文章背景与核心概要
本文提出了一个国家尺度的丹麦树种测绘综合框架,利用国家森林清查(NFI)样地与对地观测(EO)数据,评估了地理空间基础模型(Foundation Models)相比于传统特征工程在树种分类中的有效性。研究对比了两种输入表示形式:一是基于多时相 Sentinel-1 和 Sentinel-2 观测手动提取的光谱-时间特征(STF),二是利用 EO 基础模型 TESSERA 和 AlphaEarth 生成的嵌入(Embeddings)。
通过结合冠层高度数据,并在纯林和混林样地上评估了随机森林、XGBoost 以及多层感知机(MLP)分类器,研究发现基于 STF 的 MLP 取得了最高的整体分类性能。然而,在训练数据有限的情况下(少于约 25% 的样地可用),TESSERA 嵌入表现出显著的优势。最佳性能模型最终被扩展应用于国家尺度,生成了丹麦首张高分辨率(10米)树种分布图,为森林监测和生态研究提供了宝贵的开源资源。
丹麦树种测绘:光谱时间特征与地理空间基础模型嵌入的对比 (Tree Species Mapping in Denmark: A Comparison of Spectral-Temporal Features with Geospatial Foundation Model Embeddings)
arXiv ID: arXiv:2609.03480 [cs.CV]
提交时间: 2026年9月3日
作者: Alkiviadis Koukos, Spyros Kondylatos, Thomas Nord-Larsen, Lotte Nyborg, Christian Tøttrup, Kenneth Grogan
📌 摘要与核心结论 (Summary)
本文提出了一个综合框架,利用国家森林清查(NFI)样地和对地观测(EO)数据,在丹麦开展全国尺度的树种测绘。该研究通过对比树种分类的两种输入表示形式,评估了地理空间基础模型相较于传统特征工程的有效性: 1. 手动工程化的光谱-时间特征(STF),由多时相 Sentinel-1 和 Sentinel-2 观测衍生而来。 2. 由 EO 基础模型 TESSERA 和 AlphaEarth 生成的嵌入(Embeddings)。
两种表示均辅以冠层高度数据,并在纯林和混林样地上使用随机森林、XGBoost 和多层感知机(MLP)分类器进行了评估。
This paper presents a comprehensive framework for national-scale tree species mapping in Denmark using National Forest Inventory (NFI) plots and Earth Observation (EO) data. The study evaluates the effectiveness of geospatial foundation models compared to traditional feature engineering by comparing two input representations for tree species classification: 1. Manually engineered spectral-temporal features (STF) derived from multi-temporal Sentinel-1 and Sentinel-2 observations. 2. Embeddings generated by the EO foundation models TESSERA and AlphaEarth.
Both representations were complemented with canopy height data and evaluated using Random Forest, XGBoost, and Multi-Layer Perceptron (MLP) classifiers across pure and mixed forest stands.
🔍 关键发现 (Key Findings)
- 顶级性能: 基于 STF 的 MLP 实现了最高的整体分类性能,纯林和混林的宏平均 F1 分数分别达到 0.843 和 0.653。
- 基础模型的竞争力: 在纯林分类中,基于 TESSERA 嵌入训练的 MLP 表现出极具竞争力的结果,与表现最好的模型仅相差 1.1 个百分点。
- 小样本优势: 在训练数据有限(可用训练样地少于约 25%)的情况下,TESSERA 的表现始终优于基于 STF 的模型。
- 时序与多传感器价值: 与单年输入相比,多年观测系统性地提高了分类准确率。消融实验也凸显了 Sentinel-1 后向散射、光谱指数和冠层高度数据的互补优势。
- 全国尺度制图产品: 性能最佳的框架在全国范围内扩展应用,生成了丹麦第一张高分辨率(10米)树种图,经面积调整后的总体验证准确率达到 79.9%。
- Top Performance: The STF-based MLP achieved the highest overall classification performance, reaching macro F1 scores of 0.843 (pure stands) and 0.653 (mixed stands).
- Foundation Model Competitiveness: The MLP trained on TESSERA embeddings delivered highly competitive results for pure stands, coming within 1.1 percentage points of the best-performing model.
- Low-Data Advantage: TESSERA consistently outperformed STF-based models when trained on limited data (when fewer than ~25% of training plots were available).
- Temporal and Multi-Sensor Value: Multi-year observations systematically improved classification accuracy compared to single-year inputs. Ablation experiments also highlighted the complementary benefits of Sentinel-1 backscatter, spectral indices, and canopy height data.
- National-Scale Map Product: The top-performing framework was scaled nationally to produce the first high-resolution (10 m) tree species map of Denmark, achieving an area-adjusted overall validation accuracy of 79.9%.
📄 论文摘要 (Abstract)
我们利用国家森林清查样地和对地观测数据对丹麦的树种进行了测绘,同时评估了基础模型在大规模森林表征方面的潜力。我们对比了树种分类的两种替代输入表示:(i) 源自多时相 Sentinel-1 和 Sentinel-2 观测的手动工程光谱-时间特征(STF),以及 (ii) 由 EO 基础模型 TESSERA 和 AlphaEarth 生成的嵌入。这两种表示都补充了冠层高度信息。针对所有输入表示,评估了随机森林、XGBoost 和多层感知机(MLP)分类器,并对纯林和混林进行了单独评估。基于 STF 的 MLP 实现了最高的分类性能,纯林和混林的宏平均 F1 分数分别为 0.843 和 0.653。在 TESSERA 嵌入上训练的 MLP 为纯林提供了具有竞争力的性能,其结果在最佳性能模型的 1.1 个百分点以内。当可用训练样地少于大约 25% 时,TESSERA 的表现始终优于基于 STF 的模型,证明了其在有限训练数据下的显著优势。与单年输入相比,多年观测系统地改善了分类准确率,而消融实验揭示了 Sentinel-1 后向散射、光谱指数和冠层高度数据的互补贡献。表现最好的模型随后被应用于国家尺度,以生成丹麦 10 米树种图。面积调整后的验证表明,总体地图准确率为 79.9%。生成的地图作为开源产品发布,是丹麦第一张高分辨率国家树种图,为森林监测、生态研究和土地管理应用提供了宝贵的资源。
We map tree species across Denmark using National Forest Inventory plots and EO data, while evaluating the potential of foundation models for large-scale forest characterization. We compare two alternative input representations for tree species classification: (i) manually engineered spectral-temporal features (STF) derived from multi-temporal Sentinel-1 and Sentinel-2 observations, and (ii) embeddings generated by the EO FMs TESSERA and AlphaEarth. Both representations are complemented with canopy height information. Random forest, XGBoost, and Multi-Layer Perceptron (MLP) classifiers are evaluated for all input representations, with separate assessments for pure and mixed forest stands. The STF-based MLP achieves the highest classification performance, yielding macro F1 scores of 0.843 and 0.653 for pure and mixed stands, respectively. The MLP trained on TESSERA embeddings delivers competitive performance for pure stands, achieving results within 1.1 percentage points of the best-performing model. TESSERA consistently outperforms STF-based models when fewer than approximately 25% of training plots are available, demonstrating a substantial advantage under limited training data. Multi-year observations systematically improve classification accuracy relative to single-year inputs, while ablation experiments reveal the complementary contributions of Sentinel-1 backscatter, spectral indices, and canopy height data. The best-performing model is subsequently applied at the national scale to generate a 10 m tree species map of Denmark. Area-adjusted validation indicates an overall map accuracy of 79.9%. The resulting map, released as an open-access product, is the first high-resolution national tree species map of Denmark and provides a valuable resource for forest monitoring, ecological research, and land management applications.
🔗 资源与链接 (Resources & Links)
- 全文 PDF: 查看 PDF
- HTML 版本: arXiv HTML (实验性)
- 数据集与地图产品: Zenodo 存储库
- DOI: 10.48550/arXiv.2609.03480
- 许可协议: 知识共享 署名-非商业性使用-禁止演绎 4.0 国际
- Full-Text PDF: View PDF
- HTML Version: arXiv HTML (Experimental)
- Dataset & Map Product: Zenodo Repository
- DOI: 10.48550/arXiv.2609.03480
- License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
