Development of a hybrid classification technique based on deep learning applied to MSG / SEVIRI multispectral data

Abstract

The approach developed in this paper for the classification of precipitation intensities is based on deep learning of neural network. Multispectral data from the MSG satellite (Meteosat Second Generation) providing information about the cloud's physical and optical characteristics are exploited and used as inputs to a deep neural network model. The model is a combination of CNN (Convolutional Neural Network) and DMLP (Deep Multi-Layer Peceptron) which is learned and validated by comparison with the corresponding Radar data during the rainy seasons 2006/2007 and 2010/2011 respectively. The CNN extracts spatial characteristics from MSG multi-spectral images. Then, the set of spatial and multi-spectral information are used as inputs for the DMLP. The results show an improvement compared to the three other classifiers (Random Forest, Support Vector Machine and Artificial Neural Network). The CNN-DMLP method was also compared to the technique combining the three classifiers (SAR). The results indicate a percentage correct (PC) of 97% and a probability of detection (POD) of 90% for CNN-DMLP method compared to 94% and 87% for of the SAR technique, respectively. In terms of bias, the CNN-DMLP method gives 1.08 compared 1.10 for SAR technique.