AMSR2 Snow Depth Downscaling Based On Ensemble Learning Method Over The Tibetan Plateau

Snow depth is one of the most important snow parameters.It is of great significance to obtain accurate information of snow depth on the Tibetan Plateau for understanding climate change and hydrological cycle.At present,passive microwave remote sensing is the most effective method to monitor snow depth in a large scale.However,due to its low spatial resolution,passive microwave remote sensing data has poor accuracy and great uncertainty in areas with complex terrain conditions and strong heterogeneity of snow distribution,such as the Tibetan Plateau.Based on AMSR2 brightness temperature data of passive microwave remote sensing with spatial resolution of 10 km,this study constructed AMSR2 snow depth downscaling model based on integrated learning method by introducing snow cover days,topographic factors(elevation,slope,aspect,surface roughness)and geographical location(longitude,latitude).In the process of modeling,the same test set and verification set were used to construct snow depth scaling models based on random forest,Ada Boost,GBDT,XGBoost and Light GBM.Through comparative analysis,it was found that the model based on Light GBM had the best performance.Therefore,it was selected as the final model for the production of downscaling snow depth products,and the daily snow depth products with spatial resolution of 500 m on the Tibetan Plateau were produced.Furthermore,the accuracy of the snow depth products and the original AMSR2 snow depth standard products were verified and evaluated using the measured snow depth data of the stations.The main research contents and conclusions are as follows:(1)Based on AMSR2 brightness temperature data,snow cover days,terrain factors and longitude and latitude data,five snow depth downscaling models including random forest,Ada Boost,GBDT,XGBoost and Light GBM were constructed respectively.The optimal parameters of each model were selected by grid search method,and the fitting accuracy of each model was evaluated by five-fold cross validation method.The results show that the model accuracy of XGBoost and Light GBM is significantly better than the other three algorithms.The Light GBM model performed best,with an R~2 of 0.75 and RMSE of 2.87 cm,so it was finally used for downscaling snow depth modeling on the Tibetan Plateau.(2)By comparing the measured snow depth data of AMSR2 standard snow depth product with the descending snow depth product produced in this study,it is found that AMSR2 standard snow depth product significantly overestimates the snow depth in the Tibetan Plateau region.Positive Mean Error(PME)values of the rail ascending and descending products were121.12 cm and 132.01 cm respectively,which were much higher than Negative Mean Error(NME)values of-3.42 cm and-3.17 cm.The PME and NME values of the descending snow depth products generated in this study are 1.15 cm and-1.70 cm,respectively,which slightly underestimate the snow depth of the Tibetan Plateau.In terms of overall accuracy,the RMSE of the downscaling snow depth product produced in this study is 2.50,which is far better than the standard AMSR2 snow depth product(the ascending is 40.10 cm,and the descending is88.90 cm).(3)The accuracy of AMSR2 snow depth standard product and downscaling snow depth product produced in this study were comprehensively evaluated from four aspects: measured snow depth,days of snow cover,elevation and land cover type.The downscaling snow depth data produced in this study are respectively in the shallow snow area(snow depth < 5 cm),,snow cover days between 30 and 60 days,elevation between 2000 and 3000 m and land cover type of construction land has the best accuracy,RMSE of 1.43 cm,2.27 cm,1.71 cm and 1.62 cm respectively.The accuracy of AMSR2 snow depth standard product is the best in the areas where the snow depth is between 20 and 40 cm,the snow cover days are between 0 and 30 days,the elevation is between 3000 and 4000 m,and the land cover type is construction land,with RMSE of 28.60 cm,14.32 cm,11.49 cm and 9.52 cm respectively.In general,the descending snow depth products produced in this study are superior to the standard snow depth products of AMSR2 in all cases.