Combination of Bagging and Neural Network for Improving Precipitation Estimates Using Remote Sensing Data

Artificial intelligence is widely used today in several areas. Models based on machine learning have shown interesting performance in automatic classification. Although these different machine learnings have achieved good accuracy, further improvements can be made by using ensemble learning methods. In this work, we use Bagging to try to improve the classification performed by a multilayer perceptron (MLP). To do this, a large database is used, allowing the concept of Bagging to be applied. The data from the MSG geostationary satellite (second-generation weather satellite) in spatiotemporal coincidence with the data from the radar are divided into subsets. Thus, several MLPs are learned and applied. They make it possible to classify precipitation into three classes: convective, stratiform, and no-rain. The final decision corresponds to the class that appears the most, given by the majority of MLPs. The combination between Bagging and MLP shows a clear improvement compared to the use of standard MLP. The corresponding precipitation estimates are also improved.