task comparison and ablation, but ground truth indirectly derived rather than field-measured |
|---|---|
| Transferability to Taiwan | medium as a methodology; allometric parameters, species set and reference biomass would need re-derivation for Taiwan urban/subtropical forests |
| Risk of overclaiming | Do not present the biomass ground truth as field-measured; it is inferred from multi-source remote sensing. Do not claim global validation; experiments are US-only. |
與 Jacky 博論 / Review 的用途
| 博士論文 | 支撐博論把 AI 軸線從單一傳統模型推進到多基礎模型協作，並示範 LiDAR 加 OSM 多源資料在城市尺度的低成本生物量估算與時序預測路線。 |
|---|---|
| TJFS Review | 在 TJFS review 的 Ch5 機器學習段，作為 foundation model（SAM/CLIP/Informer）組合應用於城市森林生物量的代表性實證，銜接 Ch4 LiDAR 與 Ch6 多源融合。 |
| 可引用句候選 | 2024 年，Liu 等人發表的文獻中指出，把 Segment Anything Model、CLIP 與 Informer 三個深度學習基礎模型組成 3D-CiLBE 流程，可從城市 LiDAR 與 OSM 資料估算並長期預測植被生物量，植被分割 mIoU 達 0.94、物種辨識準確率 92.72%，生物量估算誤差較 2D 方法平均降低 8.2%。 |
| 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；其驗證集中於美國八城，且生物量真值為多源遙測推估而非實地量測。 |
授權與圖表重用
| Article license | CC-BY |
|---|---|
| Figure reuse policy | REUSE_ALLOWED_WITH_ATTRIBUTION_CC_BY |
| Notes | 原文版權頁標示 Copyright 2024 by the authors, Licensee MDPI, Basel, Switzerland，為 Creative Commons Attribution (CC BY) 4.0 開放取用條款。圖表可於標註出處後重用。 |
待查核清單
- 如需公開重用 Table 3 / Table 4 / Table 6，標註 CC-BY 出處即可，無版權障礙。
- 若用於台灣情境，需重新評估異速生長方程式參數 a、b 與物種清單是否適用亞熱帶城市森林。
- 可從 github.com/ForestryIIP/3DCiLBE 取得開源實作作進一步比對。
- 可選擇性檢視 Fig. 1 到 Fig. 13 的方法流程圖與結果圖以利製作自繪示意。