Research On Downscaling Of SMAP Soil Moisture Products Based On Google Earth Engine

The monitoring of the temporal and spatial changes of soil moisture is essential for the sustainable development of agriculture and forestry,water resources management,and drought monitoring.The temporal variability of soil moisture provides good information on crop water requirements and helps improve agricultural water management.The soil moisture data obtained by the traditional manual single-point actual measurement method not only consumes manpower and financial resources,but also is not representative in the region and has poor real-time performance.With the development and application of remote sensing technology,passive microwave remote sensing can provide all-weather,all-weather,and high-time resolution soil moisture data,but the spatial resolution is relatively low(several kilometers to tens of kilometers),which is suitable for large areas.Drought monitoring and water resources management in China have certain reference significance,and cannot meet the fine and medium-scale applications such as agricultural irrigation and water resources management.In order to solve the problem of the spatial data of soil moisture remote sensing products,it has become one of the research hotspots of scholars at home and abroad to use multi-source data to downscale them to generate soil moisture data sets with high spatial resolution.In this paper,the Lightning River Basin is selected as the research area,based on the Google Earth Engine(GEE)cloud platform,using the 10 km SMAP passive microwave soil moisture data provided by the platform and other remote sensing data with a resolution of 1km to analyze the correlation between characteristic parameters and soil moisture.On this basis,NDVI,EVI,LAI,LST,ET,ATI,DNWI were selected as artificial neural network(ANN),multiple linear regression(MLR),random forest(RF),support vector machine(SVM)machine learning method soil The input variables of the water downscaling study were used to verify the downscaling results using the measured soil moisture data of the site of the Lightning River Basin "Water Cycle and Energy Balance Remote Sensing Comprehensive Experiment in the Lightning River Basin",and the following conclusions were obtained:(1)Analyze the correlation between 11 surface parameters such as land surface temperature(LST),normalized vegetation index(NDVI),normalized water index(NDWI),elevation,and slope aspect in summer in June,July,and August,and soil moisture products The results show that the correlation from high to low is vegetation cover>weather> topographic factors,NDVI(R=0.600),ET(R=0.574),LST(R=-0.468),aspect(R=0.44),ATI(R=0.459)shows a good correlation and is used as an input variable in the study of machine learning downscaling methods..(2)Based on traditional regression models: global multiple non-linear regression model and local spatial weight decomposition model downscaling soil moisture products are verified by measured site data,and the R and RMSE of the two models at the network scale are respectively 0.362 and 0.139 m~3/m~3,0.507,0.102m~3/m~3;R and RMSE on the point scale are 0.342,0.067m~3/m~3,0.500,0.038m~3/m~3,respectively.Both models can achieve the purpose of improving spatial resolution in the study of soil moisture downscaling At the same time,the soil moisture after downscaling is closer to the measured value.And in the study area,the spatial weight decomposition model performs better than the multiple nonlinear regression.(3)Comparing four different machine learning algorithms of multiple linear regression(MLR),random forest(RF),artificial neural network(ANN)and support vector machine(SVM)to build a downscaling model,the results show that the MLR model has the best modeling Accuracy,three different auxiliary data combination downscaling results: R are 0.56,0.58,0.53;RMSE are 3.654 mm,3.628 mm,3.678 mm,but the robustness of the algorithm needs to be further improved;RF is the second,the same It has good soil moisture downscaling ability,and the characteristic important index calculated in the RF model shows that LST and DNVI are dominant.The ANN and SVM methods have similar downscaling accuracy(R≈0.35)in the downscaling process.The four machine learning downscaling algorithms all have underestimation of wet areas and overestimation of dry areas.The overestimation of the RF model is the most obvious..The combination of three different auxiliary data shows that the addition of terrain factors can effectively improve the accuracy of soil moisture downscaling.