Mineração de regras de associação espaço-temporais temáticas aplicada a imagens de explosões solares

Abstract

Introduction. Space weather analysis is a complex task that involves spatiotemporal data from satellite images added to data from daily bulletins. These data are characterized as time series of georeferenced images and time series of semantic data (alphanumeric data describing the images), respectively. The mining of association rules can aid in the analysis of these data as a mechanism for revealing new and useful standards for the domain expert. However, existing spatiotemporal association rules mining methods are still limited and, as a consequence, they do not adequately meet expectations for extracting patterns that relate spatiotemporal information to images and semantic data. Goal. Therefore, this work aims to support the analysis of space climate from the development of a method of mining of spatiotemporal association rules that allows to relate solar data semantics and visual. The focus is a series of solar images from satellites. Scientific contribution. A new method was developed for extracting significant patterns from satellite imagery series. Called Solar Miner, this method is composed of: a new process of Extraction Transformation Data load - directed to the solar domain - able to work and relate spatiotemporal data with image processing. A new mining algorithm for spatiotemporal association rules, capable of working with this set of data in an acceptable time. And a new classifier that uses the space-time rules to determine the future behavior of new solar data. The proposed mining algorithm advances the current state-of-the-art mining area of association rules by dividing the application of spatiotemporal constraints into two different stages of processing: spatial constraints are applied during the extraction of frequent itemsets and application of temporal constraints during the generation of spatiotemporal association rules. In this way, it is possible to obtain rules that represent the evolution of a given set of events and how they relate to each other. Finally, these rules are used by the associative classifier that was proposed in this work to predict solar behavior based on its current visual characteristics. Results. The proposed method generated rules that were used for the classification, presenting a precision of up to 87.3% in the classification of solar images, being that this value of precision varies with the characteristic extractor used to represent the images. The higher precision (87.3%) was obtained using SURF as extraction of characteristics and the less precision (82.7%) was used the Histogram as extractor of characteristics. The results obtained were analyzed by the domain expert who evaluated how effective and valid the proposed method.