Detección digital de Sistemas Convectivos de Mesoescala a partir de imágenes meteorológicas multiespectrales

Abstract

The exact identification of Mesoscale Convective Systems (MCS) is not a simple task. In this work, Support Vector Machines (SVM), Decision Trees (DT) and Random Forests (RF) non-parametric machine learning algorithms (ML) were applied to detect MCS, from a series of GOES-13 weather satellite images acquired on April 03, 2013 every half hour from 11:45 to 22:15 hours, Coordinated Universal Time (UTC), covering the Colombian territory. The results obtained by these methods were compared with a traditional method referred to as brightness temperature (BT). Accuracy assessment was conducted using STEP (shape, theme, edge, position), a method that evaluates geometric and thematic similarity between objects, using as reference a dataset of high accuracy data extracted from images of precipitation Tropical Rainfall Measuring Mission (TRMM). The aim of this study was to determine whether using information from several spectral channels of weather images, rather than from a single infrared channel (IR) as traditional techniques do, allows accurate detection of MCS. Experimental results show that the Decision Tree (DT) and Random Forest (RF) algorithms performed better than the IR-TB algorithm to detect MCS, while that the results of SVM algorithm, suggest that it use may not be favorable for practical applications. The decision criteria of the classification model yielded by DT could be replicated several times in different dates without performing visual interpretation in each image, being very useful for operational applications under the approach presented here.