Abstract. Accurate information on forest tree species composition is vital for various scientific applications, as well as for forest inventory and management purposes. Country-wide, detailed species maps are a valuable resource for environmental management, conservation, research, and planning. Here, we performed the classification of 16 dominant tree species and genera in Poland using time series of Sentinel-2 imagery. To generate comprehensive spectral–temporal information, we created Sentinel-2 seasonal aggregations known as spectral–temporal metrics (STMs) within the Google Earth Engine (GEE). STMs were computed for short periods of 15–30 d during spring, summer, and autumn, covering multi-annual observations from 2018 to 2021. The Polish Forest Data Bank served as reference data, and, to obtain robust samples with pure stands only, the data were validated through automated and visual inspection based on very-high-resolution orthoimagery, resulting in 4500 polygons serving as training and test data. The forest mask was derived from available land cover datasets in GEE, namely the ESA WorldCover and Dynamic World dataset. Additionally, we incorporated various topographic and climatic variables from GEE to enhance classification accuracy. The random forest algorithm was employed for the classification process, and an area-adjusted accuracy assessment was conducted through cross-validation and test datasets. The results demonstrate that the country-wide forest stand species mapping achieved an accuracy exceeding 80 %; however, this varies greatly depending on species, region, and observation frequency. We provide freely accessible resources, including the forest tree species map and training and test data: https://doi.org/10.5281/zenodo.10180469 (Grabska-Szwagrzyk, 2023a).

中文翻译：

---

基于 Sentinel-2 数据的波兰森林树种地图


摘要。有关森林树种组成的准确信息对于各种科学应用以及森林清查和管理目的至关重要。全国范围内详细的物种地图是环境管理、保护、研究和规划的宝贵资源。在这里，我们使用 Sentinel-2 图像的时间序列对波兰 16 个优势树种和属进行了分类。为了生成全面的频谱-时间信息，我们在 Google 地球引擎 (GEE) 中创建了 Sentinel-2 季节性聚合，称为频谱-时间指标 (STM)。 STM 是在春季、夏季和秋季 15-30 天的短期内计算的，涵盖 2018 年至 2021 年的多年观测。波兰森林数据库作为参考数据，为了获得仅纯林分的可靠样本，通过基于超高分辨率正射影像的自动视觉检查对数据进行了验证，产生了 4500 个多边形作为训练和测试数据。森林掩模源自 GEE 中可用的土地覆盖数据集，即 ESA WorldCover 和 Dynamic World 数据集。此外，我们还纳入了 GEE 中的各种地形和气候变量，以提高分类准确性。分类过程采用随机森林算法，并通过交叉验证和测试数据集进行区域调整精度评估。结果表明，全国林分物种图谱准确率超过80%；然而，这根据物种、地区和观察频率的不同而有很大差异。我们提供免费资源，包括森林树种地图以及训练和测试数据：https://doi.org/10.5281/zenodo。10180469（Grabska-Szwagrzyk，2023a）。