Reconstruction Of Soil Moisture Data In The Qinghai-Tibet Plateau Based On Machine Learning Algorithms

Soil moisture plays an important role in the balance of energy,water and matter on the global land surface.The soil moisture monitoring stations in the Tibetan Plateau are less,and the spatial distribution is uneven,the coverage is limited,and the monitoring time is not uniform.Meanwhile,the complex terrain conditions in the study area have brought great uncertainty to the observation results.In recent years,the satellite based microwave remote sensing has realized the real-time dynamic monitoring of soil moisture with the advantages of its sensitive,all-weather and short repeated cycles,which provides convenience for the study of soil moisture in large scale and long time series.However,the soil moisture data of satellite remote sensing is easily affected by sensors,inversion algorithms,cloud coverage,environment and other factors,which leads to the gap of remote sensing data and the discontinuity of space,thus limiting the application of soil moisture data in large scale and long time sequence analysis.Therefore,the reconstruction of remote sensing data and the temporal and spatial analysis of reconstructed data become a new research hotspot in the field of remote sensing application.In this study,the CCI soil moisture data product(ESA-CCI-SM)from European Space Agency(ESA),which has the longest time series,is selected for the soil moisture study in the Tibetan Plateau which is the sensitive area of global climate change.Based on the nonlinear relationship between the soil moisture and the related surface environment variables,four common machine learning algorithms are used to reconstruct the soil moisture data.The reconstruction results of the learning algorithm are quantitatively verified and compared,and the applicability of different machine learning algorithms is discussed.Based on the reconstructiondata,the spatial distribution characteristics and the variations of soil moisture in Tibetan Plateau are analyzed.The main research contents and conclusions are as follows:(1)The time sequence products of soil moisture data(ESA-CCI-SM)in the Tibetan Plateau are affected by the scanning strip and inversion algorithm,and there are a large number of missing values in the obtained data.In the growing season(5-10 month),data gap mainly occurred at the beginning and end of the growing season,that is,May and October.Data gap areas are mainly concentrated in the western part of Tibetan Plateau,namely arid and semi-arid regions.In order to improve the completeness of remote sensing monitoring data of soil moisture on Tibetan Plateau,it is an effective method to reconstruct spatio-temporal data of soil moisture.(2)Using the principle of data reconstruction based on machine learning algorithm,the surface environmental varibales,such as the normalized vegetation index NDVI,the surface temperature LST(day/night),the surface albedo Albedo(WSA/BSA),the terrain DEM,the land cover type LUCC and the longitude,latitude(Ion/lat)and so on,are closely related to the soil moisture as the auxiliary variables.By using four commonly used machine learning algorithm random forest(RF),K nearest neighbor(KNN),Bayesian(Bayes)and support vector machine(SVM),the data samples are set up and trained to predict soil moisture on the basis of a good training model.The results show that machine learning algorithm can simulate the relationship between soil moisture and related auxiliary variables and predict soil moisture.(3)On the basis of four algorithms including RF,KNN,Bayes and SVMto reconstruct the soil moisture data of the Tibetan plateau during 2005-2014,the goodness of fit of the algorithm is compared from the original data,the ground site measured data,the ERA and the ITPLDAS-SM.The results show that the reconstruction results of all kinds of algorithms are different because of the different algorithm principles.By comparing with the different angles of original data,ground site measured data,ERA and ITPLDAS-SM,the training precision of long time sequential data reconstruction of RF algorithm is better than that of other algorithms,and it has better spatial expansibility and stability.The reconstruction results can fill the data gap and better retain the true value of the original data,and reflect the variety of soil moisture.(4)The prediction results of RF algorithm with high reconstruction precision are selected as data sources of soil moisture analysis,and the distribution of soil moisture in the Tibetan Plateau during 2005-2014 is analyzed.The results show that:The average annual surface soil moisture in the growing season over the Tibetan Plateau is 0.21 m3·m-3,and the maximum value occurred in August(0.225 m3·m-3).The overall trend is gradually increasing from the northwest to the southeast,and the Tarim Basin in the north has the lowest value of soil moisture.During 2005-2014,the soil moisture was not significantly increased(0.001 m3.m-3(10a)-1),and has the largest increase in May(0.027 m3.m-3(10a)-1).The difference of soil moisture between different eco-geographical regions is significant.The highest soil moisture is VA6,and the low value is located at HIID1 and HIID2.Due to the factors such as vegetation coverage,soil characteristics and topography,the response of soil moisture to precipitation in different eco-geographical areas is also different.The highest correlation between soil moisture and precipitation is the HIBI.It also shows that the alpine meadow has a good water conservation function,which is of great significance to the hydrological cycle in the region.