Study On Drought Monitoring Modeling IntegratedMulti-source Remote Sensing Data

As a global natural disaster, Drought has been the focus of researchers around the world since long ago. In recent years, a comprehensive method Integrating ground and remote sensing data for drought monitoring and index drought construction has received lots of response and recognition of researchers. In this paper, applicability of drought index in different time scales was evaluated in the beginning and followed by drought monitor modeling with several data mining (DM) algorithms. Every model was assessed in the end of this paper for their applicability. Three algorithms were Artificial neural networks (ANN); support vector machine (SVM) and Classification and Regression Tree(CART) Respectively. And drought indices used in this paper were Standardized Precipitation Index (SPI); standardized precipitation evapotranspiration index (SPEI) and CI. The conclusions of this paper is as followed:(1) The parameters of model influence the model most. When it came to parameter optimization, an interface based on Lib-SVM tools in Matlab was used for models with SVM method. Better results were gained by this interface than trial method which was used in parameter optimization for ANN and CART modeling. During the parameter optimization, result gained by CART differs little with parameter changes.(2) Both in the precision evaluation based on drought indices calculated by meteorological data and soil moisture, results got by models established by SVM method all got smaller RMSE than the others. And in the drought monitoring maps got by SVM models in July 2010, linear disruption coincident with ridge lines was found which showed the sensitity of SVM models to the elevation in mountainous regions. Models established by CART was slightly worse compared to the other two methods.(3) With the comparison of the drought monitoring map with the remote sensing images obtained from the models, the following results was observed:CI models got smaller severe drought areas due to its consideration of longer time before assessment. On the opposite, SPEI models got bigger severe drought areas due to its consideration of higher temperature.in July. The drought index selected as dependent variable in modeling affects the model a lot.(4) According to the results of model generalization evaluation, model parameters impacted the generalization ability most. In the models established by ANN, models built with Quick building had best generalization. The parameter C in SVM method had a major impact on the generalization ability.