Improvement of rainfall estimation from MSG data using Random Forests classification and regression

In this study, a new rainfall estimation technique on 3 h and 24 h scales applied in Northern Algeria is presented. The proposed technique is based on Random Forests (RF) algorithm using data retrieved from Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI). Because the rain rate depended on the precipitation type: convective or stratiform, the RF technique is divided into two stages. The first is the classification of rainfall into three classes (no-rain, convective and stratiform) using RF classification and the second consists in assigning rain rate to the pixels belonging to the two classes (convective and stratiform) using RF regression. In classification step, spectral, textural and temporal features of clouds are used as predictor variables and the results are validated against co-located rainfall classes observed by radar. The statistical parameters score shows stronger rainfall classification performance for RF compared to the ANN and SVM. The RF regression model is validated by comparison with against co-located rainfall rates measured by a rain gauge. The results show rain rates estimated by the developed scheme are in good correlation with those observed by rain gauges.