Tesis y Trabajos de Investigación PUCP

URI permanente para esta comunidadhttp://54.81.141.168/handle/123456789/6

El Repositorio Digital de Tesis y Trabajos de Investigación PUCP aporta al Repositorio Institucional con todos sus registros, organizados por grado: Doctorado, Maestría, Licenciatura y Bachillerato. Se actualiza permanentemente con las nuevas tesis y trabajos de investigación sustentados y autorizados, así como también con los que que fueron sustentados años atrás.
Ingresa a su web: Repositorio Digital de Tesis y Trabajos de Investigación PUCP

Explorar

Filtros

20241

Ajustes

Autorsearch.filters.applied.operator.equals: search.filters.author.Machacca Puma, Roger

Resultados de búsqueda

Mostrando 1 - 1 de 1
  • No hay miniatura disponible
    ÍtemTexto completo enlazado
    Forecasting volcanic eruptions based on massive seismic data processing. Application to Peruvian volcanoes
    (Pontificia Universidad Católica del Perú, 2024-11-08) Machacca Puma, Roger; Tavera Huarache, Hernando Jhonny; Lesage, Philippe
    This dissertation investigates the potential improvement of volcanic eruption understanding and forecasting methods by using advanced data processing techniques to analyze large datasets at three target volcanoes (Piton de la Fournaise (PdlF) (France), Sabancaya, and Ubinas (Peru)). The central objective of this study is to search for possible empirical relationships between the pre-eruptive behavior of the accelerated increase in seismic activity using the Failure Forecast Method (FFM) and velocity variations measured by Coda Wave Interferometry (CWI), since both observations are reported to be independently associated with medium damage. The FFM is a deterministic method used to forecast volcanic eruptions using an empirical relationship of increased and accelerated evolution of an observable (e.g., volcano-seismic event rates). The event rates used with FFM in this study were generated using two Deep Learning (DL) based models. The detection model (VSDdeep) is based on EQTransformer and the classification model (VSCdeep) consists of a simple convolutional neural network that uses the short-time Fourier transform of the detected signals as input data. VSDdeep, trained on ∼16.3 k volcano-seismic events, outperforms previous DL-based models, achieving an accuracy of 97.68%. The VSCdeep model was trained on two datasets, one for effusive volcano (7 classes) and a second for explosive volcanoes (10 classes), and achieved accuracies of 96.55% and 90.5%, respectively. The combination of the two DL-based models detects and classifies 1.5 times more volcano-tectonic (VT) events than the catalog provided by the local observatories. A Bayesian approach of FFM was applied to study the 27 eruptions recorded between 2014- 2021 at PdlF volcano. The analysis shows that 23 (85.2%) of the precursory sequences are suitable for retrospective application of the FFM. Eight eruptions fulfilled the reliability criteria. Only seven eruptions (25.93%) were successfully predicted in the real-time scenario, but when the reliability criteria are met, the successful prediction rate increases to 87.5%. For Sabancaya volcano, the FFM cannot be applied because the explosions are not preceded by significant increases in seismicity. In the case of Ubinas volcano, LP event rates were used, with a successful forecasting rate of 4.55% with real-time criteria of the 330 explosions analyzed, showing a low forecasting rate for these two Peruvian volcanoes. We report long-term (over 22 years) apparent velocity variations (AVV) that appear to be related to the frequency of occurrence of magmatic intrusions. However, the simple dike-intrusion model tested in this study does not explain this long-term pattern. The short-term pre-eruptive velocity variations generally show two phases: A first phase corresponding to a slight velocity decrease ∼5 days before the eruption, and a second phase of sudden velocity decrease one day before the eruption. The precursory behavior of AVV indicates that 12 eruptions (44.4%) were preceded by AVV ≥ 0.15% observed at least one day before the eruption. Two models were tested to explain the pre-eruptive velocity variations. One is based on the cumulative rock damage associated with VT activity. The other one takes the effect of the dike intrusions into account. However, in both cases, the comparison of the observed and modeled AVV amplitudes shows low regression coefficients. This indicates that the generation of velocity variations is complex and that these simple models alone cannot explain the observations. The statistical analysis of accelerated VT event rates and AVV precursors at PdlF volcano indicates that 37% of the eruptions were preceded by both precursors, 48.2% by one of the two precursors, and 14.8% were not preceded by either. These findings suggest that both precursors, accelerated VT event rates and AVV, can serve as potential tools for early detection of volcanic unrest at this volcano.