Retrieved Snow Depth Over The Tibetan Plateau Based On AMSR2 Passive Microwave Data

The Tibetan Plateau(TP)is one of the three stable snow-covered regions in China.The climate change in northern and southern China,and in South Asia deeply influenced by the abnormal changes in snow cover on the TP.In addition,there are also distribute important pastoral area of TP,and snow research is very crucial for snow disaster monitoring and early warning in pastoral areas.It is also known as the“Roof of the World”,“Asian Water Tower”and“Third Pole”.Many rivers in China originate from TP,and it is headwater for the Yangtze River,the Yellow River,the Lancang River,and the Yarlung Zangbo River.The TP is also a vital mid-latitude and high-altitude mountainous area in the world.Recently,many researchers have paid more and more attention to the changes of snow in mountainous areas.Passive microwave remote sensing data is widely used in snow depth(SD)retrieval and snow water equivalent estimation,and it is also an effective and extensive SD retrieval method.However,a series of SD products have poor inversion results on SD inversion over the TP.Therefore,using passive microwave remote sensing data for the new SD inversion over the TP is expected to provide a reference for future SD inversion with complex terrain and topography,and to provide reliable support for snow cover monitoring and snow disaster early warning in the TP.In this study,the measured SD by the national meteorological station of China and field measured SD were used to evaluate the AMSR2 SD products in the TP.Based on the new generation of AMSR2 passive microwave remote sensing data,multi-parameter models and machine learning models(random forest,support vector machine and BP artificial neural network)were used to achieve SD inversion for the TP.The multi-parameter model introduced brightness temperature parameters(Tb),brightness temperature difference parameters(TBD),reciprocal of brightness temperature parameters(reTB),natural logarithmic of brightness temperature parameters(lnTB),terrain parameters(Elevation,Ele),and geographic parameters(Longitude,Lon;Latitude,La).The inversion factors selected by genetic algorithm and simulated annealing algorithm were used to establish machine learning SD inversion model.(1)The accuracy of AMSR2 SD products on the TP is poor.The root mean square error(RMSE)of AMSR2 SD products exceeds 10 cm,the ascending products and descending products are 11.82 cm and 13.34 cm,respectively.The inversion results of ascending products on the shallow snow(<10 cm)are more accurate over the TP,and the average SD of descending products is closer to the measured average SD.Grassland by low vegetation is mainly land cover type of the TP.The ascending products show large errors on the underlying surface of the grassland,and it's RMSE exceeds 10 cm.The average elevation of the TP exceeds 4 000 m,and the accuracy of the ascending products are better than that of the descending products in the area above at elevation 3800 m.(2)The accuracy of newly constructed multi-parameter SD inversion models are improved compared to the AMSR2 products.When the ground-measured SD is less than 30 cm,the RMSE and the mean absolute error(MAE)of developed model are below 7 and 6 cm,respectively,and the BIAS is approximately 1 cm.The SD inversion results are slightly poor in the western part of the TP due to complex terrain and thick snow cover.In conclusion,this novel SD inversion model is more applicable and accurate than AMSR2 SD products.(3)Compared to the multi-parameter model,the machine learning model(SARF)has further improved accuracy.The RMSE of optimal machine learning model SARF is 5.25 cm,and the coefficient of decision(R~2)for the model is 0.67.The introduction of multiple parameters in a machine learning model will improve the accuracy of SD inversion in complex terrain to some extent,such as terrain parameters(elevation)and geographic parameters(latitude).Compared with the genetic algorithm(GA),the factor-trained random forest model selected by the simulated annealing algorithm(SA)has higher accuracy.For areas with complicated terrain such as the TP,the combination of the brightness temperature difference between the low-frequency channel and the mid-high-frequency channel may be more suitable for the SD inversion of the random forest algorithm.(4)The stability of the machine learning model for SD inversion on the TP is superior to the multi-parameter model.The accuracy of the machine learning model for various SD is better than the multi-parameter model.The RMSE for SDs above 40 cm is only about 10 cm,while the multi-parameter model exceeds 20 cm.The multi-parameter model still has limitations on the inversion of thick snow(>30 cm),especially in the area of the Himalayas,while machine learning model has less error in the SD inversion of weather stations in the Himalayas.