Titolo della tesi: An innovative approach based on Deep Learning - Convolutional Neural Networks for the integrated analysis of the effects on the Territory and Infrastructures due to Natural Hazards, applied to Mt. Etna (Italy); Case study Galeras volcano (Colombia)
Volcanic eruptions eject huge quantity of magmatic hot materials that represent a major danger for the surrounding rural and metropolitan areas and need a continuous survey. Remote sensing monitoring proves to have remarkable potentialities for tracking the evolution of pyroclastic plumes and lava flows during volcanic activity, in addition deep Learning is a powerful technique in artificial vision that can offer great results, beating human capacity.
A methodology for observing and quantifying eruptive processes on Mt. Etna (Italy) was already developed by integrating satellite and radar data and images acquired from in-situ visible and thermal video cameras from ground-based network (Etna_NETVIS), demonstrating the possibility to obtain in almost real-time updated maps of an ongoing lava eruption.
Based on the above mentioned developments, in this P.h.D thesis a similar methodology was applied proposing an innovative approach creating a synergy among Deep learning, monitoring technics and human resources, to mapping pyroclastic plumes and flows phenomena produced by an explosive eruption, like those produced by the Galeras volcano.
The evolution of the ash column were mapped training RGB In-situ and satellite images in different Deep learning models based in Convolutional Neural Networks (CNN), those models allowed to make image classification and semantic segmentation, it should be noted that those images were previously labeled, normalized and standardized using Image Editor and python PIL lib; And on the other hand, to generate 3D training datasets a eruptive virtual scenes ware created using Blender 2.9.2, those 3D input dataset were used to train using a novel Generative adversarial networks 3DGAN model It allow to build a reconstruction 3D from from 2D images, all this process allowed to calculate geometrical parameters as height, density and fallout/flow direction. By adopting satellite images we enhance the observational capability of standard surveillance activities based on ground data. The data acquired from ground-based sensor networks allows to downscale the information derived from satellite data and to integrate the satellite datasets in case of incomplete coverage or missing acquisitions. The performed analysis were published into WebGis in clear and understandable format that enables to assess the integrated capability of ground-based and satellite sensors to monitor the evolution of fast evolving pyroclastic plumes and flows propagating above and along the volcano slopes, for tracking ashes fallout and hot flows and forecasting their dangerous impact in rural and metropolitan areas.