Validation Samples for SFAC

The 2,072 samples of secondary and 3000 samples of stable forest, respectively were used to assess the accuracy of the results produced by each algorithm and ensemble. Samples for validation were selected randomly from the 7th National Forest Resources Inventory (NFRI) and were compared to the secondary forest maps produced above. The candidate points were visually examined using “Landsat Time Series Explorer”, a shared Application on GEE (https://jstnbraaten.users.earthengine.app/view/landsat-timeseries-explorer). In addition, historical imagery from Google Earth (https://earth.google.com/), GF-6 panchromatic/multispectral (PMS) images (a high-resolution Chinese satellite) (https://data.cresda.cn/#/2dMap) helped to distinguish stable and secondary forest samples. A total of 2,072 validation samples of secondary forest age ranging from 0 to 34 were defined by the re-interpreted approach mentioned above. Over 3,000 candidates of stable forests were systematically sampled from stable forest maps for validation. The classification of these samples of stable forest was ensured by filtering through many public land cover products. As shown in Table 1, these datasets included AGLC-2000-2015, GLC_FCS, FNF, GLC, CLUD, and GFCC. The categorization of the samples as stable forest was ensured by processing using Python, ArcGIS 10.6, and GEE. The 3,000 samples of the stable forest were then completed after manually removing pixels at imperfect sites. The value 1 of class in the data presents the secondary forest, value 0 of class presents the stable forest samples.

本研究分别选取2072个次生林（secondary forest）样本与3000个稳定林（stable forest）样本，用于评估各算法及集成模型生成结果的精度。验证样本从第七次全国森林资源清查（National Forest Resources Inventory, NFRI）中随机抽取，并与前文生成的次生林分布图进行对比。候选样本通过谷歌地球引擎（Google Earth Engine, GEE）上的共享应用"Landsat时间序列浏览器（Landsat Time Series Explorer）"开展目视解译验证，该应用的访问链接为"https://jstnbraaten.users.earthengine.app/view/landsat-timeseries-explorer"。此外，还借助谷歌地球（Google Earth）历史影像（访问链接："https://earth.google.com/"）、中国高分六号（GF-6）全色/多光谱（Panchromatic/Multispectral, PMS）高分辨率卫星影像（访问链接："https://data.cresda.cn/#/2dMap"）来区分稳定林与次生林样本。通过上述重新解译方法，共定义了2072个林龄范围为0至34年的次生林验证样本。从稳定林分布图中系统抽样获取了3000余个稳定林候选样本用于验证。这些稳定林样本的分类正确性通过多套公开土地覆盖产品筛选得到保障，如表1所示，涵盖AGLC-2000-2015、GLC_FCS、FNF、GLC、CLUD及GFCC。样本的稳定林类别判定通过Python、ArcGIS 10.6及GEE联合处理完成。在人工剔除存在缺陷的像元后，最终完成3000个稳定林样本的筛选。数据集中类别标签为1代表次生林，标签为0则代表稳定林样本。