**7. Conclusions**

*Updates in Volcanology – Transdisciplinary Nature of Volcano Science*

202 out of 319 Holocene volcanoes did not have seismic, deformation or gas monitoring in 2013 [7]. A complex-valued CNN was proposed to extract areas with land shapes similar to given samples in interferometric synthetic aperture radar (InSAR), a technique widely applied in volcano monitoring. An application was presented grouping similar small volcanoes in Japan [149]. InSAR measurements have great potential for volcano monitoring, especially where images are freely available. ML methods can be used for the initial processing of single satellite data. Processing of potential unrest areas can then fully exploit integrated multi-disciplinary, multisatellite datasets [7]. The Copernicus Programme of the European Space Agency (ESA) and the European Union (EU) has recently contributed by producing the Sentinel-2 multispectral satellites, able to provide high resolution satellite data for disaster monitoring, as well as complementing previous satellite images like Landsat. The free access policy also promotes an increasing use of Sentinel-2 data, which is often processed by ML techniques such as SVM and RF [150]. A transfer learning strategy was applied to ground deformation in Sentinel-1 data [151] and a range of pretrained networks was tested, finding that AlexNet [152] is best suited to this task. The positive results were checked by a researcher and fed back for model

The global volcano monitoring platform MOUNTS (Monitoring Unrest from Space) uses multisensor satellite-based imagery (Sentinel-1 Synthetic Aperture Radar SAR, Sentinel-2 Short-Wave InfraRed SWIR, Sentinel-5P TROPOMI), ground-based seismic data (GEOFON and USGS global earthquake catalogs), and CNN to provide support for volcanic risk assessment. Results are visualized on an open-access website. The efficiency of the system was tested on several eruptions (Erta Ale 2017, Fuego 2018, Kilauea 2018, Anak Krakatau 2018, Ambrym 2018, and

Debris flow events are one of the most widespread and dangerous natural processes not only on volcanoes but more in general in mountainous environments. A methodology was recently proposed [154] that combines the results of deterministic and heuristic/probabilistic models for susceptibility assessment. RF models are extensively used to represent the heuristic/probabilistic component of the modeling. The case study presented is given by the Changbai Shan volcano,

Mapping lava flows from satellite is another important remote sensing application. RF was applied to 20 individual flows and 8 groups of flows of similar age using a Landsat 8 image and a DEM of Nyamuragira (Congo) with 30 m resolution. Despite spectral similarity, lava flows of contrasting age can be well discriminated

The hazard related to landslides at volcanoes is also significant. DNN models were proposed for landslide susceptibility assessment in Viet Nam, showing considerable better performance with respect to other ML methods such as MLP, SVM, DT and RF [156]. The use of DNN approach could be therefore an interesting approach

Muon imaging has been successfully used by geophysicists to investigate the internal structure of volcanoes, for example at Etna (Italy) [157]. Muon imaging is essentially an inverse problem and it can profit from the application of ML tech-

map volcanoes also on other planets. A ML paradigm was designed for the identification of volcanoes on Venus [159]. Other studies have used topographic data, such as DEM and associated derivatives obtained from orbital images, to detect and

classify manually labeled Martian landforms including volcanoes [160].

Combinations of supervised and unsupervised ML techniques have been used to

**118**

updating.

China [154].

Piton de la Fournaise 2018–2019) [153].

and mapped by means of image classification [155].

niques, such as ANN and CA [158].

for the landslide susceptibility mapping of active volcanoes.

ML techniques will have an increasing impact on how we study and model volcanoes in all their aspects, how we monitor them and how we evaluate their hazards, both in the short and in the long term. The increasing number of monitoring equipment installed on volcanoes on one side provides more and more data, on the other often causes their real time processing unfeasible especially when most needed i.e., during unrest and eruptions. Here ML will show its best usefulness, as it can provide the perfect tools to sift through big data to identify subtle patterns that could indicate unrest, hopefully well before eruptions. One important issue is the one of generalization. We must go towards the construction of ML models that can be applied on different volcanoes, for instance when previous data is not available for training specific models. The concepts of transfer learning can be important here.

The routine use of ML tools at the different volcano observatories should be promoted by providing easy installation procedures and easy integration into existing monitoring systems. Open source software should be always chosen whenever possible. On the other hand, observatories should provide good open training data to ML developers, researchers and data scientists in order to improve the models in a virtuous circle. An easy availability of open access data, both from the ground and from satellites should be exploited for building reliable training sets in the different fields of volcanology. This will allow "scientific competition" between research groups using different ML approaches and make a direct comparison of results easier, like it is common in other disciplines where "standard" training datasets are available for download to everybody.
