Sample data for simulation in Jing-Jin-Ji region

Vegetation distribution simulations could help to understand vegetation distribution patterns and trends, but it is difficult to accurately simulate the distribution of vegetation especially in regions that are heavily affected by human disturbance. Climate, topographic, and spectral data were used as input predictor variables of four machine learning models, including the random forest (RF), decision tree (DT), support vector machine (SVM) and maximum likelihood methods, in three vegetation classification units, including the vegetation group, vegetation type, and formation and subformation, in the Jing-Jin-Ji region, which is one of the most developed regions in China. A total of 2789 vegetation points were used for model training, and 974 vegetation points were used for model assessment. The result showed that the random forest method was the best of the four models and could simulate the distribution of the vegetation in all three classification units well. Kappa coefficients indicated that the random forest method had the highest prediction ability in regard to vegetation type, followed by vegetation group, formation and subformation. Five predictor variables, including 4 climate variables (annual mean temperature, max temperature of warmest month, min temperature of coldest month and annual precipitation) and 1 geospatial variable (elevation), were the most important for three vegetation classification levels. The winter surface albedo of band 4, the slope and the three summer spectral variables (the summer surface albedo of bands 2 and 6 and the summer brightness index) could also increase the accuracy of vegetation classification to some extent.

植被分布模拟有助于理解植被分布格局与变化趋势，但精准模拟植被分布颇具挑战，在受人类活动强烈干扰的区域尤为如此。本研究以中国最发达区域之一的京津冀（Jing-Jin-Ji）地区为研究区，选取气候、地形与光谱数据作为4种机器学习模型的输入预测变量，所涉模型包括随机森林（random forest, RF）、决策树（decision tree, DT）、支持向量机（support vector machine, SVM）以及最大似然法；实验设置了3类植被分类单元，分别为植被群、植被型、群系与亚群系。研究共采集2789个植被样点用于模型训练，974个植被样点用于模型性能评估。实验结果表明，随机森林模型在4种模型中表现最优，可精准模拟3类分类单元下的植被分布。卡帕系数（Kappa coefficient）分析显示，随机森林模型对植被型的预测能力最强，其次依次为植被群、群系与亚群系。在3类植被分类层级中，5个预测变量的重要性最为突出，其中包括4个气候变量（年平均气温、最热月最高气温、最冷月最低气温与年降水量）以及1个地理空间变量（海拔高程）。波段4的冬季地表反照率、坡度，以及3个夏季光谱变量（波段2与波段6的夏季地表反照率、夏季亮度指数）也可在一定程度上提升植被分类精度。