Detección de alertas pre-eruptivas volcánicas basada en aprendizaje incremental

Abstract

Background: Volcanoes are geological structures that generate emergency situations for those who live in their environment. The risks to which the population is exposed (earthquakes, flows, explosions, emissions of gases and ashes, etc.) cause morbidity and high mortality due to the size of large eruptions. Indirectly, volcanic events can cause socioeconomic deterioration, damage to vital transport lines and infrastructures, in general, alter the living conditions of the populations involved. For this reason, volcano monitoring is a key task to detect volcanic anomalies in real time and act accordingly. On the other hand, from the computer science we know that the automatic learning techniques have been positioned as tools to solve various real-life problems, such as classification and detection of intruders, monitoring of industrial processes, among others. These techniques have better performance when data from the a priori domain is available, where an automatic classification algorithm is trained with the data set obtaining a model capable of creating a classification or prediction with a high percentage of accuracy.

Objectives: To develop a system that allows the detection of pre-eruptive alerts from the detection of volcanic anomalies, which is able to deal with data flows coming from its monitoring stations, allowing this detection to maintain an acceptable precision.

Methods: It is proposed to use an atypical value detection algorithm that implements incremental learning so as not to store all the examples of the data flow coming from the deformation and volcanic geochemical stations and to update the model function every time changes occur. In this way, it is intended to calculate the outliers in real time and generate the respective alerts, which will be classified by experts of volcanic monitoring as the case may be.

Results: The present proposal delivered as results a series of data sets that involve vulcanological information pertaining to the areas of volcanic monitoring of geochemistry and deformation. These data were collected through inclinometry and carbon dioxide stations, located near the Puracé volcano (department of Cauca), in addition, a prototype is delivered capable of detecting in real time the different anomalies generated in the volcano.

Conclusions: The domain of application used in the present investigation showed that using the RDE algorithm (Recursive Density Estimation) in volcanic monitoring is a good option to find outliers and generate alerts that allow experts to know the anomalies that are occurring in the volcano in real time.