CLDAS Soil Moisture Downscaling Based On Ensemble Learning Method And Its Application In Drought Monitorin

Soil moisture is the key parameter of water cycle and energy balance in terrestrial ecosystems.The land surface data assimilation system can obtain spatio-temporal continuous soil moisture data,but its further application is limited due to its low spatial resolution.Therefore,the soil moisture data of the China Meteorological Administration Land Surface Data Assimilation System(CLDAS-V2.0)were used to analyze the temporal and spatial distribution characteristics of soil moisture at different depths(0～10,10～40 and 40～100cm)in North China in 2019.The maximum information coefficient and gray correlation analysis method were introduced to objectively quantify the impact of downscaling factors on soil moisture,using 0-10 cm soil moisture as a representative layer.The downscaling factors with higher impact were optimized on this basis.A downscaling model of soil moisture from 0to 10 cm was constructed using four methods,including gradient boosting machine,deep feedforward neural network,random forest and Stacking ensemble learning,to downscale the spatial resolution from 6 km to 1 km.The accuracy of the downscaling results were evaluated based on station observation data,and a comparative study was carried out on the soil relative moisture calculated by the data before and after downscaling and the drought monitoring in typical areas of Henan Province.The main conclusions are as follows:(1)In terms of time,the distribution of 0～10 cm soil moisture in North China had a very significant annual change.As the soil layer thicken,the curve fluctuation tend to gradually smooth out.Spatially,the soil moisture in each layer shows the spatial distribution characteristics of high in the east,low in the west,high in the south and low in the north.The maximum information coefficient and grey correlation analysis method are introduced to objectively quantify the impact of downscaling factors on 0～10 cm soil moisture.Integrating the results of correlation coefficient,maximum information coefficient and gray correlation analysis method with the actual locational conditions in North China,the study selected surface temperature(LST),surface albedo(Albedo),elevation(DEM),soil texture(Clay,Silt,Sand),and normalized difference water index(NDWI)as downscaling factors for soil moisture downscaling.(2)In North China,the four soil moisture downscaling results of gradient boosting machine,deep feedforward neural network,random forest,and Stacking ensemble learning had similar patterns to the spatial distribution of the original soil moisture.The soil moisture in the south and coastal areas was high,while that in the middle and north was low,and the average soil moisture was more than 0.2 m~3/m~3.All four different downscaling methods had effectively improved the spatial resolution and accuracy of CLDAS soil moisture products,which may be related to the fact that the downscaling model input parameters also include soil moisture information at some stations.The downscaling process had assimilation,so that the downscaling can further improve the data accuracy at the same time.The absolute deviation of all four methods was smaller than that of CLDAS products,which improved the overestimation phenomenon to a certain extent.The accuracy of downscaling soil moisture was in the order of Stacking ensemble learning,random forest,deep feedforward neural network,and gradient boosting machine.The downscaling results of both the original soil moisture and the four different downscaling methods can better reflect the daily variation characteristics of soil moisture.However,there was an overestimation of the original soil moisture on most days,and the four downscaling results were underestimated to a certain extent.Overall,the Stacking ensemble learning method was optimal.Compared with the original soil moisture,the average correlation coefficient was increased by 0.13,and the error and deviation were also reduced.(3)In terms of spatial distribution,both CLDAS＿Rsm(Rsm,Relative soil moisture)and Stacking＿Rsm showed a trend of decreasing from east to west,and the low value areas were mainly located in the south and west of North China.The average relative soil moisture was Stacking＿Rsm was better than CLDAS＿Rsm decreased by 8%.Stacking＿Rsm effectively improved the stations with larger root mean square error in the western and northern regions.In terms of time series,the overall curve trends and fluctuation frequencies of CLDAS＿Rsm and Stacking＿Rsm were more consistent with the observed daily averages,but CLDAS＿Rsm was higher overall.Over the whole growing season,the overall average improvement of Stacking＿Rsm correlation coefficient was 0.087,the root mean square error was lower than that of CLDAS＿Rsm,and the overall reduction of deviation was 0.094.In the monthly scale variation,Stacking＿Rsm outperformed CLDAS＿Rsm in all three indicators of correlation coefficient,root mean square error and deviation.(4)The study analyzed and compared the drought distribution characteristics in Henan Province from April to October 2019.The distribution range of CLDAS＿Rsm partial dry areas was significantly smaller than that of observations and Stacking＿Rsm.While the spatial variation trend of Stacking＿Rsm remained basically consistent with that of observations,its spatial distribution details were richer.At the ten day scale,the drought frequency in the central part of North China was low,and the frequency of drought in the surrounding area was high.The centers of drought frequency were mainly located in Nanyang,Zhumadian and Zhoukou in Henan Province,and Zhangjiakou area and northern Chengde in Hebei Province.The frequency of light drought was significantly higher than the other three levels,and all levels of drought occurred.Combined with the spatial distribution of drought intensity under the ten-day scale,it can be seen that droughts was more frequent in northeastern,central,eastern,southwestern and southeastern parts of North China,but severe and exceptional droughts occurred less frequently.Thus,the drought intensity was light drought.In southern Henan Province,western Hebei Province,and northern areas,droughts were frequent and the drought conditions were basically light-moderate drought with small-scale severe drought and above.