Abstract Information about forest stand species distribution is essential for biodiversity modelling, forest disturbances, fire hazard and drought monitoring, biomass and carbon estimation, detection of non-native and invasive species, as well as for planning forest management strategies. High temporal and spectral resolution remote sensing data from the Sentinel-2 mission enables the derivation of accurate and timely maps of tree species in forests in a cost-efficient way. However, there is still a lack of studies regarding forest stand species mapping for large areas like the Polish Carpathian Mountains (approx. 20,000 km2). In this study, we aimed to develop a workflow to obtain forest stand species maps with machine learning algorithms applied to multi-temporal Sentinel-2 products and environmental data at regional scale. Using variable importance techniques - Variable Importance Using Random Forests (VSURF) and Recursive Feature Elimination (RFE) - we assessed three Sentinel-2 Best Available Pixel composites (April, July and October), eight annual spectral-temporal metrics (STM; mean, minimum, maximum, standard deviation, range, first quartile, third quartile and interquartile range), and four environmental topographic variables (elevation, slope, aspect, distance to water bodies), i.e. 114 variables in total. Following a variable importance assessment, we produced maps of eleven tree species with the use of three Machine Learning algorithms: Random Forest (RF), Support Vector Machines (SVM) and Extreme Gradient Boosting (XGB) on nine different variable subsets, i.e. in total 27 classifications. The results showed that SVM outperformed the other two classifiers - the highest overall accuracy exceeded 85% for SVM classification of all variables (86.9%), and 64 variables (85.6%). Including elevation information improved the accuracies. From the best five classifications we created a final ensemble map (overall accuracy of 86.6%) and a precision map based on the Simpson Index, which indicates where the five models agree. This ensemble approach allowed us to determine that the lowest precision occurred in foothills and basins with lower forest cover, in the areas with lack of good quality imagery, and at the borders of stands with homogenous species composition. On the other hand, the highest precision occurred in regions with homogenous stands with high forest and canopy cover. The study demonstrates the potential of Sentinel-2 imagery and topographic data in mapping forest stand species in large mountainous areas with high accuracy. Furthermore, it demonstrates the usefulness of the ensemble approach, which enables to assess the classification precision.

中文翻译：

---

使用 Sentinel-2 图像和波兰喀尔巴阡山脉的环境数据对林分物种制图的机器学习算法进行评估

摘要 关于林分物种分布的信息对于生物多样性建模、森林干扰、火灾和干旱监测、生物量和碳估计、非本地和入侵物种的检测以及规划森林管理策略至关重要。来自 Sentinel-2 任务的高时间和光谱分辨率遥感数据能够以具有成本效益的方式推导出准确和及时的森林树种地图。然而，关于波兰喀尔巴阡山脉（约 20,000 平方公里）等大面积林分物种制图的研究仍然缺乏。在本研究中，我们旨在开发一种工作流程，通过应用于区域尺度的多时 Sentinel-2 产品和环境数据的机器学习算法获取林分物种地图。使用可变重要性技术 - 使用随机森林 (VSURF) 和递归特征消除 (RFE) 的可变重要性 - 我们评估了三个 Sentinel-2 最佳可用像素组合（4 月、7 月和 10 月）、八个年度光谱时间指标（STM；平均值、最小值、最大值、标准差、范围、第一四分位数、第三四分位数和四分位数间距）和四个环境地形变量（高程、坡度、坡向、与水体的距离），共 114 个变量。在对变量重要性进行评估之后，我们使用三种机器学习算法生成了 11 种树种的地图：随机森林 (RF)、支持向量机 (SVM) 和极限梯度提升 (XGB) 在九个不同的变量子集上，即总共27个分类。结果表明，SVM 优于其他两个分类器——SVM 分类所有变量（86.9%）和 64 个变量（85.6%）的最高总体准确率超过 85%。包括高程信息提高了准确性。从最好的五个分类中，我们创建了最终的集成图（总体准确率为 86.6%）和基于辛普森指数的精度图，这表明了五个模型的一致之处。这种整体方法使我们能够确定最低精度出现在森林覆盖率较低的山麓和盆地、缺乏高质量图像的区域以及具有同质物种组成的林分边界处。另一方面，最高的精度出现在具有高森林和冠层覆盖的同质林分区域。该研究证明了 Sentinel-2 图像和地形数据在高精度绘制大面积山区林分物种方面的潜力。此外，它展示了集成方法的有用性，它能够评估分类精度。