文献类型：期刊文献

英文题名：Mapping Dryland Ecosystems Using Google Earth Engine and Random Forest: A Case Study of an Ecologically Critical Area in Northern China

作者：Li, Shuai[1,2,3] Guo, Pu[1,2] Sun, Fei[3] Zhu, Jinlei[1,2] Cao, Xiaoming[1,2] Dong, Xue[3] Lu, Qi[1,2]

第一作者：李帅

通信作者：Lu, Q[1];Lu, Q[2]

机构：[1]Chinese Acad Forestry, Inst Desertificat Studies, Beijing 100091, Peoples R China;[2]Chinese Acad Forestry, Inst Ecol Conservat & Restorat, Beijing 100091, Peoples R China;[3]Chinese Acad Forestry, Expt Ctr Desert Forestry, Inner Mongolia Dengkou Desert Ecosyst Natl Observa, Bayannur 015200, Peoples R China

年份：2024

卷号：13

期号：6

外文期刊名：LAND

收录：;Scopus(收录号：2-s2.0-85197251891);WOS:【SSCI(收录号:WOS:001256350900001)】；

基金：This research was funded by National Nonprofit Institute Research Grant of Chinese Academy of Forestry (grant number CAFYBB2021MC002-02), National Nonprofit Institute Research Grant of Chinese Academy of Forestry (grant number CAFYBB2021ZB003-04), and Science & Technology Fundamental Resources Investigation Program (grant number 2023FY100703)

语种：英文

外文关键词：dryland ecosystem mapping; Google Earth Engine; random forest algorithm; Landsat 8 OLI; object-based segmentation; feature optimization

摘要：Drylands are characterized by unique ecosystem types, sparse vegetation, fragile environments, and vital ecosystem services. The accurate mapping of dryland ecosystems is essential for their protection and restoration, but previous approaches primarily relied on modifying land use data derived from remote sensing, lacking the direct utilization of latest remote sensing technologies and methods to map ecosystems, especially failing to effectively identify key ecosystems with sparse vegetation. This study attempts to integrate Google Earth Engine (GEE), random forest (RF) algorithm, multi-source remote sensing data (spectral, radar, terrain, texture), feature optimization, and image segmentation to develop a fine-scale mapping method for an ecologically critical area in northern China. The results showed the following: (1) Incorporating multi-source remote sensing data significantly improved the overall classification accuracy of dryland ecosystems, with radar features contributing the most, followed by terrain and texture features. (2) Optimizing the features set can enhance the classification accuracy, with overall accuracy reaching 91.34% and kappa coefficient 0.90. (3) User's accuracies exceeded 90% for forest, cropland, and water, and were slightly lower for steppe and shrub-steppe but were still above 85%, demonstrating the efficacy of the GEE and RF algorithm to map sparse vegetation and other dryland ecosystems. Accurate dryland ecosystems mapping requires accounting for regional heterogeneity and optimizing sample data and feature selection based on field surveys to precisely depict ecosystem patterns in complex regions. This study precisely mapped dryland ecosystems in a typical dryland region, and provides baseline data for ecological protection and restoration policies in this region, as well as a methodological reference for ecosystem mapping in similar regions.