Deep Learning for volcanic risk assessment

The large amount of lava outflows during effusive eruptions can cause profound morphological changes, affecting both the natural and inhabited environment, destroying buildings, agricultural fields and important infrastructures such as roads, power lines, aqueducts and even modified the coastline. The ongoing demographic congestion around volcanic structures increases the potential risks and costs that lava flows represent and leads to a growing demand for implementing effective risk mitigation measures. Therefore, it is important to assess the elements at risk in volcanic areas to establish the mitigation actions to reduce the lava flow risk. Risk management for volcanoes is not just an emergency response to save lives but is also important in terms of economic loss. However, the collection of data regarding exposed elements surrounding the volcanoes is a lengthy and time-consuming process but utilizing the satellite images together with several machine learning techniques helps address this goal. We propose a cloud based platform in Google Colab using Land Use and Land Cover (LULC) classifiers to automatically assess the elements at risk by exploiting freely available high spatial resolution satellite images. This procedure allows to get an updated map of elements at risks in volcanic areas worldwide and will allow to routinely update the exposure map and thus risk map. In fact, up-to-date risk maps are fundamental to reaching the optimal decision in case of any hazard and crisis and can help us add or delete critical zones around the volcano. Using the freely available Sentinel 2-Multispectral Instrument (MSI) images and deep learning models, we aim to test the LULC applicability to a variety of volcanic areas whilst comparing the performances of two Convolutional Neural Network (CNN) architectures, namely VGG16 and ResNET50.