基于反射率和亮度温度的锡林郭勒积雪深度估算

积雪是我国西北干旱半干旱区重要的水资源，也是影响全球气候变化的重要因子之一。 目前光学影像反射率和雷达亮温数据是积雪遥感领域的主要数据，本文首次结合两类遥感数据估 算积雪深度，并比较偏最小二乘法和机器学习算法（人工神经网络、支持向量机和随机森林算法） 在积雪深度估算方面的表现。以锡林郭勒盟 2012—2015 年积雪深度数据为例，基于反射率和亮度 温度相结合的积雪深度估算精度优于单个数据源，且随机森林算法表现最好，均方根误差为 2.93 cm，满足实际应用的需求。研究结果对我国西北地区水资源分布、生态环境评估等研究具有重要 意义。

Estimation of snow depth based on reflectance and bright temperature in Xilin Gol League

Snow is an important water resource in arid and semi-arid regions of northwest China, which is also one of the important factors that affect the global climate change. Snow depth is the main parameter of the water resources and its estimation using a remote sensing technique has a vital effect on the assessment of resources and disease of northern China. The reflectance and brightness temperature data were the main data for snow remote sensing. Taking the snow depth data of 2012—2015 for Xilin Gol League, Inner Mongolia, China as an example, The MODIS/Terra satellite reflectance data and FY-3B brightness temperature data for the estimation of snow depth are combined for the first time in this study. The estimation accuracy for partial least squares and machine learning (artificial neural network, support vector machine and random forest) are then compared. When only reflectivity data is used, best accuracy is achieved by the random forest algorithm (RMSE=3.65 cm), followed by the neural network and support vector machine. The stochastic forest algorithm still performs best (RMSE=3.12 cm) when only the brightness and temperature data are used. However, no obvious difference is found with the artificial neural network. Using the partial least squares method exhibits the worst performance among the methods used. For the four methods, the accuracy of estimating snow depth using brightness temperature data is higher than that using reflectivity data, which may be attributed to the fact that reflectivity data is easy to saturate. Comprehensive research shows that the snow depth estimation accuracy based on the combination of reflectance and brightness temperature was better than that with single data. The random forest algorithm obtained the best performance, which is able to meet the needs of practical application. However, the accuracy of the three machine learning algorithms is not much different, with RMSE of 2.93- 3.92 cm, which is far lower than that of the partial least squares method (RMSE=7.12 cm). Although partial least squares method can eliminate the collinearity of variables, its performance in snow depth estimation is poor, which may be attributed to its difficulty in adapting to complex snow scenes. Results of this paper have important significance for the study of water resources distribution and ecological environment assessment in northern China.