Comparison Between SVR and SVM in Rainfall Estimation from Remote Sensing Data

This study highlights the importance of rainfall estimates from remote sensing data for better water resource management for a smart city. The machine vector support model is used to estimate precipitation from observations from the MSG (Meteosat Second Generation) geostationary satellite in the visible to infrared channels. In this work, we tested the support vector machine model in regression (Support Vector Regression called hereafter SVR) and in classification (support vector machine called hereafter SVM) to estimate the precipitations in three scales, namely, daily scale, monthly scale and seasonal scale. SVR and SVM models are learned and evaluated using measurements from rain gauges. In the case of SVM, we classified precipitation from MSG images into three classes, namely, convective, stratiform, no rain. To estimate precipitation, a rain rate is calculated for each class. In the case of SVR, a regression is carried out between the MSG observations and the amounts of precipitation measured by rain gauge. In both cases, for the seasonal scale, the estimation results are well correlated with the rain gauge measurements. The correlation coefficient is 87% for SVM and 86% for SVR. However, in the short term, the best estimates are obtained using regression unlike classification, which shows limited results. On the other hand, in the long term, the estimation of precipitation by classification shows better performance than the estimation by regression.