Soil Texture Data Fusion Based On Ensemble Learning And Its Impact On Noah-MP Soil Temperature And Humidity Simulation

As a very important surface characteristic parameter,soil texture directly affects soil hydrological parameters such as saturated soil water content,saturated hydraulic conductivity and saturated soil matrix potential,as well as soil thermodynamic parameters including soil heat capacity and soil thermal conductivity.Using high quality soil texture data can effectively improve the accuracy of model simulation.Therefore,this study proposes a method to simulate and develop regional soil texture data in China based on ensemble learning,which uses meteorological data,terrain data,vegetation data and other soil forming factors as covariates,and regression simulation of sand,silt and clay data is carried out based on Gradient Boosted models(GBM).Meantime,Generalized Linear Model(GLM)is used as meta learner to fuse the GBM simulation results with the existing high-quality soil texture data to obtain the regional soil texture data of China for Noah-MP.The data are applied to CLDAS/Noah-MP system to study the effect of soil texture change on the simulated soil temperature and humidity.The main conclusions are as follows:(1)The precipitation,surface temperature,NDVI,MODIS seven bands of surface reflectance,land use type,longitude and latitude,elevation,slope,aspect and soil type as covariate factors,are used for GBM model to simulate the regional sand,silt,clay content in China.GLM was inturduced to fuse GBM simulation results,soil mechanical composition data based on random forest and soil mechanical composition data based on polygon linking method.The root mean square errors(RMSE)of sand,silt and clay are 0.127,0.098 and 0.075,respectively.Compared with GBM results,RMSE of sand,silt and clay are reduced by 0.022,0.015 and 0.028,respectively.Among them,the improvement of clay is the most obvious.Using the soil texture triangle map made by the US Department of Agriculture(USDA),the fusion results were calculated as soil texture.Compared with the Noah-MP model's own soil texture(FAO),it was found that the proportion of different soil textures in the two soil texture types was quite different,especially the spatial distribution of sand,silt,clay soil and sandy clay content.On the whole,GBM/GLM soil texture has a more reasonable spatial distribution.(2)In this study,the soil texture obtained in this paper(GBM/GLM)and the soil texture of the model(FAO)for CLDAS/Noah-MP were designed to simulate the regional soil moisture in China at the depth of 0-10cm and 10-40cm in 2014.The results showed that the change of soil texture could significantly affect the simulation results of soil moisture.The simulation effect of GBM/GLM was better than FAO.The correlation coefficients of 0-10cm and 10-40cm depth were increased by 0.035 and 0.036,respectively,and RMSE was decreased by 0.004 m~3/m~3 and0.003 m~3/m~3,respectively.From the bias of simulation,GBM/GLM improves the over estimation in the 0-10cm soil layer in summer by FAO to a certain extent,but generally under estimates in the depth of 10-40cm,which depends on the improvement of the model parameterization scheme.For the eight regions of China,it is found that the accuracy of GBM/GLM has been improved in most regions,especially in the southeast region,the correlation coefficients of the first layer and the second layer are increased by 0.016 and 0.041respectively,and RMSE is reduced by 0.008 m~3/m~3 and 0.005 m~3/m~3 respectively.(3)The two experiments of GBM/GLM(by this paper)and FAO for CLDAS/Noah-MP were designed to simulate effects of the soil texture on regional soil temperature in China at the depth of 0-10cm in 2014.The results showed that the difference between the two groups of experimental simulation values ranged from-0.25?to 0.13?.The soil texture changes could affect the simulation results of Noah-MP for soil temperature.The correlation coefficient of GBM/GLM was 0.937,and RMSE was 0.9091?.Compared with FAO,the correlation coefficient increased by 0.004,and the root mean square error decreased by 0.0311?.On the daily scale,the correlation coefficient of GBM/GLM is higher than that of FAO in 308 days,and RMSE is also lower than that of FAO in most days.On the monthly scale,except February,November and December,the results of CLDAS/Noah-MP simulation are slightly overestimated.RMSE of GBM/GLM was lower in January,February and October December.On the seasonal scale,the effect of the two groups of experiment simulation is the best in autumn and worst in winter,and the correlation coefficient is only 0.84,which is obviously underestimated.The correlation coefficients of the other three seasons are greater than 0.98.On the whole,GBM/GLM soil texture improved the CLDAS/Noah-MP soil temperature simulation except in summer.