The Prediction Of Flash Droughts Over Xiang River Basin Based On SVM-EnKF And Multi-source Remote Sensing Soil Moisture

Flash drought is a rapid-onset drought,which can greatly threaten the agricultural production and economic development.As a new type of drought,the spatio-temporal characteristics and developmental mechanism of flash drought are still not completely understood.Compared with the traditional drought,it has rapid onset time and short lead time,which brings unique challenges for its monitoring,prediction,and mitigation.Therefore,the purpose of this study is to investigate the spatio-temporal characteristics,driving factors and the prediction of flash drought,taking Xiang River Basin as an example,and to reduce losses and risks for agricultural production.Reliable soil moisture is the key for the prediction of flash drought.Firstly,three remote sensing soil moisture products,namely SMAP,ASCAT and AMSR2,are evaluated over Xiang River Basin.Based on the evaluation results,products with relatively higher accuracy among these products are selected for the data fusion process with Bayesian model averaging（BMA）algorithm,ensemble Kalman filter（En KF）algorithm and simple average method,respectively.Secondly,the meteorological variables are adopted as the inputs of the support vector machines（SVM）for the multi-layer（0-5cm,0-10 cm,10-40 cm and 40-100cm）soil moisture prediction.And to improve the prediction efficiency of SVM,the above three remote sensing soil moisture and the fusion soil moisture are added individually in the inputs.Besides,the En KF technique is applied to be coupled with SVM to probe its capability in improving the performance on soil moisture prediction.The effectiveness of coupling En KF technique with SVM and the applicability of the remote sensing products in soil moisture prediction are investigated by comparing the soil moisture prediction of multiple cases.Finally,the spatio-temporal characteristics and the influencing factors of flash droughts over Xiang River Basin are investigated,compared with that over Wei River Basin,which is in different climate conditions;the development mechanism of flash drought is explored by analyzing the anomalies of meteorological variables in the evolution of flash drought;based on the comparison among the above designed cases,SVM and SVM-En KF are applied to predict flash drought using SMAP and meteorological variables as inputs.The main conclusions are as follows:（1）SMAP shows the highest accuracy in the Xiang River Basin among the three selected remote sensing soil moisture products,and ASCAT performs better than the AMSR2 at both basin and grid scales.The fusion data obtained based on the three fusion algorithms show high accuracy both at the grid scale and the basin scale,which can better capture the dynamic characteristics of the reference data.Compared with the original soil moisture,the accuracy of the fused soil moisture is obviously improved.（2）The soil moisture prediction efficiency of SVM with meteorological variables as inputs is satisfactory for the surface-layers（0-5cm and 0-10cm）soil moisture,while poor for the deeplayers（10-40 cm and 40-100cm）.Adding SMAP or fusion soil moisture as input to SVM can improve the soil moisture prediction efficiency,with more than 36% increase in the R-value,at least 140% in NSE,and at least 6% reduction in RMSE for every layer.However,adding ASCAT or AMSR2 has no obvious improvement for its performance.Coupling En KF can significantly improve the performance of SVM in the soil moisture prediction of both surface and deep layers.The increase in R-value and NSE are respectively above 80% and 50%,while the reduction in BIAS and RMSE is respectively more than 90% and 63%.Adding remote sensing soil moisture products as inputs can no further improve the performance of SVM-En KF.However,SVM-En KF can also eliminate the influence of extreme values of remote sensing soil moisture and predict soil moisture accurately.（3）The flash droughts are sub-month to month scale events over Xiang River Basin,with almost all events are in summer and autumn.The deficit on precipitation and the high temperature are the main driving factors for flash drought development,the anomalies of which in a relatively short-period（15 days to 25 days）is closely related to the development of flash drought.The prediction of flash drought conducted using SVM and SVM-En KF,with meteorological variables and SMAP as inputs,has an overall satisfactory performance.In the training stage and the test stage,more than 70% of events can be captured by both models.However,false alarms exist in the prediction result using both models,with false alarm ratio（FAR）ranges from 0.06 to 0.30.