Application of Dempster-Shafer theory for optimization of precipitation classification and estimation results from remote sensing data using machine learning

Abstract

In this work, a technique based on Dempster-Shafer Theory (DST) is proposed to combine three classifiers, namely Artificial Neural Network (ANN), Support Vector Machine (SVM) and Random Forest (RF) in order to improve the precipitation estimation over the North of Algeria. Data from SEVIRI (Spinning Enhanced Visible and Infrared Imager) radiometer on board MSG (Meteosat Second Generation) is used. First, the three classifiers were applied for the classification of instantaneous precipitation scenes into three classes (stratiform class, convective class or no-rain class). Secondly, to apply the DST and to calculate the mass functions, normalized probabilities from the classifications are determined. The mass functions of the different classifiers were combined using DST. Maximum plausibility or maximum belief are calculated and used for the final assignment decision.The DST-ANN/SVM/RF method shows interesting performances compared to the separate use of the three classifiers. The proposed method has also been compared to some techniques combining two, three and six classifiers based on machine learning. The results show that Dempster-Shafer Theory performs better than other combination techniques. The estimation results are well correlated with the reference radar measurements, with a correlation coefficient of 0.96 for the developped technique against 0.94 for the technique combining six classifiers, 0.90 for the technique combining three classifiers, 0 0.87 for the technique combining two classifiers and 0.85 for the technique with a single classifier.