Research On Estimation Of Cropland Surface Soil Moisture Based On Remote Sensing And Ground Measurements

Soil moisture is the key factor for predicting agricultural drought and monitoring crop growth status.In current soil surface moisture estimation practices,commonly used regression models combined with optical/near infrared/thermal infrared remote sensing data and ground truth measurements for soil moisture retrieval show some advantages in acquiring specific regional soil moisture details and spatial patterns.However,the accuracy of most models is largely determined by the quantity of training samples and the limitation of correlation extraction in regression models,which are unsuitable for larger-scale projections(e.g.,national scale).This paper solves these issues according to the following aspects.First,to solve the difficulty in the indirect soil moisture estimation by using multi-spectral data,the cropland surface soil moisture method was proposed by modifying the classic 'Universal Triangle' method and considering the crop growth stages.Specifically,we extended the two parameters-fractional vegetation coverage(FVC)and land surface temperature(LST)to three parameters-FVC,LST and land surface albedo(LSA)in the'Universal Triangle' method.Considering the dependence of soil moisture estimation on corn growth stages,we incorporated five growth stages-the seeding,seedling,growing,maturing,and harvesting stages in the regression models for soil moisture estimation.These models were built for each stage using five years(2006-2010)of soil moisture data obtained through field measurements combined with daily MODIS based land parameters at 1-km spatial resolution.Our results show that the extended triangle method is strongly correlated to in-situ soil moisture(2011-2012)at each corn growth stage.Moreover,these correlations are dependent on vegetation cover and growth stage.The correlation coefficients between field-measured and model-estimated soil moisture were 0.73,0.67,0.66,0.74 and 0.63 for each growth stage,respectively.The estimated soil moisture were more accurate than the 'Universal Triangle' based soil moisture estimates and SMOS-BEC soil moisture Level 3 products.These findings show that our model is feasible method for estimating surface soil moisture accurately at the 1-km scale for cornfields in the study area.In addition,to estimate surface soil moisture throughout China's cropland,we propose a deep learning regression model.A deep learning based model on a deep feedforward neural network(DFNN)was proposed to estimate soil moisture over China's cropland.The deep learning model has the advantage in representing nonlinearity and modeling complex relationships from large scale data.To illustrate the model,15 years of moderate-resolution imaging spectroradiometer(MODIS)based variables were used as input features:FVC,LST and LSA.Next,in situ soil moisture measurements were linked with model input features at the same location and time,and 37,888 pairs of samples were obtained for model building.Then,model tuning was carried out with 3 parameters:the numbers of hidden layers(Hiddens),Neurons,and Epochs.The result indicates that the semi-physical learning model captured the complex relationship between remote sensing variables and surface soil moisture with average R2 = 0.8939 and MSE = 0.0006 in China.The best model parameters is Hiddens=8,Neurons=500 and Epochs=3500.Our study suggests the potential capacity of the deep learning model for operational applications of soil moisture estimation on national scale.Finally,considering the urgement of near real time soil moisture,we combined direct satellite measured top of atmosphere radiance data with in situ soil moisture measurements to estimate surface soil moisture.According to the relationship between soil moisture and soil spectral reflectance,we selected twelve bands of the second-generation satellite optical/thermal sensor of MODIS-Visible Infrared Imaging Radiometer Suite(VIIRS)raw data records(RDR),which includes ten bands of top of the atmosphere reflectance bands and two thermal emissive bands.Four years(2012-2015)of VIIRS RDR were used as input parameters of DFNN based model for soil moisture estimation.The results of this study demonstrate that the estimated models captured the complex relationship between the remote sensing variables and the in situ surface soil moisture measurements with the adjusted correlation coefficient of R2=0.9812 and root mean square error(RMSE)of 0.0118 in China.These results were more accurate than the SMAP active radar soil moisture products and the GLDAS 0-10 cm depth soil moisture data.Our study suggests that multi sensors(optical/thermal)based method have potential for extending multi sensors(optical/thermal/microwave)applications in the upscaling of in situ surface soil moisture data.