EVENT DOSSIER
Danish researchers used national forest survey plots and remote sensing data, combined with basic models such as TESSERA, to generate component maps and construct the first high-precision national tree species distribution map with a resolution of 10 meters. The study compared spectral temporal features with input representations of traditional base models, and classified evaluations were conducted with crown height information. Results showed that the multi-layer perceptron based on spectral temporal features achieved macro F1 scores of 0.843 and 0.653 in pure and mixed forests, respectively; the TESSERA embedding model performed better than traditional methods when the training plots accounted for less than 25%. Multi-year observations significantly improved classification accuracy, and the final tree species map had an overall accuracy of 79.9%. It has been released as an open-access product, serving forest monitoring, ecological research, and land management.
丹麦研究人员利用国家森林调查样地和遥感数据，结合基础模型生成分量图，构建了首个高分辨率全国树种分布图。研究对比了人工构建的光谱时序特征与 TESSERA、AlphaEarth 两种基础模型嵌入的输入表示，并辅以冠高信息，采用随机森林、XGBoost 及多层感知机进行分类评估。结果显示，基于光谱时序特征的多层感知机在纯林和混林中分别达到 0.843 和 0.653 的宏观 F1 分数；TESSERA 嵌入模型在纯林上表现具有竞争力，且在训练样地少于 25% 时优于传统方法。多年度观测显著提升了分类精度，最终生成的 10 米分辨率树种图整体准确度为 79.9%，已作为开放获取产品发布，服务于森林监测、生态研究及土地管理。