The process of forest degradation, along with deforestation, is the second greatest producer of global greenhouse gas emissions. A key challenge that remains unresolved is how to quantify the critical threshold that distinguishes a degraded from a non-degraded forest. We determined the critical threshold of forest degradation in mature stands belonging to the temperate evergreen rain forest of southern Chile by quantifying key forest stand factors characterizing the forest degradation status. Forest degradation in this area is mainly caused by high grading, harvesting of fuelwood, and sub-canopy grazing by livestock. We established 160 500-m² plots in forest stands that represented varied degrees of alteration (from pristine conditions to obvious forest degradation), and measured several variables related to the structure and composition of the forest stands, including exotic and native species richness, soil nutrient levels, and other landscape-scale variables. In order to identify classes of forest degradation, we applied multivariate and machine-learning analyses. We found that richness of exotic species (including invasive species) with a diameter at breast height (DBH) < 10 cm and tree density (N, DBH > 10 cm) were the two composition and structural variables that best explained the forest degradation status, e.g., forest stands with five or more exotic species were consistently found more associated with degraded forest and stands with N < 200 trees/ha represented degraded forests, while N > 1,000 trees/ha represent pristine forests. We introduced an analytical methodology, mainly based on machine learning, that successfully identified the forest degradation status that can be replicated in other scenarios. In conclusion, here by providing an extensive data set quantifying forest and site attributes, the results of this study are undoubtedly useful for managers and decision makers in classifying and mapping forests suffering various degrees of degradation.

中文翻译：

---

在巴塔哥尼亚温带雨林中使用多元和机器学习方法评估森林退化

森林退化和森林砍伐是全球第二大温室气体排放源。一个尚未解决的关键挑战是如何量化区分退化森林与非退化森林的关键阈值。我们通过量化表征森林退化状况的关键林分因素，确定了属于智利南部温带常绿雨林的成熟林分的森林退化临界阈值。该地区的森林退化主要是由于高等级、砍伐薪材和牲畜在冠层下放牧造成的。我们建立了 160 500-m ²代表不同程度变化（从原始条件到明显的森林退化）的林分地块，并测量了与林分结构和组成相关的几个变量，包括外来和本地物种丰富度、土壤养分水平和其他景观-尺度变量。为了确定森林退化的类别，我们应用了多变量和机器学习分析。我们发现胸径 (DBH) < 10 cm 和树木密度 ( N , DBH > 10 cm) 的外来物种（包括入侵物种）的丰富度是最能解释森林退化状况的两个组成和结构变量，例如，具有五种或更多外来物种的林分始终被发现与退化的森林和林分与N  < 200 棵树/公顷代表退化的森林，而N  > 1,000 棵树/公顷代表原始森林。我们引入了一种主要基于机器学习的分析方法，该方法成功地确定了可以在其他场景中复制的森林退化状态。总之，通过提供量化森林和场地属性的广泛数据集，本研究的结果无疑有助于管理人员和决策者对遭受不同程度退化的森林进行分类和绘图。