Satellite-Based Rainfall Estimation through Semi-supervised Learning

Satellite images have been extensively used for rainfall estimation predictive models based on pattern recognition techniques, even so unsupervised and supervised. However, most of these kind of data are unlabeled, and the acquisition of labeled data for a learning problem often requires a skilled human agent to manually classify training examples. In this paper we introduce the use of semi-supervised support vector machines for rainfall forecasting using images obtained from visible and infrared NOAA satellite channels. The semi-supervised learners combine both labeled and unlabeled data to perform the classification task. Two experiments were performed, one involving traditional SVM and other using semi-supervised SVM (S³VM). Comparisons among artificial neural networks using multilayer perceptrons are also presented. The S³VM approach outperforms SVM in our experiments, with can be seen as a good methodology for rainfall satellite estimation, due to the large amount of unlabeled data. The accuracies obtained for SVM and S³VM were, respectively, 90.6% and 95.96%.