Deep Learning for Weather Classification from a Meteorological Device

Abstract

In order to realize climate prediction or weather forecast, there exist large and expensive meteorological stations which monitor various weather variables, such as temperature, wind speed, humidity, etc. However, these prediction systems are deployed to monitoring urban zones or large population areas. Therefore, predictions for communities far from urbanization are in the practice, imprecise. Currently, these inaccurate predictions for the climate changes, affect the agriculture in several areas, due to inadequate planning by the farmer, which is based on a priori knowledge that the inhabitants have with the experience from the observation of the climate behavior, action that is highly imprecise and unpredictable. Therefore, in this work, for a more accurate weather forecast in rural regions, a portable low-cost meteorological device is proposed, which using suitable sensors measures and record the weather variables such as temperature, humidity of the air, luminosity, rain, humidity, and atmospheric pressure. Using these weather data, a classification of the target class set {rainy, cloudy, sunny}, is made based on the parameters obtained through the device. Then, using a combinations of assembled algorithms with deep learning, optimum results are obtained with the following classifiers: MultiClassClassifier$+$Multilayer perceptron, using the sampling criteria 2/3-1/3, cross validation and representative sample. The classification results are comparable and competitive, with respect to those reported in the state-of-the-art, and stands out by distinguishing the target classes with a high degree of precision.