文章基本信息

标题：A 30 m terrace mapping in China using Landsat 8 imagery and digital elevation model based on the Google Earth Engine
本地全文：下载
作者：Cao, Bowen ; Yu, Le ; Naipal, Victoria 等
期刊名称：Earth System Science Data (ESSD)
印刷版ISSN：1866-3508
电子版ISSN：1866-3516
出版年度：2021
卷号：13
期号：5
页码：2437-2456
DOI：10.5194/essd-13-2437-2021
出版社：Copernicus
摘要：The construction of terraces is a key soil conservation practice on agricultural land in China providing multiple valuable ecosystem services. Accurate spatial information on terraces is needed for both management and research. In this study, the first 30 m resolution terracing map of the entire territory of China is produced by a supervised pixel-based classification using multisource and multi-temporal data based on the Google Earth Engine (GEE) platform. We extracted time-series spectral features and topographic features from Landsat 8 images and the Shuttle Radar Topography Mission digital elevation model (SRTM DEM) data, classifying cropland area (cultivated land of Globeland30) into terraced and non-terraced types through a random forest classifier. The overall accuracy and kappa coefficient were evaluated by 10 875 test samples and achieved values of 94 % and 0.72, respectively. For terrace class, the producer's accuracy (PA) was 79.945 %, and the user's accuracy (UA) was 71.149 %. The classification performed best in the Loess Plateau and southwestern China, where terraces are most numerous. Some northeastern, eastern-central, and southern areas had relatively high uncertainty. Typical errors in the mapping results are from the sloping cropland (non-terrace cropland with a slope of ≥ 5 ∘ ), low-slope terraces, and non-crop vegetation. Terraces are widely distributed in China, and the total terraced area was estimated to be 53.55 Mha (i.e., 26.43 % of China's cropland area) by pixel counting (PC) method and 58.46 ± 2.99 Mha (i.e., 28.85 % ± 1.48 % of China's cropland area) by error-matrix-based model-assisted estimation (EM) method. Elevation and slope were identified as the main features in the terrace/non-terrace classification, and multi-temporal spectral features (such as percentiles of NDVI, TIRS2, and BSI) were also essential. Terraces are more challenging to identify than other land use types because of the intra-class feature heterogeneity, interclass feature similarity, and fragmented patches, which should be the focus of future research. Our terrace mapping algorithm can be used to map large-scale terraces in other regions globally, and our terrace map will serve as a landmark for studies on multiple ecosystem service assessments including erosion control, carbon sequestration, and biodiversity conservation. The China terrace map is available to the public at https://doi.org/10.5281/zenodo.3895585 (Cao et al., 2020).