祁连山区域30m土地覆盖分类产品数据集（1985-2017）V1.0

本数据集包括祁连山区域1990年至2017年每5年一期的30m土地覆盖分类产品。该产品首先利用Landsat-8/OLI构造2015年时间序列数据，针对各类地物随时间变化呈现的NDVI时间序列曲线不同，对不同地物特征进行知识归纳，设定提取规则不同地物信息，得到2015年的土地覆盖分类图。分类系统参考了IGBP分类系统和FROM_LC分类系统，共分为耕地、林地、草地、灌丛、湿地、水体、不透水面、裸地、冰川和积雪共10大类。由Google Earth高清影像和实地调研数据进行精度评价，得出2015年土地覆盖分类产品的总体精度高达92.19%。以2015年的土地覆盖分类产品为底图，按各类别的比例选取大量样本，基于Google Earth Engine平台的Landsat系列数据和强大地数据处理能力，利用深度学习的思想，选取随机森林分类器，对波段信息和NDVI、MNDWI、NDBI等指数进行样本训练，生产出1985-2017年每5年一期的土地覆盖分类产品。对2套2015年的分类产品进行比较，得出基于Google Earth Engine平台生产的土地覆盖分类产品与基于时间序列方法得到的分类产品具有很好的一致性。总之，祁连山核心区的土地覆盖数据集具有较高的总体精度，且基于Google Earth Engine平台样本训练的方法能够在时间和空间上对现有的分类产品进行扩展，每5年一期的频次能够在长时间序列上反映更多的土地覆盖类型变化信息。

This dataset contains 30m-resolution land cover classification products for the Qilian Mountains region, updated every 5 years from 1990 to 2017. Initially, the 2015 land cover classification map was produced by constructing time-series datasets using Landsat-8/OLI imagery acquired in 2015. Given that different land cover types exhibit distinct NDVI time-series profiles over time, knowledge induction was conducted for the characteristics of various ground objects, and targeted extraction rules were established to retrieve information of different land features, thus generating the 2015 land cover classification map. The classification system was developed with reference to both the IGBP land cover classification system and the FROM_LC classification system, comprising 10 categories in total: cropland, forest land, grassland, shrubland, wetland, water body, impervious surface, bare land, glacier, and snow cover. Accuracy evaluation was carried out using high-resolution Google Earth imagery and field survey data, revealing that the overall accuracy of the 2015 land cover classification product reached 92.19%. Taking the 2015 land cover classification product as the base map, a large number of samples were selected based on the proportional distribution of each category. Leveraging Landsat series data and the robust data processing capabilities of the Google Earth Engine (GEE) platform, and drawing on deep learning principles, the random forest classifier was adopted to train the model with band information and indices including NDVI, MNDWI, and NDBI, thereby producing land cover classification products updated every 5 years from 1985 to 2017. A comparison of the two sets of 2015 classification products showed that the land cover classification product generated via the GEE platform has excellent consistency with that derived from the time-series method. In summary, the land cover dataset of the core Qilian Mountains region boasts high overall accuracy. Moreover, the sample training method based on the GEE platform can expand existing classification products both temporally and spatially. The 5-year update frequency can reflect more information on land cover type changes across long time series.