Abstract: Grassland coverage is an important index to evaluate grassland growth status,and the establishment of high-precision coverage estimation model is the key to carry out dynamic monitoring of natural grassland. In this study,we established four remote sensing estimation models of grassland coverage based on the measured data from 2003 to 2018 and 55 indices including topography and soil. The optimal models were obtained by comparison,and the spatial-temporal dynamic changes of grassland coverage in the study area from 2001 to 2019 were analyzed. The results showed as follows:Ten variables with high importance were selected by minimum absolute compression variable screening method. Among them,the correlation between Ratio Vegetation Index and grassland vegetation coverage is better than that between other single-factor models and grassland vegetation coverage. Among the three machine learning models,random forest is superior to artificial neural network and support vector machine,with R² and root mean square errors of 0.68 and 12.75%,respectively. The grassland coverage model constructed by machine learning was superior to the single-factor model,with R² increased by 0.09~0.16 and RMSE decreased by 1.52%~2.81%. From 2001 to 2019,grassland coverage increased from west to east and from north to south in the study area,with 55.4% of the areas showing an increasing trend and 44.6% of the areas showing a decreasing trend.

Key words: Grassland coverage, Random forest, Moderate-resolution Imaging Spectroradiometer, Variable selection

摘要： 草地盖度是评价草地生长状况的重要指标，构建高精度的盖度估测模型是开展天然草地动态监测的关键。本文利用2003—2018年实测数据和地形、土壤等55个指标，建立了4种草地盖度遥感估测模型，通过比较得到研究区草地盖度遥感估测最优模型，并分析了2001—2019年研究区草地盖度时空动态变化。结果表明：由最小绝对压缩变量筛选方法选出10个有较高重要性的变量。其中，比值植被指数与草地植被盖度间的相关性优于其他单因素模型与草地植被盖度间的相关性。在3种机器学习模型中，随机森林优于人工神经网络与支持向量机模型，其R²和均方根误差分别为0.68和12.75%。机器学习方法构建的草地盖度模型优于单因素模型，其R²可提高0.09~0.16，RMSE降低1.52%~2.81%；2001—2019年，研究区草地盖度整体上呈现出自西向东、自北向南增加的趋势，呈增加趋势的面积占比为55.4%，呈减少趋势的面积占比为44.6%。

关键词: 草地盖度, 随机森林, MODIS, 变量筛选