摘要
青藏高原因其复杂的地形地势和和积雪分布使得多种雪深算法未达到理想的精度。基于新一代被动微波数据AMSR2(Advanced Microwave Scanning Radiometer2),应用随机森林算法(Random Forest,RF)将亮温(Brightness Temperature,BT)和亮温差(Brightness Temperature Difference,BTD)作为参数输入,并将高程和纬度参数引入雪深反演模型中,经过模拟退火算法进行有效反演因子筛选,构建了基于随机森林算法的青藏高原雪深反演模型。结果表明:与AMSR2全球雪深产品相比,随机森林算法的拟合优度(R2)由0.41提升至0.60,均方根误差(Root Mean Square Error,RMSE)由7.36cm降至4.88cm,偏差(BIAS)由3.24cm减小至-0.16cm,随机森林雪深反演模型在青藏高原的精度更高;青藏高原平均海拔超过4000m,当海拔大于青藏高原平均海拔时,随机森林算法的反演效果最差,但RMSE仅为3.78cm,BIAS仅为-0.09cm;高原南部(25°~30°N)因其复杂的地势和相对较少的气象站点使得反演效果较差,RMSE为5.94cm,BIAS为-0.39cm;青藏高原的主要土地覆盖类型为草地,随机森林算法在草地的RMSE约为3cm,BIAS接近0cm。
Due to the complex terrain and snow distribution of the Tibetan Plateau(TP),many snow depth retrieval algorithms have not reached the ideal accuracy.Based on the new generation of AMSR2 passive microwave data,the TP was used as the study area in this study,the Brightness Temperature(BT)and the Brightness Temperature Difference(BTD)were input as parameters,and the elevation and latitude parameters were also introduced into the SD retrieval model.The simulated annealing algorithm(SA)was used to feature selection,and the SD retrieval model of the TP that based on the RF algorithm was constructed.The results showed that the RF algorithm had higher accuracy over the TP compared with the AMSR2 global SD products;the goodness of fit(R^2)were increased from 0.41 to 0.60,the root mean square error(RMSE)were decreased from 7.36 cm to 4.88 cm,and the BIAS were decreased from 3.24 cm to-0.16 cm,respectively.The BTD of low frequency are more suitable for SD retrieval with RF algorithm over the TP.When the elevation approached or exceeded the average elevation of the TP by 4000 m,the retrieval result was not well,but the RMSE was only 3.78 cm,and the BIAS was only-0.09 cm.The southern part of the plateau(25°~30°N)had poor retrieval results due to its complex topography and relatively few meteorological stations,with the RMSE of5.94 cm and the BIAS of-0.39 cm.The grassland,bare land and farmland with the low vegetation cover area had a maximum RMSE of 3.19 cm,and BIAS of-0.49 cm.The main land cover type on the TP was grassland,and RMSE and BIAS of the random forest algorithm on the grassland underlay were about 3 cm and 0 cm,respectively.
作者
王健顺
王云龙
周敏强
刘畅宇
黄晓东
WANG Jianshun;WANG Yunlong;ZHOU Minqiang;LIU Changyu;HUANG Xiaodong(State Key Laboratory of Grassland Agro-ecosystems/College of Pastoral Agriculture Science and Technology,Lanzhou University,Lanzhou 730020,China;School of Geographical Sciences,Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《冰川冻土》
CSCD
北大核心
2020年第3期1077-1086,共10页
Journal of Glaciology and Geocryology
基金
国家自然科学基金项目(41691330,41971293)
科技部基础资源调查专项(2017FY100501)
教育部长江学者和创新团队发展计划人才项目(IRT_17R50)资助。
