Improving satellite rainfall estimation from MSG data in Northern Algeria by using a multi-classifier model based on machine learning

Abstract

The primary focus in this paper is the estimation of precipitation from MSG images (Meteosat Second Generation) using a machine learning-based multi-classifier model. Learning and validation of the multiclass model is performed using the correspondences between MSG satellite data and radar data. To do this, six classifiers were first combined in order to exploit the full potential of each of these classifiers. These are Random Forest (RF1), Artificial Neural Network (ANN), Support Vector Machine (SVM), Naive Bayesian (NB), Weighted k-Nearest Neighbors (WkNN), and the Kmeans ++ algorithm (Kmeans). The application of these classifiers makes it possible to carry out a classification at level 1. A pixel can therefore be assigned to more than one class by the different classifiers. We calculated six certainty coefficients from these classification results. To refine these results, in a second step, a classification at level 2 was performed using the Random Forest classifier (RF2) taking the certainty coefficients as input parameters. Six classes of precipitation intensities are thus obtained: very high precipitation intensities, moderate to high precipitation intensities, moderate precipitation intensities, light to moderate precipitation intensities, light precipitation intensities and no rain. Comparisons between the results of the multi-classifiers model and those obtained by the classifiers used separately show a clear improvement in the quality of classification. Using multiple linear regressions, precipitation rates for the different classes were determined using data from the rain gauges. To validate the model, precipitation amounts were estimated and compared to actual rain gauge data and then compared to the results of a few precipitation estimation methods. The results are very interesting and show superior performance for the elaborated model. Indeed, the coefficient of correlation has a value of 0.93 for developed scheme against 0.46 to 90 for the different techniques presented here for comparison. Also, a better Bias (2.2 mm) and a better RMSD (12 mm) were obtained for developed scheme while the other methods indicate bias between −11 mm to 16 mm and RMSD higher than 14 mm.