Satellite observations are widely used to investigate, monitor, and forecast volcanic activity. Spaceborne thermal infrared (TIR) measurements of high-temperature volcanic features improve our understanding of the underlying processes and our ability to identify reactivation of activity, forecast eruptions, and assess hazards. In particular, thermal changes, indicative of subtle pre-eruptive volcanic thermal activity, have been observed. Over the last several decades, different approaches have been explored to detect and estimate the temperature above background of these thermal anomalies. The most common approach relies on a spatial statistical analysis based on a scene by scene choice of the background temperature region. Satisfactory results have also been shown by using time series anomaly detection algorithms based on statistical profiling approach. Artificial intelligence (AI) is growing sharply in different remote sensing fields because of its capability to automatically learn patterns from the data. Here, we develop an AI approach to automatically detect volcanic thermal features by using spatiotemporal information from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), LANDSAT-8 Thermal Infrared Sensor (TIRS) and MODerate-resolution Imaging Spectroradiometer (MODIS) data acquired over several decades. Our goal is to exploit AI techniques using a combination of high temporal and high spatial resolution satellite data to improve thermal volcanic monitoring and detect very low-level anomalies caused by pre-eruptive activity. We use both low-spatial, high-temporal resolution MODIS data to detect hotter thermal features at short time scales; as well as high-spatial low-temporal resolution ASTER TIR and TIRS data to detect more subtle thermal changes otherwise missed by MODIS. Our analysis is conducted in Google Earth Engine (GEE), a cloud computing platform with fast access and processing of satellite data. The comparison with a new statistical algorithm is documented in a companion abstract in this session.

During an explosive eruption, large volumes of ash and gases are ejected into the atmosphere, forming a volcanic plume which is transported by the wind. The dispersion of volcanic ash in atmosphere represents a threat for aviation safety, whereas the tephra fallout, together with gas emission, may strongly affect population health and damage to environment and infrastructure as well. Volcanic monitoring from space offers now a powerful tool to quantify hazards on both population and air traffic and gain insight into processes and mechanism of violent explosive eruptions. Here we propose a machine learning (ML) algorithm that exploits the Thermal Infrared (TIR) bands of the images acquired by the sensor Spinning Enhanced Visible and InfraRed Imager (SEVIRI), on board Meteosat Second Generation (MSG) geostationary satellite, to identify the components of ash and SO2 gas in a volcanic plume. The detection and assessment of volcanic ash clouds has been performed applying the brightness temperature difference (BTD) approach, between bands at 10.8 μm and 12.0 μm, which highlights the presence of thin volcanic ash, while the algorithm for the SO2 retrieval is based on the contributions given by the bands at 10.8 μm and 8.7 μm. Combining the latter two bands with the 10.8 μm band in the RGB channels, it is possible to create an Ash RGB image, used both day and night for the detection and monitoring of volcanic ash and sulphur dioxide gas. The advantage of the machine learning algorithm is to detect and extract automatically these features from an Ash RGB image. As test cases, we considered the sequence of explosive eruptions occurred at Etna volcano (Italy) in early 2021, which produced very long and high plume columns. Thanks to the high temporal resolution of SEVIRI (one image every 15 minutes), it was possible to visualize and to follow the plumes, from their formation to their complete dispersion in the atmosphere. The comparison of our ML algorithm with the consolidated procedure based on a RGB channels combination in the visible (VIS) spectral range showed a good agreement.

Despite significant advances in monitoring of the development of active lava flow fields, many challenges remain. Timely field surveys of active lava flows could improve our understanding of the development of flow fields, but data of sufficient accuracy, spatial extent and repeat frequency have yet to be acquired. Satellite remote sensing of volcanoes is very useful because it can provide data for large areas with a variety of modalities ranging from visible to infra-red and radar. Satellite sensing can also access remote locations and hazardous regions without difficulty. Radar and multispectral satellite sensing data have been shown that can be combined to map heterogeneous lava flows using machine learning techniques, but a robust general model trained with several different lava compositions has to be developed. Here, we propose a robust, automatic approach based on machine learning techniques for analysing open-access satellite data in order to map lava flows in near-real time applicable to different kind of lava with different thermal components (i.e., incandescent, cooling and cooled lava component). We built a neural network model and trained it with a set of satellite images (e.g., Sentinel-1 SAR, Sentinel-2 MSI and Landsat 8 OLI/TIRS) of recent lava flows, and the relative labels of the lava and background regions. In this way, the trained model becomes capable to detect and map lava flows and to classify any new image, when available. The relative output is a segmented image with lava and background classes, obtained without an analysis made by a human operator. This approach allows to segment lava flows with both hot spot and cooling parts, and to recognize lava flows with different characteristics in near-real time. The results obtained during the long sequence of short-lived eruptive events occurred at Mt. Etna (Italy) between 2020 and 2021 are shown.

Satellite remote sensing is becoming an increasingly essential component of volcano monitoring, especially at little-known and remote volcanoes where in-situ measurements are unavailable and/or impractical. Moreover the synoptic view captured by satellite imagery over volcanoes can benefit hazard monitoring efforts. By monitoring, we mean both following the changing styles and intensities of the eruption once it has started, as well as nowcasting and eventually forecasting the areas potentially threatened by hazardous phenomena in an eruptive scenario. Here we demonstrate how the diversity of remote sensing data over volcanoes and the mutual interconnection between satellite observations and numerical simulations can improve lava flow hazard monitoring in response to effusive eruption. Time-averaged discharge rates (TADRs) obtained from low spatial/high temporal resolution satellite data (e.g. MODIS, SEVIRI) are complemented, compared and fine-tuned with detailed maps of volcanic deposits with the aim of constraining the conversion from satellite-derived radiant heat flux to TADR. Maps of volcanic deposits include the time-varying evolution of lava flow emplacement derived from multispectral satellite data (e.g. EO-ALI, Landsat, Sentinel-2, ASTER), as well as the flow thickness variations, retrieved from the topographic monitoring by using stereo or tri-stereo optical data (e.g. Pléiades, PlanetScope, ASTER). Finally, satellite-derived parameters are used as input and validation tags for the numerical modelling of lava flow scenarios. Here we show how our strategy was successfully applied to several remote volcanoes around the world.