case study of a volcano

• Current Eruptions
• Smithsonian / USGS Weekly Volcanic Activity Report
• Bulletin of the Global Volcanism Network
• Weekly Report 20th Anniversary
• Holocene Volcano List
• Pleistocene Volcano List
• Country Volcano Lists
• Volcano Search
• Eruption Search
• Webservices
• Database Information
• Image Collections
• Video Collections
• Theme Collections
• Keyword Collections
• Student Art Gallery 2024
• St. Helens 40th Anniversary
• Frequent Questions
• Information Sources
• Google Earth Placemarks
• This Dynamic Planet
• Eruptions, Earthquakes & Emissions Application
• Volcano Numbers
• Volcano Naming
• How to Cite
• Terms of Use

Report on La Palma (Spain) — October 2021

Bulletin of the Global Volcanism Network, vol. 46, no. 10 (October 2021) Managing Editor: Edward Venzke. Edited by A. Elizabeth Crafford. La Palma (Spain) First eruption since 1971 starts on 19 September; lava fountains, ash plumes, and lava flows

Please cite this report as: Global Volcanism Program, 2021. Report on La Palma (Spain) (Crafford, A.E., and Venzke, E., eds.). Bulletin of the Global Volcanism Network , 46:10. Smithsonian Institution. https://doi.org/10.5479/si.GVP.BGVN202110-383010

28.57°N, 17.83°W; summit elev. 2426 m

All times are local (unless otherwise noted).

Multiple eruptions have occurred during the last 7,000 years at the Cumbre Vieja volcanic center on La Palma, the NW-most of the Canary Islands. The eruptions have created cinder cones and craters, and produced fissure-fed lava flows that reached the sea a number of times. Eruptions recorded since the 15th century have produced mild explosive activity and lava flows that damaged populated areas, most recently at the southern tip of the island in 1971. During the three-week eruption in October-November 1971, eruptive activity created a new cone, Teneguia, that had as many as six active vents (CSLP 90-71), and blocky lava flows that reached the sea on the SW flank.

A new eruption began at La Palma on 19 September 2021 in an area on the SW flank of the island about 20 km NW of the 1971 eruption, after a multi-year period of elevated seismicity. Two fissures opened and multiple vents produced lava fountains, ash plumes, and flows that traveled over 5 km W to the sea, destroying hundreds of properties in their path (figure 2). Activity through the end of September is covered in this report with information provided by Spain’s Instituto Geographico Nacional (IGN), the Instituto Volcanologico de Canarias (INVOLCAN), the Steering Committee of the Special Plan for Civil Protection and Attention to Emergencies due to Volcanic Risk (PEVOLCA), maps from Copernicus EMS, satellite data, and news and social media reports.

Precursor seismicity. In early July 2017 IGN enhanced their Volcanic Surveillance Network at La Palma to include four GPS antennas, five seismic stations, and four hydrochemical groundwater control points. A seismic swarm of 68 events located on the southern third of the island was recorded during 7-9 October 2017. It was the first of a series of seismic swarms recorded during 2017-2021 (table 1) located in the same general area. This first swarm was followed by a similar set of events a few days later during 13-14 October. The magnitudes of the events during October 2017 (given as MbLg, or the magnitude from the amplitude of the Lg phase, similar to the local Richter magnitude) ranged from less than 1.5 to 2.7, and they occurred over a depth range of 12-35 km. The next seismic swarm of similar characteristics occurred during February 2018, followed by a smaller swarm of seven microseismic events recorded in the same area one year later, on 12 February 2019.

Table 1. Precursor seismicity episodes at La Palma between October 2017 and late June 2021 were all located in the southern third of the island. Magnitude is reported by IGN as MbLg, or the magnitude from the amplitude of the Lg phase, similar to the local Richter magnitude. Data courtesy of IGN Noticias.

By the time the next seismic swarm began in July 2020, IGN had a network of 13 seismic stations installed around the island. There were 160 located events that occurred during 24 July-2 August 2020 with magnitudes of 1.2-2.5 and depths of 16-39 km. Reprocessing of the previous data indicated a distribution of seismicity for the three series (October 2017, February 2018, and July 2020) in a wide strip in an east-west direction, although the October 2017 series occurred at a shallower depth and with the epicenters more concentrated. IGN noted similarities between the February 2018 and July-August 2020 events in terms of location and magnitude (figure 3). Another very similar swarm of 602 detected events was recorded during 23-26 December 2020, with most events located on the western slope of Cumbre Vieja. Two swarms on 21 January and 25 June 2021 had fewer events but similar depths and magnitudes to the earlier events.

Renewed seismicity began on 11 September 2021. The number, frequency, and magnitude of the events all increased over the next several days, while the depth of the events grew shallower. On 13 September a multi-agency scientific committee raised the Alert Level to Yellow (the second lowest level on a four-color scale) for the municipalities of El Paso, Los Llanos de Aridane, Mazo, and Fuencaliente de la Palma. IGN noted a migration of the seismicity toward the W side of the island on 14 September (figure 4). The accumulated surface deformation between 12 and 14 September measured 1.5 cm from the island GNSS network. Seismic activity on 15 September continued to migrate slightly NW at depths of around 7-9 km; in addition, 20 shallow earthquakes of 1-3 km depth were recorded. The accumulated deformation had reached 6 cm by 15 September. As of 0930 on 16 September 50 shallow earthquakes between 1-5 km depth had been located and the maximum vertical deformation was around 10 cm in the area of the seismicity. During 16-18 September seismic activity decreased, but a 3.2 magnitude earthquake located at 100 m depth was felt by the local population. Intense surface seismicity (between 0-6 km) increased in the early hours of 19 September and numerous earthquakes were felt by the local population (figure 4). The maximum accumulated deformation increased to 15 cm in the area close to the seismicity by 1100 on 19 September, and the eruption began about five hours later.

Eruption begins 19 September 2021. A fissure eruption began at 1510 local time (1410 UTC) on 19 September after the intense seismic and deformation activity that began on 11 September. Observers near the eruption site in the area of Cabeza de Vaca, in the municipality of El Paso, witnessed a large explosion with ejecta that produced a gas-and-ash plume. Strombolian activity was accompanied by phreatomagmatic pulses along two 100-m-long N-S fissures about 200 m apart. INVOLCAN scientists observed seven vents along the fissures during the initial stage of the eruption (figure 5). Multiple tall lava fountains fed flows downslope to the W, igniting fires. The PEVOLCA steering committee briefly raised the Alert Level to Orange, and then to Red by 1700 for high-risk municipalities directly affected by the eruption. About 5,500 people evacuated with no injuries reported, and authorities recommended that residents stay at least 2 km from the vents. INVOLCAN scientists determined an average flow rate of 700 m/hour and lava temperatures of around 1,075°C at the start of the eruption (figure 6).

The Toulouse VAAC issued the first ash advisory for the La Palma eruption about 90 minutes after it began. They reported ongoing lava fountains and an ash plume to about 1 km altitude. The plume drifted SW at less than 1.5 km altitude, while SO 2 emissions were reported drifting ESE at 3 km altitude. Later that day, they noted continuing intense lava fountains and ashfall in the vicinity of the volcano. The next day ash emissions drifted S at 2.4 km altitude. Sulfur dioxide emissions were measured by satellite instruments beginning on 19 September; they increased dramatically and drifted hundreds of kilometers E and SE toward the NE coast of Africa over the next few days (figure 7). Ongoing ash emissions rose to 4.6 km altitude later on 20 September. The first Sentinel-2 satellite images of the eruption appeared on 20 September showing a strong point source thermal anomaly partly covered by meteoric clouds (figure 8).

The first map of the new flow on 20 September produced by IGN in partnership with Copernicus Emergency Management Service (EMS) showed that the main channel of the lava flow had traveled more than 3 km W. The flows had covered about 1 km 2 and destroyed an estimated 166 buildings (figure 9). A report of the PEVOLCA Scientific Committee indicated that activity on 20 and 21 September was concentrated at four main vents that produced parallel flows with an average flow rate of 200 m/hour; the maximum flow thickness was 10-12 m (figure 10). Strong lava fountaining continued both days and ash fell in the vicinity of the vents. By 0814 on 21 September an updated Copernicus EMS map showed that 350 homes had been covered by lava and the flow field had expanded to 1.54 km 2 . A few hundred more residents evacuated as lava advanced towards Tacande; bringing the number of evacuees to about 5,700. One lava flow branch was advancing slowly S at a rate of 2 m/hour. An ash cloud was observed later that day on the W flank of the volcano slowly drifting SW at 2.4 km altitude. Sulfur dioxide emissions were present over the SE part of the island and were visible at Gomera Island, 80 km SE. Late in the day, ash was observed in satellite imagery about 50 km W of the volcano, while intense lava fountaining continued at the source vent (figure 11).

Activity during 22-25 September 2021. Ash emissions during 22 and 23 September drifted SW and S from 0-3 km altitude, and NE and E from 3-5 km altitude (figure 12); ashfall up to 3 cm thick was reported downwind. An SO 2 plume was also noted drifting NE in satellite imagery. PEVOLCA reported on 23 September that two relatively slow-moving lava flows continued to advance downslope from the vent (figure 13). The northernmost flow was moving at 1 m/hour and was 12 m high and 500 m wide in some places. The southern flow, which surrounded Montaña Rajada, was moving at 4-5 m/hour and about 10 m high. The overall flow was 3.8 km long and 2.1 km from the coast (figure 14). By late on 23 September reports indicated 420 structures had been destroyed and the flow covered just under 2 km 2 .

Lava fountains rose hundreds of meters above the summit crater of the new cone early on 24 September 2021 (figure 15). IGN reported an increase in explosive activity on 24 September that was accompanied by a sharp increase in tremor amplitude. This was followed a short while later by the opening of two new vents on the NW flank of the cone; the fast-moving flows merged into one and produced a new flow over top of the earlier flows. Part of the upper section of the S flank of the cone collapsed on 24 September and briefly caused flow speeds to increase to 250-300 m/hour overnight before slowing to an average speed of 40 m/hour. Due to the fast-moving flow, an evacuation order was issued in the early afternoon for Tajuya, Tacande de Abajo, and part of Tacande de Arriba, affecting 300-400 people. Three airlines also suspended flights to La Palma. The Toulouse VAAC reported ash plumes throughout the day. Ash plumes drifted SW below 3 km altitude and E and SE at 3-5.2 km altitude and resulted in significant ashfall in numerous locations by the next morning (figure 16). Pilots also reported ash near Tenerife and over La Gomera.

By 25 September there were three active vents in the crater and one on the flank of the cone (figure 17), and two active lava flows. On 25 and 26 September dense ash emissions (figure 18) closed the airport and produced ashfall not only in the municipalities near the eruption, but also on the eastern slope of the island; it was reported in Villa de Mazo, Breña Alta and Breña Baja, and Santa Cruz de La Palma or Puntallana. Plumes were drifting SW at altitudes below 1.5 km and NE between 1.5 and 3.9 km altitude over a large area. Mapping by Copernicus EMS indicated that the ashfall covered an area of 13 km 2 (figure 19).

Activity during 26-28 September 2021. During the evening of 26 September jets of lava up to 1 km high were visible from La Laguna and some explosions were strong enough to be felt within 5 km of the vent (figure 20). The main, more northerly lava flow overtook the center of Todoque, in the municipality of Los llanos de Aridane, which had been evacuated several days earlier. It crossed the highway (LP-213) in the center of town and continued 150 m W. It was initially moving at about 100 m/hour, was 4-6 m high, and the front was about 600 m wide, but it slowed significantly after crossing through Todoque, and the height grew to 15 m; it was located about 1,600 m from the coast. The more southerly flow continued moving at about 30 m/hour and was about 2.5 km long.

The PEVOLCA Scientific Committee determined that the volume of erupted material from the beginning of the eruption on 19 September until 27 September was about 46.3 m 3 . By early on 27 September the front of the flow was close to the W side of Todoque Mountain (figure 21), and reports indicated that 589 buildings and 21 km of roads had been destroyed by the 2.5 km 2 of lava. A seismic swarm on the morning of 27 September was located at about 10 km depth in the same area of the previous seismicity below the vent. In addition, pulses of tremor coincided with pulses of ash emissions. A new flow appeared on the N flank of the cone during the afternoon and partly covered previous flows through the center of Todoque, reaching about 2 km from the coast (figure 22). Ash emissions were more intermittent on 27 and 28 September, drifting SW to 1.5 km altitude and NE to 4.3 km altitude in sporadic pulses associated with lava fountains.

The new flow moved through the upper outskirts of Todoque and had reached the road to El Pampillo on the border of the municipalities of Los Llanos and Tazacorte, about 1 km from the coast, early on 28 September (figure 23). A plume with moderate to high ash concentration rose to 5.2 km altitude and extended up to 25 km W. The altitude of the plume increased to 6.1 km drifting E later in the day. A significant SO 2 cloud was clearly identifiable in satellite imagery in a 75 km radius around the island. In addition, satellite instruments measured very large plumes of SO 2 drifting hundreds of kilometers E, S, and N over the next several days (figure 24).

Activity during 28-30 September 2021. Effusive activity continued with a sharp decrease in tremor during the day on 28 September. By evening, sustained fountaining was continuing at the N flank vent, while pulsating jets from three vents within the main crater produced strong effusion into both lava flows. The volume of the cone that had formed at the vent was estimated by PEVOLCA to be 10 million m 3 . Around 2300 local time on 28 September the main lava flow passed on the S side of Todoque Mountain and entered the sea in the area of Playa de Los Guirres in Tazacorte. A continuous cascading flow of lava fell over the cliff (figure 25) and began to form a lava delta. By dawn on 29 September the delta was growing out from the cliff, producing dense steam explosions where the lava entered the sea (figure 26).

By nightfall on 29 September vigorous Strombolian activity was continuing at the pyroclastic cone, and the main lava flow was active all the way to the sea, with a growing delta into the ocean. Ash emissions continued on 29 and 30 September, rising in pulses to 5.2 km altitude and drifting SE, changing to S, SW, and finally NW. Sentinel-2 satellite imagery comparing 25 and 30 September showed the growth of the lava flow during that interval (figure 27). Strombolian and flow activity continued at the fissure vent on 30 September with new surges of activity sending fresh pulses of lava over existing flows (figure 28). The ocean delta continued to grow and reached a thickness of 24 m by the end of 30 September. Mapping of the flow indicated that 870 buildings had been destroyed and the flow covered 3.5 km 2 by midday on 30 September (figure 29).

Late on 30 September 2021 two new vents emerged about 600 m NW of the base of the main cone. They created a new flow about 450 m away from, and parallel to, the main flow that crossed a local highway by the next morning and continued moving W (figure 30). Multiple vents also remained active within and on the flank of the main cone. As of 1 October, the front of the delta was 475 m out from the coastline and 30 m deep. IGN concluded that the volume of material erupted through the end of September was approximately 80 million m 3 .

Geological Summary. The 47-km-long wedge-shaped island of La Palma, the NW-most of the Canary Islands, is composed of two large volcanic centers. The older northern one is cut by the steep-walled Caldera Taburiente, one of several massive collapse scarps produced by edifice failure to the SW. On the south, the younger Cumbre Vieja volcano is one of the most active in the Canaries. The elongated volcano dates back to about 125,000 years ago and is oriented N-S. Eruptions during the past 7,000 years have formed abundant cinder cones and craters along the axis, producing fissure-fed lava flows that descend steeply to the sea. Eruptions recorded since the 15th century have produced mild explosive activity and lava flows that damaged populated areas. The southern tip of the island is mantled by a broad lava field emplaced during the 1677-1678 eruption. Lava flows also reached the sea in 1585, 1646, 1712, 1949, 1971, and 2021.

Information Contacts: Instituto Geographico Nacional (IGN) , C/ General Ibáñez de Íbero 3, 28003 Madrid – España, (URL: https://www.ign.es/web/ign/portal, https://www.ign.es/web/resources/volcanologia/html/CA_noticias.html); Instituto Volcanologico de Canarias (INVOLCAN) (URL: https://www.involcan.org/, https://www.facebook.com/INVOLCAN, Twitter: INVOLCAN, @involcan); Steering Committee of the Special Plan for Civil Protection and Attention to Emergencies due to Volcanic Risk (PEVOLCA) , (URL: https://www3.gobiernodecanarias.org/noticias/los-planes-de-evacuacion-del-pevolca-evitan-danos-personales-en-la-erupcion-volcanica-de-la-palma/); NASA Global Sulfur Dioxide Monitoring Page , Atmospheric Chemistry and Dynamics Laboratory, NASA Goddard Space Flight Center (NASA/GSFC), 8800 Greenbelt Road, Goddard, Maryland, USA (URL: https://so2.gsfc.nasa.gov/); Copernicus EMS (URL: https://emergency.copernicus.eu/, https://twitter.com/CopernicusEMS ); Sentinel Hub Playground (URL: https://www.sentinel-hub.com/explore/sentinel-playground); Cabildo La Palma (URL: https://www.cabildodelapalma.es/es/algunas-de-las-imagenes-de-la-erupcion-volcanica-en-la-palma); El Periodico de Cataluny, S.L.U. (URL: https://www.elperiodico.com/es/fotos/sociedad/erupcion-palma-imagenes-12093812/12103264). Corporación de Radio y Televisión Española (RTVE) (URL: https://rtve.es, https://img2.rtve.es/imagenes/casas-todoque-alcanzadas-lava-este-miercoles-22-septiembre/1632308929494.jpg); Tom Pfeiffer , Volcano Discovery (URL: http://www.volcanodiscovery.com/); Volcanes de Canarias (URL:https://twitter.com/VolcansCanarias/status/1441711738983002114); Agence France-Presse (AFP) (URL: http://www.afp.com/ ); Bristol Flight Lab , University of Bristol, England (URL: www.https://flight-lab.bristol.ac.uk, https://twitter.com/UOBFlightLab).

• Share full article

Advertisement

Supported by

When Kilauea Erupted, a New Volcanic Playbook Was Written

Scientists learned lessons from the 2018 outburst on the island of Hawaii that are changing how responders prepare for eruptions in other places.

By Robin George Andrews

Back in the summer of 2018, Wendy Stovall stood and stared into the heart of an inferno.

Hawaii’s Kilauea volcano had been continuously erupting in one form or another since 1983. But from May to August, the volcano produced its magnum opus , unleashing 320,000 Olympic-size swimming pools’ worth of molten rock from its eastern flank.

Dr. Stovall, the deputy scientist-in-charge at the U.S. Geological Survey’s Yellowstone Volcano Observatory , recalls moments of being awe-struck by the eruption’s incandescence: lava fountains roaring like jet engines, painting the inky blue sky in crimson hues. But these briefly exhilarating moments were overwhelmed by sadness. The people of Hawaii would suffer hundreds of millions of dollars in economic damage. The lava bulldozed around 700 homes. Thousands of lives were upended. Even the headquarters of the Hawaiian Volcano Observatory itself, sitting atop the volcano , was torn apart by earthquakes early in the crisis.

Like many volcanologists who were there during the eruption, Dr. Stovall is still processing the trauma she witnessed. Sadness is not quite the right word to describe what she feels, she said: “Maybe it’s an emotion that I don’t even have a word for.”

But not only trauma has resulted from the crisis: It has also produced something of a sea change in the way scientists and their emergency services partners are able to respond to volcanic emergencies.

During Kilauea’s devastating outburst, responders found novel ways to deploy drones and used social media to help those in the lava’s path. They also achieved more ineffable insights into how to keep cool in the face of hot lava. And this pandemonium of pedagogical experiences will prove valuable in times to come. The United States is home to 161 active or potentially active volcanoes — approximately 10 percent of the world’s total. When — not if — a Kilauean-esque outburst or something more explosive takes place near an American city, scientists and emergency responders will be better prepared than ever to confront and counter that volcanic conflagration.

A Patchwork of Fire

We are having trouble retrieving the article content.

Please enable JavaScript in your browser settings.

Thank you for your patience while we verify access. If you are in Reader mode please exit and  log into  your Times account, or  subscribe  for all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?  Log in .

Want all of The Times?  Subscribe .

Volcano World

• Central Andes (Under Construction)
• Hawaiian Volcanism
• Mount Saint Helens
• Volcanoes of Japan
• Extraterrestrial Volcanoes
• Teaching Resources
• Mount St Helens Living Lab
• Plate Tectonics
• Rocks and Minerals
• Prehistoric Earth
• Types of Volcanoes
• Eruption Styles
• Submarine Volcanoes
• Columnar Jointing
• Hydrovolcanism
• Volcanogenic Tsunamis
• Volcanic Minerals
• Volcanic Gases
• Volcanic Lightning
• Volcanic Sounds
• Volcanic Hazards
• GVP Current Eruption Reports
• US Volcano Activity Map
• Volcano Factoids
• Building Volcano Models
• Interviews with Volcanologists
• Working on Volcanoes
• Art Gallery
• Volcano Games
• Mount St. Helens by Helicopter
• Virtual Altiplano
• Virtual Altiplano Bolivia
• Around Santorini
• Climbing Mt Asama
• Driving Kilauea Rim Drive
• Exploring the Canary Islands
• Hiking Mount Ijen
• To the Top of Beerenberg
• Volcano Folklore
• Volcanic Parks
• Volcano Gallery

You are here

Case studies.

Left:  Mt. Pinatubo eruption plume, July 1991, from Clark Air Base control tower.  Photo by J.N. Marso, U.S. Geological Survey.

The effects of several historic eruptions have been observed and the impacts of larger, prehistoric eruptions can be estimated.

Estimates of the fraction of sunlight transmitted through stratigraphic aerosols after major eruptions. Roza refers to a flood basalt eruption in the northwestern United States. Graph from Rampino and others (1988).

The pages in this section explore the following case studies for their impact on global climate

Impact of some major historic eruptions.

Data from Rampino and Self, 1984.

• Mt. Pinatubo, Philippines - 1991
• El Chichon, Mexico - 1982
• Krakatau - 1883
• Tambora, Indonesia, 1815
• Laki, Iceland - 1783
• Toba, Indonesia, 75,000 years ago

Contact Info

VW is a higher education, k-12, and public outreach project of the Oregon Space Grant Consortium administered through the Department of Geosciences at Oregon State University.

Disclaimers and Policies | Login

• 0 Shopping Cart

Eyjafjallajokull Case Study

What is Eyjafjallajokull?

Eyjafjallajokull is a volcano located in Iceland. The name is a description of the volcano with Eyja meaning island; fjalla meaning mountain; and jokull meaning glacier. You can find out how to pronounce Eyjafjallajokull on the BBC website .

Eyjafjallajökull consists of a volcano completely covered by an ice cap. The ice cap covers an area of about 100 square kilometres (39 sq mi), feeding many outlet glaciers.

What type of volcano is Eyjafjallajokull?

The mountain itself, a composite (stratovolcano) volcano, stands 1,651 metres (5,417 ft) at its highest point and has a crater 3–4 kilometres (1.9–2.5 mi) in diameter, open to the north.

When did Eyjafjallajokull erupt?

Eyjafjallajokull erupted between March and May 2010.

Why did Eyjafjallajokull erupt?

Iceland lies on the Mid-Atlantic Ridge, a constructive plate margin separating the North American and Eurasian plates. The two plates move apart due to ridge push along the Mid-Atlantic Ridge. As the plates move apart, magma fills the magma chamber below Eyjafjallajokull—several magma chambers combined to produce a significant volume of magma below the volcano. Eyjafjallajokull is located below a glacier.

The Eyjafjallajökull volcano erupted in 920, 1612 and again from 1821 to 1823 when it caused a glacial lake outburst flood (or jökulhlaup). It erupted three times in 2010—on 20 March, April–May, and June. The March event forced a brief evacuation of around 500 local people. Still, the 14 April eruption was ten to twenty times more powerful and caused substantial disruption to air traffic across Europe. It caused the cancellation of thousands of flights across Europe and to Iceland.

How big was the eruption of Eyjafjallajokull?

The eruption was only three on the volcanic explosivity index (VEI). Around 15 eruptions on this scale usually happen each year in Iceland. However, in this case, a combination of a settled weather pattern with winds blowing towards Europe, very fine ash and a persistent eruption lasting 39 days magnified the impact of a relatively ordinary event. The eruptions in March were mainly lava eruptions. On 14 April, a new phase began, which was much more explosive. Violent eruptions belched huge quantities of ash into the atmosphere.

The eruption of Eyjafjallajokull

What were the impacts of the eruption? (social / economic / environmental – primary and secondary effects)

Primary effects : As a result of the eruption, day turned to night, with the ash blocking the sun. Rescuers wore face masks to prevent them from choking on ash clouds.

Homes and roads were damaged, services were disrupted, crops were destroyed by ash, and roads were washed away. The ash cloud brought European airspace to a standstill during the latter half of April 2010 and cost billions of euros in delays. During the eruption, a no-fly zone was imposed across much of Europe, meaning airlines lost around £130m per day. The price of shares in major airlines dropped between 2.5 and 3.3% during the eruption. However, it should be noted that imports and exports are being impacted across European countries on the trade front, so the net trade position was not affected markedly overall.

Secondary effects : Sporting events were cancelled or affected due to cancelled flights. Fresh food imports stopped, and industries were affected by a lack of imported raw materials. Local water supplies were contaminated with fluoride. Flooding was caused as the glacier melted.

International Effects: The impact was felt as far afield as Kenya, where farmers have laid off 5000 workers after flowers and vegetables were left rotting at airports. Kenya’s flower council says the country lost $1.3m a day in lost shipments to Europe. Kenya exports typically up to 500 tonnes of flowers daily – 97% of which is delivered to Europe. Horticulture earned Kenya 71 billion shillings (£594m) in 2009 and is the country’s top foreign exchange earner. You can read more about this on the Guardian website .

What opportunities did the eruption of Eyjafjallajokull bring?

Despite the problems caused by the eruption of Eyjafjallajokull, the eruption brought several benefits. According to the Environmental Transport Association, the  grounding of European flights prevented some 2.8 million tonnes of carbon dioxide into the atmosphere (according to the Environmental Transport Association).

As passengers looked for other ways to travel than flying, many different transport companies benefited. There was a considerable increase in passenger numbers on Eurostar. It saw a rise of nearly a third, with 50,000 extra passengers travelling on their trains.

Ash from the Eyjafjallajökull volcano deposited dissolved iron into the North Atlantic, triggering a plankton bloom, driving an increase in biological productivity.

Following the negative publicity of the eruption, the Icelandic government launched a campaign to promote tourism . Inspired by Iceland was established with the strategic intent of depicting the country’s beauty, the friendliness of its people and the fact that it was very much open for business. As a result, tourist numbers increased significantly following the campaign, as shown in the graph below.

Foreign visitor arrivals to Iceland

What was done to reduce the impact of the eruption of Eyjafjallajokull?

In the short term, the area around the volcano was evacuated.

European Red Cross Societies mobilised volunteers, staff and other resources to help people affected directly or indirectly by the eruption of the Eyjafjallajökull glacier volcano. The European Red Cross provided food for the farming population living in the vicinity of the glacier, as well as counselling and psychosocial support, in particular for traumatised children. Some 700 people were evacuated from the disaster zone three times in the past month. In one instance, people had to flee their homes in the middle of the night to escape from flash floods.

The European Union has developed an integrated structure for air traffic management. As a result, nine Functional Airspace Blocks (FABs) will replace the existing 27 areas. This means following a volcanic eruption in the future, areas of air space may be closed, reducing the risk of closing all European air space.

Premium Resources

Please support internet geography.

If you've found the resources on this page useful please consider making a secure donation via PayPal to support the development of the site. The site is self-funded and your support is really appreciated.

Related Topics

Use the images below to explore related GeoTopics.

Previous Topic Page

Topic home, sunda strait tsunami indonesia case study 2018, iconbox title.

Click to add your own text here

Share this:

• Click to share on Twitter (Opens in new window)
• Click to share on Facebook (Opens in new window)
• Click to share on Pinterest (Opens in new window)
• Click to email a link to a friend (Opens in new window)
• Click to share on WhatsApp (Opens in new window)
• Click to print (Opens in new window)

If you've found the resources on this site useful please consider making a secure donation via PayPal to support the development of the site. The site is self-funded and your support is really appreciated.

Search Internet Geography

Latest Blog Entries

Pin It on Pinterest

• Click to share
• Print Friendly

Case Study: Volcanoes ( SL IB Geography )

Revision note.

Geography Lead

Case Study: Mount Merapi

Mount merapi earthquake facts.

• Name – Mount Merapi
• Location – Java, Indonesia
• Date – 25th October–30th November 2010
• Magnitude – VEI 4
• Plate boundary – Destructive plate boundary where the Indo-Australian plate is subducting below the Eurasian plate
• Type of volcano – Stratovolcano or composite

Location of Mount Merapi

Impacts of the 2010 Eruption of Mount Merapi, Indonesia

Factors affecting vulnerability

• The number of deaths, injuries and displacement of population was high during and after the eruption
• People were vulnerable to the impacts of the hazard
• People refused to leave their homes, which made them more vulnerable to the impacts of the eruption
• Caring responsibilities for elderly parents
• Responsibilities for livestock
• Long-term residency and a subsequent unwillingness to leave
• Cultural beliefs
• Population density in the area has increased 
• Local people don't always believe that scientific monitoring is accurate, relying instead on traditional warning signs
• Communication regarding the dangers of the eruption was slow and ineffective

Case Study: Cumbre Vieja

La palma, spain.

• Part of the Canary Islands, La Palma is located in the Atlantic Ocean off the coast of North Africa
• The Canary Islands are an [popover id="RAr2r~3MbVY7biGB" label=''autonomous region"] of Spain
• There are 33 volcanoes across the Canary Islands, 10 of which are in La Palma

Cumbre Viejo earthquake facts

• Name – Cumbre Viejo
• Location – La Palma, Spain
• Date – 19th September–December 2021
• Magnitude – VEI 2 or 3
• Plate boundary – Magma plume (hotspot)
• Type of volcano – Cinder cone (basaltic lava)

Location of Cumbre Vieja Volcano

Primary impacts

• Almost 1500 houses were destroyed by the lava flow
• Over 1500 other buildings such as churches, shops and schools were destroyed
• The lava flow cut across the coastal highway and covered 1000 hectares
• The water supply was cut off for almost 3000 people
• 400 hectares of banana farms were destroyed
• Almost 1300 hectares of land were affected
• There was one death

Secondary impacts

• Air traffic was suspended on a number of occasions due to ash in the atmosphere
• Over 1000 people were evacuated after the eruption began on the 19th September
• A further 5600 people were evacuated over the next few weeks
• About 20,000 people were exposed to the eruption and its effects
• Evacuation plans
• Suspension of air traffic
• La Palma also has a Local Hazard Mitigation Plan , which aims to reduce the impacts of any hazard event
• People are encouraged to have an emergency kit ready in case of eruptions
• Insurance cover means that recovery from a hazard event is more rapid
• La Palma has well-trained and equipped emergency services

You've read 0 of your 10 free revision notes

Get unlimited access.

to absolutely everything:

• Downloadable PDFs
• Unlimited Revision Notes
• Topic Questions
• Past Papers
• Model Answers
• Videos (Maths and Science)

Join the 100,000 + Students that ❤️ Save My Exams

the (exam) results speak for themselves:

Did this page help you?

Author: Bridgette

After graduating with a degree in Geography, Bridgette completed a PGCE over 25 years ago. She later gained an MA Learning, Technology and Education from the University of Nottingham focussing on online learning. At a time when the study of geography has never been more important, Bridgette is passionate about creating content which supports students in achieving their potential in geography and builds their confidence.

La Palma Eruption 2021

Date: Sept. 19, 2021 Type:    Volcanoes Region :  Africa , Canary Islands Info & Resources: 

• View maps & data products for the La Palma eruption on the NASA Disasters Mapping Portal
• NASA Disasters program resources for volcanoes
• Latest updates from the Instituto Geologico y Minero de Espana (IGME)
• Latest updates from the Smithsonian Global Volcanism Program
• Latest updates from the Instituto Volcanológico de Canarias (INVOLCAN)
• Educational story map of La Palma data products & visualizations, developed by Esri

UPDATE Oct. 13, 2021

View fullscreen on the NASA Disasters Mapping Portal

Researchers working with the NASA ROSES A.37 project “ Day-Night Monitoring of Volcanic SO2 and Ash for Aviation Avoidance at Northern Polar Latitudes ” developed this animation of sulfur dioxide (SO2) clouds from the La Palma eruption using satellite data from NASA / NOAA Suomi-NPP and NOAA-20 Ozone Mapping and Profiler Suite (OMPS) spectrometers. Both satellites fly similar near-polar orbits, but are about 50 minutes apart. NOAA-20 OMPS measures with higher ground resolution. Using two satellites allows researchers to make more frequent, precise observations to identify hazardous densities of volcanic gases and aerosols.  

The above animation shows SO2 column density in Dobson Units (1 DU = 2.69 x 1016 SO2 molecules /cm2) from Sept. 19 – 30, 2021. S02 is used to indicate the presence of volcanic gases and also as a proxy for volcanic aerosols (sulfuric acid or vog and ash), which can negatively affect air quality for people living in the region, as well as potentially damage aircraft flying through the volcanic clouds. Credits: NASA  

Update Oct. 4, 2021

On Sept. 19, 2021, the Cumbre Vieja volcano on the island of La Palma in the Canary Islands started erupting after remaining dormant for 50 years. Since the initial eruption, the volcano has seen several Strombolian explosions , significant emissions of ash and gas, and multiple vents spewing molten lava down the mountain and into surrounding regions. According to the latest media reports over 800 buildings have been destroyed and about 6,000 people evacuated from the area.

The NASA Earth Applied Sciences Disasters program area has activated efforts to monitor the eruption and provide Earth-observing data and analysis in support of risk reduction and recovery for the eruption. The program is in contact with colleagues from the Instituto Geologico y Minero de Espana ( IGME ) and the Institut de Physique du Globe de Paris ( IPGP ) to share knowledge and data for situational awareness. 

These efforts are being supported by the NASA ROSES A.37 research projects “ Day-Night Monitoring of Volcanic SO2 and Ash for Aviation Avoidance at Northern Polar Latitudes ” and “ Global Rapid Damage Mapping System with Spaceborne SAR Data .”

Related Impact

Connect with the Disasters Program

With help from NASA’s Earth-observing satellites, our community is making a difference on our home planet. Find out how by staying up-to-date on their latest projects and discoveries.

Stay Connected

An ash plume rises from the active volcano on Whakaari/White Island in New Zealand on December 9, 2019.

Why the New Zealand volcano eruption caught the world by surprise

The explosive event was “the actual worst-case scenario,” geologists report.

On Monday, at 2:11 p.m. local time, an explosive eruption rocked White Island, a small volcanic isle in New Zealand’s Bay of Plenty. A series of violent blasts rang out, flinging ash 12,000 feet into the sky and showering the volcano’s floor with hot debris, before everything fell silent a handful of minutes later.

It quickly became apparent that lives were lost. A number of tourists were on the island at the time, several of whom were right next to the volcano ’s active vent. At the time of writing, five deaths have been officially recorded, with several more people still unaccounted for. Recent reconnaissance flights over the island, also known by its Māori name Whakaari, have found no signs of life .

For this sort of volcanic paroxysm, “this was probably the actual worst-case scenario,” says Shane Cronin , a volcanologist and Earth scientist at the University of Auckland.

Understandably, this explosion took many by surprise. But for this volcano, and for the type of eruption style involved, it was nothing out of the ordinary: Similar eruptions, though not everyday occurrences, have happened at many volcanoes all over the world, and they will continue to appear without much warning. ( Find out the most dangerous volcanoes in the U.S., according to geologists .)

“This is a case of people being in the wrong place at the wrong time,” Janine Krippner , a volcanologist at the Smithsonian Institution’s Global Volcanism Program. “It’s horrible when it happens, but it will continue to happen over and over again.”

So why was the eruption at Whakaari/White Island so unpredictable and deadly? Let’s look at the facts.

Invisible warning signs

Whakaari/White Island is the summit of a complex submarine volcano. According to the Smithsonian Institution’s Global Volcanism Program , it is highly ebullient, engaging in a variety of eruption styles. Many feature moderate explosions.

As a result of its hyperactivity, and the frequency with which tourists visit, Whakaari/White Island is heavily monitored. Scientists were keeping a close eye on it to try and spot any behaviors that could potentially indicate the volcano was gearing up for something explosive.

Volcanologists with GNS Science, a New Zealand-based consulting group, spotted some localized surface deformation a few weeks earlier, says Geoff Kilgour , a scientist with the group. Deformation can be indicative of subsurface pressure changes caused by moving superheated liquids, gases, or magma. But in this case, the activity didn’t indicate that any major build-up of pressure was happening.

Still, monitoring efforts and reports from tourist companies also picked up some geyser-like convulsions at the time, along with an uptick in gas emissions and seismic rumblings. So, authorities did raise the volcano’s alert level. Alert level rises don’t mean that an eruption is inevitable, and in many cases, no eruption is forthcoming. Unfortunately, this time was a deadly exception.

Volcanic pressure cooker

Magma happens to sit close to the surface at Whakaari/White Island, and the molten rock constantly degases and heats up the plentiful supply of groundwater.

“You’ve got a really complex witches’ brew there,” Cronin says.

You May Also Like

Dramatic photos show La Palma volcano’s ongoing eruption

A rare and puzzling ‘domino effect’ triggered 4 powerful quakes in Afghanistan

Earth is a geological oddball in our solar system. This is why.

In addition, the magma’s path to the surface at this vent can get blocked, sometimes through the precipitation of minerals. That can lead to gases getting trapped underneath, where they keep accumulating and getting heated, creating a pocket of superheated elements somewhat like a pressure cooker. The sizzling water is often trying to boil off into steam, but it remains a liquid because of the immense pressure.

That means any crack in this geologic pressure cooker can produce a savage, speedy decompression event. The liquid water flashes to steam, expanding its volume 1,700 times in a heartbeat. The expansion energy is enough to shatter rocks and carve out geologic scars. When the pressure at the vent is released, a decompression wave rockets down into the volcano’s throat, where it encounters more pressurized water. Sometimes, the shockwave hits the magma, turning what is a steam-based outburst into a magma-driven eruption.

Even a steam blast is lethal if you stand too close to the vent, Cronin says. But the flying rocks, wet debris jets, and scorching air can also cause life-threatening harm.

And the cascade of events happens so fast that no one has time to react. The bulk of this eruption was a collection of impulsive bursts, each just tens of seconds long. The whole thing was pretty much over in two minutes, Kilgour says. The situation is especially dangerous on Whakaari/White Island, since the volcanic island is so small.

“There aren’t a lot of places you can be on White Island without being very close to the vent,” says Loÿc Vanderkluysen , a volcanologist at Drexel University.

Uncertain future

This eruption style—driven at least initially by steam—occurs at volcanoes all over the world. A somewhat similar steam blast happened at Japan’s Mount Ontake in 2014, killing 63 hikers . Crucially, these volcanic events don’t provide any definitive warning signs. No matter where in the world they occur, “no one has shown that they are able to forecast this kind of activity,” Kilgour says.

In this case, bad luck played a role, too. A similarly vicious eruption took place at Whakaari/White Island in 2016, but it happened at night. This one happened in the afternoon, while tourists were present. ( Here’s why people continue to live near active volcanoes .)

It’s unclear how this tragedy will affect the ability of tourists to visit this privately owned island in the long run. Scientists will continue to monitor and study the island closely, and Cronin and his colleagues are creating miniature hydrothermal explosions in a lab to try and better understand what causes them.

For now, though, there is much that remains unknown about these volcanic blasts, and it’s all the more important for everyone to understand the dangers and uncertainties involved in visiting an active volcano like Whakaari/White Island.

“It’s hard to portray uncertainty to tourists,” Kilgour says. “And there’s a lot of uncertainty about volcanic eruptions, especially these rapid-onset, or essentially unpredictable, events.”

What is grimly certain is that these types of volcanic explosions will continue to pose potentially deadly risks. “Of all the volcanoes that I’ve visited that are easily accessible to tourists,” Vanderkluysen says, “I don’t think there were any where I heard absolutely zero horror stories.”

Related Topics

• VOLCANOLOGY
• NATURAL DISASTERS

The world’s loneliest volcano may hold something truly rare. We went to find it.

Extreme heat is the future. Here are 10 practical ways to manage it.

Volcanoes blow smoke rings. Here's how they do it.

The island of Santorini is hiding an explosive secret

These breathtaking natural wonders no longer exist

• Best of the World
• Environment
• Paid Content

History & Culture

• History & Culture
• Out of Eden Walk
• Mind, Body, Wonder
• Terms of Use
• Privacy Policy
• Your US State Privacy Rights
• Children's Online Privacy Policy
• Interest-Based Ads
• About Nielsen Measurement
• Do Not Sell or Share My Personal Information
• Nat Geo Home
• Attend a Live Event
• Book a Trip
• Inspire Your Kids
• Shop Nat Geo
• Visit the D.C. Museum
• Learn About Our Impact
• Support Our Mission
• Advertise With Us
• Customer Service
• Renew Subscription
• Manage Your Subscription
• Work at Nat Geo
• Sign Up for Our Newsletters
• Contribute to Protect the Planet

Copyright © 1996-2015 National Geographic Society Copyright © 2015-2024 National Geographic Partners, LLC. All rights reserved

Case Studies Highlighting Impacts of Volcanic Ashfall, Gas & Vog

Case studies of past impacts and mitigation strategies for specific eruptions are given here. Sector specific information from these case studies also appear under their relevant topic headings (topics on the left).

Each case study begins with a brief overview discussing the size and volume of ash dispersed where known or approximated. Specific impact & mitigation information is organized into the following categories (where it is reported):

• Agriculture – Plants & Animals – includes livestock, pastoral land, horticulture and forestry.
• Health – direct and indirect from exposure.
• Infrastructure – may be summarised or some or all of the following are detailed depending on complexity
• Equipment & Communications
• Power supply
• Transportation
• Water & Wastewater
• Cleanup & Disposal
• Remobilization and coping with long-term ash – includes water and wind remobilised ash.
• Eyewitness &ndash Accounts from eruptions where available.
• Emergency management – monitoring, response during the eruption and recovery post eruption.

Please contact the Ash Web Team if you would like to contribute additional case study information. We are always looking for additional information.

New study reveals high risk of overdose deaths in Cuyahoga County among those using drugs when they’re alone

In Cuyahoga County, the stark reality of the opioid crisis is that most drug overdose victims die alone, with no one nearby to help.

A recent study , done in partnership with Case Western Reserve University and Cuyahoga County, highlights the critical need for “targeted harm-reduction strategies” in Northeast Ohio, where the opioid epidemic continues to claim lives at nearly twice the national average.

Those strategies include the distribution of Naloxone (an opioid antagonist that can reverse the effects of an overdose), and increasing the availability of medication-assisted treatment options and fentanyl test strips.

The research, using data from the  Cuyahoga County Medical Examiner’s Office , examined overdose deaths between 2016 and 2020, focusing on people using drugs when they were alone.

The study revealed that a staggering 75% of overdose victims were using drugs alone, a behavior strongly associated with increased mortality. Key findings indicate that individuals using drugs alone were more likely to be at home and less likely to receive life-saving interventions like naloxone, said  Daniel Flannery , the Dr. Semi J. and Ruth Begun Professor and director of the  Begun Center for Violence Prevention Research and Education .

“Being informed is crucial—knowledge equips you to take action,”  Flannery  said. “It’s about reviving someone in need, and if that’s not possible, contacting emergency services immediately. The chances of a fatal outcome significantly increase when there’s no one around to help.”

New policies and community efforts must prioritize reaching individuals at risk of using alone to curb the devastating impact of the opioid crisis, said  Vaishali Deo , research associate at the Begun Center and co-principal investigator in the research.

“Our findings underscore the urgent need for innovative harm-reduction strategies aimed at those most vulnerable—those using drugs alone,” Deo said. “Interventions must focus on reducing isolation and improving access to emergency medical care to prevent further loss of life.”

The research findings were published by the National Institutes of Health’s  National Library of Medicine .

Additional insights

• In Cuyahoga County, from 2016 through 2020, there were 2,944 unintentional overdose deaths for those over 18 years old. That’s 54 deaths per 100,000 residents. The national average is 28 overdose deaths per 100,000 residents.
• The study further details the demographics and circumstances surrounding overdose deaths in Cuyahoga County from 2016 to 2020. Most were non-Hispanic (94.9%), white (72.2%) and male (71.3%), with a significant portion 35 to 64 years old. Most lived in the City of Cleveland. Over half attained at least a high school diploma.
• Toxicology reports revealed that synthetic opioids, specifically illicitly manufactured fentanyl, was present in 72.7% of the deaths. Cocaine and heroin were also significant contributors, found in 41.6% and 29.6% of cases, respectively. Nearly 80% of overdose deaths involved the use of multiple substances.
• Despite the presence of bystanders in more than half the cases, most victims (74.9%) were using drugs alone at the time of their fatal overdose, mainly at home. Emergency medical services responded to most of the incidents, yet over 60% of victims were pronounced dead at the scene—highlighting the critical timing needed for interventions like naloxone, which was administered in just 28.6% of the cases.

Deo and Flannery were joined in the research by Sarah Fulton, a research associate at the Begun Center, and Manreet K. Bhullar, a senior forensic epidemiologist at the Cuyahoga County Medical Examiner’s Office, and Thomas P. Gilson, chief medical examiner of Cuyahoga County.

“These findings paint a stark picture of the opioid crisis in our community,” Gilson said. “The tragic reality is that too many people are dying alone, and we must act swiftly to implement lifesaving measures that can prevent these unnecessary deaths.”

For more information, contact Colin McEwen .

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 23 August 2024

Fingerprints of past volcanic eruptions can be detected in historical climate records using machine learning

• Johannes Meuer   ORCID: orcid.org/0009-0000-1998-6699 1 ,
• Claudia Timmreck   ORCID: orcid.org/0000-0001-5355-0426 2 ,
• Shih-Wei Fang   ORCID: orcid.org/0000-0003-3763-7753 2 , 3 , 4 &
• Christopher Kadow   ORCID: orcid.org/0000-0001-6537-3690 1  

Communications Earth & Environment volume  5 , Article number:  455 ( 2024 ) Cite this article

144 Accesses

2 Altmetric

Metrics details

• Atmospheric science
• Natural hazards
• Palaeoclimate

Accurately interpreting past climate variability, especially distinguishing between forced and unforced changes, is challenging. Proxy data confirm the occurrence of large volcanic eruptions, but linking temperature patterns to specific events or origins is elusive. We present a method combining historical climate records with a machine learning model trained on climate simulations of various volcanic magnitudes and locations. This approach identifies volcanic events based solely on post-eruption temperature patterns. Validations with historical simulations and reanalysis products confirm the identification of significant volcanic events. Explainable artificial intelligence methods point to specific fingerprints in the temperature record that reveal key regions for classification and point to possible physical mechanisms behind climate disruption for major events. We detect unexpected climatic effects from smaller events and identify a northern extratropical footprint for the unidentified 1809 event. This provides an additional line of evidence for past volcanoes and refines our understanding of volcanic impacts on climate.

Similar content being viewed by others

Climate change now detectable from any single day of weather at global scale

Training machine learning models on climate model output yields skillful interpretable seasonal precipitation forecasts

Pushing the frontiers in climate modelling and analysis with machine learning

Introduction.

Volcanic eruptions have been a major source of past climate variability 1 , 2 , 3 , causing regional to global cooling or warming, and altered dry-wet patterns depending on their location and strength 4 , 5 . Their impact on climate is a major source of uncertainty in predicting climate on seasonal to decadal timescales 6 , 7 . Notably, the Tambora eruption of 1815 stands as a stark illustration of its profound effect, evidenced by extreme cold records across Europe and North America in 1816 8 . These distinct volcanic climate effects can be identified in proxies and applied to reconstruct past volcanic events. However, uncertainties in reconstructing past events stem from the complex interaction between the signal and the inherent stochastic internal variability within our singular ensemble member - the Earth. In simpler terms, the challenge lies in untangling the influences of natural variability within our planet’s system, which can obscure the clarity of historical signals.

Traditionally, high-resolution analysis of ice cores combined with other palaeo-data has been used as an indicator of past volcanic eruptions 5 , 9 , 10 . However, the location of the volcanic source often remains unknown, and reconstructed climate variations cannot be accurately attributed to a specific volcanic event. A prominent example is the unidentified historical eruption of 1809, which is half the size of the well-known Tambora eruption of 1815 11 . This eruption was detected in ice core data 30 years ago 12 , but was not recorded in historical documents. Major surface cooling in 1810, probably in response to the eruption, has been detected in instrumental observations 13 , 14 and proxy records 15 , 16 , 17 and has in combination with the Tambora eruption led to strong climate responses 18 . Indeed, the cooling signature is spatially heterogeneous and inconsistent across different data sources 19 . This means that additional evidence is needed to have reliable attribution of reconstructed climate anomaly patterns to a (specific) volcanic eruption.

Recently, artificial neural networks (ANNs) have been used to distinguish between the forced and unforced (UF) signatures of ensemble climate simulations. Barnes et al. 20 have shown that ANNs can be trained to predict the year, given a map of annual mean temperature or precipitation from forced climate model simulations. In this way, ANNs can learn to identify forced patterns of change against a background of climate noise and model differences, providing reliable indicators of the underlying signal. Toms et al. 21 identify oceanic variability regions that contribute to predictability on decadal timescales in a fully coupled Earth system model. In addition, both approaches apply neural network explainability techniques, such as layer-wise relevance propagation (LRP), to help visualize these spatial patterns and highlight areas of disagreement between observed and simulated patterns. In addition, Kadow et al. 22 demonstrate the potential of using ANNs to reconstruct climate information of the past via transfer learning from numerical simulations. These studies have shown the potential of machine learning techniques as a valuable tool for identifying and reconstructing anomaly patterns in climate data.

In our study, due to the complex nature of climate data, we use a convolutional neural network (CNN) 23 architecture instead of simpler techniques such as linear regression. CNNs are well suited to capture the non-linear relationships and intricate spatio-temporal patterns in climate systems, such as the non-linear evolution of ENSO following volcanic eruptions. Their spatial invariance properties allow them to recognize patterns regardless of their position in the input data, which is crucial for global climate data. In addition, CNNs can automatically learn relevant features from raw data, whereas linear regression relies on manually constructed features that may miss important aspects of the data. This makes CNNs a robust and accurate choice for classifying volcanic eruption locations.

Data-driven identification of volcanic fingerprints in large ensemble simulations

To classify whether a volcanic eruption occurred and where it was located in the tropics or extratropics of each hemisphere, we train a CNN on the first post-volcanic boreal summer surface temperature anomalies.

Our CNN is constructed and evaluated on a suite of Max-Planck Institute Earth-System Model 1.1 24 large ensemble simulations with idealized volcanic radiative forcing (EVA-Ens 25 , 26 ), which are derived from the historical experiments of the Max-Planck-Institute Grand Ensemble (MPI-GE) 27 . The volcanic radiative forcing is prescribed in EVA-Ens and encompasses a combination of three eruption locations: NHE (northern hemispheric extra-tropical eruptions, 30°N–90°N), TR (tropical eruptions, 30°S–30°N), SHE (southern hemispheric extra-tropical eruptions, 30°S–90°S) and four different strengths, determined by the amount of sulphur injected into the atmosphere: 5, 10, 20, 40 Terra grams Sulphur (Tg S), and a UF (no volcanic forcing) scenario. The detailed setup is illustrated in Fig.  1 (see also ‘Methods’).

a Generating training and validation data for the classifier. The stratospheric aerosol optical depth (AOD) fields are shown for different volcanic eruption scenarios, northern hemispheric extratropical (NHE), tropical (TR), and southern hemispheric extratropical (SHE). These are prescribed in an earth system model (MPI-ESM1.1 24 ) to simulate the climate evolution after a volcanic eruption. b The convolutional neural network uses only the average boreal summer surface temperature anomalies one year after the volcanic eruption. The input layer is connected to three consecutive convolutional layers (Conv) + max-pooling layers (Pool), followed by two fully connected layers (FC). c The accuracy of the approach, given global surface temperature anomalies, by predicting each of the 100 ensemble members by 100 different trained models, dividing the accuracy scores into sulphur injection (volcanic strength) and the four predicted labels (NHE, TR, SHE, and UF). The overall average score is estimated from the mean of all correctly labelled samples.

The classification of excluded training data demonstrates strong performance, validated by the percentages depicted in Fig.  1 c. These percentages represent the ratio of correctly classified members to the total members within each ensemble. The correctly classified members are obtained by evaluating the excluded one member from each ensemble during training and repeating this process for each ensemble member. A detailed explanation can be found in the ‘Methods’ section. For eruptions greater than or equal to 10 Tg S, the classifications are almost perfect (>98%). The 5 Tg S eruption demonstrates accuracy exceeding 70% for both NHE and SHE classifications. This finding (detailed in Supplementary Table  1 ) carries two implications: Firstly, the signal-to-noise ratio for the 5 Tg S ensembles is smaller than for the other experiments and therefore the results are often influenced by internal variability, making it difficult to distinguish them from the UF scenarios 28 , 29 . Secondly, the NHE and SHE ensembles exhibit a closer resemblance to the TR ensemble compared to the larger eruptions. Notably, UF cases are correctly identified with 92% accuracy, underscoring our CNN’s capability to discern volcanic events based on subsequent boreal summer surface temperature anomaly patterns.

To understand how well our CNN performs on more realistic events, we assess the performance using 100 members from the historical runs of the MPI-GE, which is conducted with the same model as the EVA-Ens (Fig.  2 ). The volcanic forcing in the MPI-GE follows the Coupled Model Intercomparison Project Phase 5 (CMIP5) 30 , 31 protocol including different eruption locations and seasons (Fig.  2 a), which differs from the idealised volcanic forcing in EVA-Ens with three fixed locations and fixed timing mimicking the 1991 Pinatubo eruption (Fig.  1 a).

a The volcanic radiative forcing in the historical runs of the MPI-GE 27 in terms of monthly zonal mean AOD. b The probability of eruption location is predicted by our machine learning model. The distribution shows how many of the members were classified into the respective location, NHE, TR, and SHE, over the total ensemble members from each year of 1860–2000 for the Grand Ensemble. c Same as ( b ), but shows the percentage for the larger volcanic eruptions in numbers.

The input to the CNN are seasonal mean summer temperature anomalies (see ‘Methods’ for details), given for each year from 1860 to 2000. The CNN accurately detects the 1883 Krakatau eruption, identifying 87% (Fig.  2 c) of the members as a tropical eruption given the 1884 summer global surface temperature anomalies. As the climatic effects of volcanic eruptions diminish over several years, lower accuracies are found for 1884 (Fig.  2 b with darker colours indicating a high classification rate). The 1991 Pinatubo eruption, the 2nd largest eruption in the historical period in terms of radiative forcing after the Krakatau in 1883, is accurately identified by classifying 90% of the members as tropical eruptions. The Agung eruption of 1963, located in the tropics, is predominantly classified as SHE. This is consistent with the southerly shift of the tropical eruption’s aerosol load, into the southern hemisphere in the MPI-GE, which resembles more the AOD distribution of the southern hemisphere’s extra-tropical eruptions rather than the tropical ones in the training data set. Similarly, the CNN classifies most members of the tropical El Chichón eruption in 1982 as NHE, which can be attributed to the AOD displaying a pattern which is more similar to the northern hemispheric extra-tropical eruptions in the EVA-Ens.

Detection of historical volcanic events in observational records

How well does our CNN classify volcanic events based on observed temperature anomalies? To test the general applicability of our data-driven approach, we apply our CNN to six datasets (see ‘Methods’ for details). We focus on the European Reanalysis 5 (ERA5) 32 from 1960 to 2000 since there were four major volcanic eruptions in the period and only smaller eruptions before.

To estimate the uncertainty of our approach, we train 100 different CNNs, excluding each simulated ensemble member from the training. The ensemble accuracy of our approach results from the number of members identified as volcanic events by the 100 CNNs (see ‘Methods’ for details). Figure  3 a compares the mean accuracy of 100 CNN models on ERA5 data with the global mean stratospheric AOD from the CMIP6 volcanic forcing compilation 33 , 34 and the Global Space-based Stratospheric Aerosol Climatology (GloSSAC v2.0) dataset 35 ).

a The percentages of detected eruptions, shown as bars (red: NHE, purple: TR, and blue: SHE), are based on global surface temperature anomaly grids from ERA5 32 reanalysis. The monthly global mean AOD from CMIP6 66 climate models and the stratospheric aerosol observations (GloSSAC2.0 35 ) are in black. b The input mean surface temperature anomalies of the first post-volcano summer for 1963 Agung, 1982 El Chichón, and 1991 Pinatubo from ERA5 reanalysis. c The relevance heatmap of the corresponding surface temperature anomaly retrieved by LRP.

Our CNN model successfully identifies the presence or absence of eruptions for most years. The 1963 Agung eruption is again classified as a SHE instead of a tropical eruption with a hit rate of 100%, consistent with the classification for MPI-GE forced by the asymmetric AOD. In 1969, the figure shows a high probability ( ~80%) for an NHE eruption, which is not prominent in the CMIP6 AOD. However, previous studies 36 , 37 show that the 1968 tropical Fernandina eruption (0°S) had a large aerosol emission, exceeding, for example, the 1974 Fuego eruption.

The 1974 eruption of Fuego, which is located in the tropics, is classified as SHE in 1975 (>90%) and correctly as TR in 1976 (100%). This may be due to the fact that the eruption took place at the end of 1974 and had less of an impact on the climate in the first summer than in the second. Interestingly, the 1982 El Chichón eruption is correctly classified as TR (>90%), in contrast to the MPI-GE where the asymmetric AOD misleads our CNN. Furthermore, the eruption was already detected in 1982, which is consistent with the fact that El Chichón erupted in the spring of 1982 and already had an impact on global temperatures that summer 38 .

The 1991 Pinatubo eruption is successfully classified as tropical with an accuracy of  ~80% as TR and  ~20% as NHE. This high accuracy of identification not only demonstrates the applicability of our CNN, but also confirms the current understanding of the impact of volcanoes on global climate, since our classification model is constructed solely from physics-based numerical climate simulations with idealised volcanic forcing.

The results of the other reanalysis and observational datasets are shown in Supplementary Fig.  1 . The averaged result from all data shows a similar behaviour to that of ERA5. However, the differences in the individual data highlight the difficulties in classifying volcanic events, especially for the years preceding the Pinatubo eruption (1990 and 1991), where volcanic behaviour was detected in several cases. In addition, the HadCRUT5 and GISTEMP4 analyses for Pinatubo are classified as an NHE eruption, indicating problems with the lower resolution of these datasets.

A detailed examination of the decision-making mechanisms employed by our CNN could further uncover the reasons behind the accurate localization of certain eruptions in ERA5, such as El Chichón and Pinatubo, in contrast to the less consistent localization of Agung. To do so, we employ LRP 39 . This approach generates heatmaps that highlight the regions contributing to the machine learning model’s prediction (see ‘Methods’ for details).

Figure  3 b shows the following boreal summer surface temperature anomalies of the 1963 Agung (JJA 1964), 1982 El Chichón (JJA 1983), and 1991 Pinatubo (JJA 1992) eruptions from the ERA5 reanalysis, where the 1982 El Chichón and 1991 Pinatubo eruptions are mostly identified as TR eruptions, but the 1963 Agung eruption is incorrectly identified as a SHE eruption. For the 1964 input anomalies, the heatmap (Fig.  3 c) highlights the considerable focus of the CNN on negative anomalies in the tropical Pacific, southern Atlantic, and southern Indian Ocean—typical features associated with southern hemispheric extra-tropical eruptions (Supplementary Fig.  2 ). In the case of El Chichón, the map shows a substantial El-Niño event 40 in the tropical Pacific, to which the CNN only pays little attention. The attention is on regions with negative anomalies, equally in both hemispheres, a feature of tropical eruptions (Supplementary Fig.  2b ). This can also be seen in the 1992 temperature anomalies for Pinatubo, but in different regions (Fig.  3 b, c). In general, the heat maps point to the regions where the signal is strong. These relevant regions have also been identified in reanalysis data 38 and paleo reanalysis data (e.g. ref. 14 ).

Unlocking mysteries of the early 19th century

The good performance of the CNN model offers the potential to go even further back in time to classify and identify past volcanic eruptions in surface temperature reconstructions. One of the challenges is the limited availability of historical global surface temperature reconstructions 41 . Here we utilized the 20CR-v3 dataset 42 , the only global surface temperature reconstruction we found that is available back to 1806 with its experimental extension. The dataset uses only surface observations of synoptic pressure to assimilate global surface temperature grids and does not include information on external forcings such as volcanoes, solar radiation, greenhouse gases, and anthropogenic aerosols. This presents an additional challenge to our machine learning model in detecting volcanic fingerprints, as they are not artificially introduced. In addition, 20CR-v3 provides uncertainty estimates from 80 ensemble members using an ensemble Kalman filter, which facilitates uncertainty estimation for our approach. Over the 19th century, 12 major volcanic eruptions occurred according to the evolv2k-ENS 11 , which is an ensemble reconstruction of volcanic stratospheric sulphur injection and stratospheric AOD. We compared our machine learning model predictions with the reconstructed AOD field from evolv2k (Fig.  4 a). The accuracy of the CNN model was about 87% for identifying non-volcanic events, given an AOD threshold of 0.005, and 54% for identifying volcanic events (Fig.  4 b): 6 (6) out of 12 were clearly classified as volcanic (non-volcanic) events. The CNN had difficulties with the Icelandic high-latitude eruptions (1873 Grimsvötn and 1875 Askja) that were not included in our training data. The CNN also detects an NHE volcanic eruption in the summer of 1882 although the tropical eruption of Krakatau occurred in the summer of 1883. This could possibly be attributed to the weaker eruptions of Fuego in the summer of 1880 or Mayon in the summer of 1881. The Tarawera eruption in 1886 was not identified due to the previous Krakatau eruption, whose global cooling probably exacerbated the temperature anomalies calculated for 1887 (see ‘Methods’ for details). An interesting discovery is the 1831 eruption, which our model identifies as NHE. Although Sigl et al. 43 attribute the forcing to the tropical volcano Babuyan, there is some debate as to whether this eruption occurred near Sicily 44 , 45 , which would be consistent with our prediction.

a The AOD field is based on ref. 11 with the eruptions of the 19th century identified by Sigl et al. 43 . The unknown location of the first volcano eruption is marked as a dashed red line. The probability distributions from the predictions of our machine learning model were made on the 80 members of the 20CR-v3 reanalysis. b List of the largest volcano eruptions from the 19th century and the corresponding class probabilities, NHE, TR, SHE, and UF, estimated by our classifier.

We further focus on the early 19th century, which was one of the coldest periods of the past 500 years 3 , 46 , caused mainly by two strong volcanic eruptions. The unidentified eruption of 1809, that is likely located in the tropics 9 , 11 , 19 , and the well-known Tambora eruption of 1815 47 . The Tambora eruption is clearly classified by the CNN as TR with a score of 95% calculated from the 80 members of the 20CR-v3. The CNN also suggests a volcanic eruption in 1814 with a score of 79% as the Tambora eruption in April 1815 already influenced summer temperatures in the same year. Hence CNN predicts that a volcanic eruption had happened a year before.

The 1809 eruption was identified as NHE with a score of 83% (Fig.  5 a). Based on the surface temperature anomaly pattern in the experimental 20CR-v3 data, our CNN, therefore, suggests that the unidentified eruption from 1809 most likely has its aerosol loading concentrated in the northern extra-tropical region indicating that it was either a northern extratropical eruption or a tropical one but with an asymmetric forcing similar to the historic 1982 El Chichón eruption.

a Estimated average global AOD field from the evolv2k reconstruction 11 (top) and detected eruptions with predicted probabilities of eruption location (bottom) given the 80 members of the 20CR-v3 reanalysis 42 . b Average LRP heatmaps were retrieved from the 80 20CR-v3 reanalysis members of 1810 and 1816 and from 100 members of the 20 Tg S EVA-Ens of NHE and TR.

To analyse the classification of the 1809 event as NHE, we compare the average heatmaps generated from the 80 members of 20CR-v3 from summer 1810 with the heatmaps from summer 1816 for the 1815 Tambora eruption, as well as the heatmaps for the 20 Tg S NHE and TR EVA-Ens (Fig.  5 b). The 1810 and NHE heatmaps prominently highlight regions such as Greenland, northwest Africa, Central Asia and a part of central Russia. Interestingly, these areas do not show the same level of relevance in the 1816 and TR heatmaps, suggesting that the classification for 1809 is more similar to the 20 Tg S NHE EVA-Ens.

It should be noted that the surface temperature considered for the 1809 classification is under the experimental phase of the 20cr-v3 reanalysis. We also tested our CNN on the assimilation of tree-ring data 48 , which only covers the NH extratropics (30°N–90°N) and has a distinct temperature anomaly pattern compared to the 20CR-v3 (Fig.  6 ). In order to match these grids, we trained a new CNN model on simulated EVA-Ens surface temperature anomalies covering the same latitude range from 30°N to 90°N. Although this model achieves a lower skill score (72% compared to 92%), the results are still quite good, especially for larger eruptions. The classification of the 1809 eruption remains as NHE, with a skill of 99% confirming the results of the CNN trained on global data.

a The results on EVA-Ens of the machine learning model trained on surface temperature anomalies ranging from 30°N to 90°N, dividing the accuracy scores into sulphur injection (volcano strength) and the four predicted labels (NHE, TR, SHE, and UF). The total average score is estimated from the mean over all correctly labelled samples. b Tree-ring temperature anomalies that were assimilated using the MPI-ESM 48 and the corresponding relevance map.

The good performance of the north-extratropical-only CNN model on assimilated tree-ring data also highlights the application of our method not only on a global scale but also on a large hemispheric scale. This offers new pathways to identify volcanic eruptions and their hemispheric location in large-scale temperature reconstructions beyond the common era.

Overall, our data-driven approach demonstrates the feasibility of classifying past volcanic events from observational datasets but it has some limitations. As the eruptions in our training dataset all start only in July from only three locations, events with different eruption seasons or locations which were not included in our training simulations are more difficult to identify correctly, such as the high latitude and winter eruptions. Extended training data would therefore be essential to improve the classification skill.

With the potential to accurately detect past large eruptions and their possible eruption locations from observational data, CNN provides an additional line of evidence for identifying past volcanic events, helping the community to reconcile the historical record with climate responses. By using only temperature reconstructions from observed fields, our data-driven method also shows that our simulated volcanic fingerprints for training are reasonable. Even though we have a limited number of events in recent decades for which we have better observations, the study does indeed support the current understanding of volcanic impacts on climate, as well as their implementation in future projections and social impacts. In addition, this research enables historical investigations of past volcanic activity and its impact on human civilisation, ecosystems and the environment. As exemplified by the unidentified 1809 eruption, it has the potential to reveal previously overlooked or poorly documented volcanic events, enriching our understanding of the climate effects of historical eruptions.

Training data

For training the CNN, we used seasonal mean climate anomalies from different large ensemble simulations of idealized volcanic eruptions of different locations and strengths (EVA-Ens). The CNN is trained with boreal summer mean global surface temperature anomalies one year after the volcanic eruption. The idealized volcanic forcing experiments were performed with the Max-Planck-Institute Earth-System-Model (MPI-ESM1.1-LR), which is an intermediate version between the MPI-ESM CMIP5 version 24 and CMIP6 version 49 . The MPI-ESM1.1-LR has an atmospheric horizontal resolution of 1.8° with 47 vertical levels up to 0.01 hPa and a nominal ocean resolution of 1.5° with 64 vertical levels.

The EVA-Ens experiments were designed as sensitivity experiments to the Max Planck Institute Earth System Model Grand Ensemble (MPI-GE, 27 ) historical run, where the prescribed historical volcanic forcing was replaced in 1991 by more idealized volcanic forcing data. The EVA-Ens simulations were branched from 100 members of the historical experiments of the MPI-GE in January 1991 and for a time span of 3 years. We create large 100-member ensembles from combinations of four strengths (40 Tg S, 20 Tg S, 10 Tg S, and 5 Tg S) and three locations (tropical and northern/southern extra-tropical), for a Northern Hemisphere summer eruption in June 1991 25 , 26 , 28 . In addition, we also run one 100-member ensemble experiment without any volcanic forcing. To put the strengths into perspective, the recent Pinatubo eruption from 1991 had an estimated sulphur emission between 5 Tg S and 10 Tg S 50 . Other forcings, such as solar variations, greenhouse gases and anthropogenic aerosols remain for all experiments unchanged and are the same as in the historic simulation of the MPI-GE. Each of these simulations provides global monthly grids ranging from January 1991 to December 1993.

The volcanic forcing is prescribed in the MPI-ESM by zonally and monthly mean optical parameters (aerosol extinction, single scattering, albedo and asymmetry factor). In the MPI-GE historical runs, the PADS dataset 30 , 31 is used for the volcanic aerosol forcing of historic eruptions. For our idealized volcanic experiments, we compiled the radiative forcing with the easy volcanic aerosol (EVA) forcing generator 51 . EVA is based on a parameterized three-box model of stratospheric transport and scaling relationships that calculates stratospheric aerosol optical properties from eruption time and latitude, estimated stratospheric sulphur emission and wavelength. The stratospheric AOD fields generated with EVA for different emission strengths and locations can be seen in Supplementary Fig.  4 .

To ensure an even distribution of output classes in our training data, we included 100 members from the years 1991, 1992, 1993, and 1994 of the unperturbed runs to cover the no-volcano case. In total, this results in a dataset of 1600 training samples.

Reanalysis and observation of data

Our assessment included an ensemble of four reanalyses and two observational analyses. An overview of all datasets used for validation and evaluation can be seen in Supplementary Table  1 . Regarding the reanalyses, the ERA-5 reanalysis 32 is a comprehensive global atmospheric reanalysis produced under the auspices of the Copernicus Climate Change Service at the European Centre for Medium-Term Weather Forecasts. It covers the period from January 1940 to the present and provides hourly estimates of atmospheric, land and ocean climate variables at a spatial resolution of 0.25°/0.25°, covering 137 levels from the Earth’s surface to 80 km altitude. The 20CR-v3 42 is a four-dimensional weather reconstruction dataset covering the period 1836–2015, with an experimental phase covering the period 1806–1835, developed by NOAA’s Physical Sciences Laboratory. This reanalysis incorporates only surface pressure observations into NOAA’s Global Forecast System to predict various climate variables, including temperature, pressure, winds, humidity, solar radiation, and cloud cover. The data are presented as daily estimates with a spatial resolution of 0.7°/0.7° and include 80 ensemble members designed to estimate uncertainty. The Japanese 55-year reanalysis (JRA55) 52 , by the Japanese Meteorological Agency, integrates historical observations through a sophisticated operational data assimilation system. The dataset covers the period from 1958 to 2024, providing data at 3-h intervals and a spatial resolution of 1.25°/1.25°. In addition, NCEP1 53 is the result of a collaboration between the National Centers for Atmospheric Prediction (NCEP) and the National Center for Atmospheric Research and covers the period from 1948 to the present. The reanalysis uses multiple numerical weather prediction models, such as the global forecast system, the climate forecast system and others to assimilate observations. The data is given at 6-hourly, daily and monthly intervals with a spatial resolution of 1.875°/1.875°.

HadCRUT5 54 , a joint product of the Met Office Hadley Centre and the Climatic Research Unit at the University of East Anglia, contains historical surface temperature anomalies on a global scale, covering the period from 1961 to 1990. This dataset is available from 1850 onwards and provides monthly grids with a spatial resolution of 5°/5°. The most recent HadCRUT.5.0.2.0 analysis includes a total of 200 ensembles from which we only considered the ensemble mean. Finally, the GISTEMP4 55 , 56 analysis represents an estimate of global surface temperature changes produced by the National Aeronautics and Space Administration from 1880 to the present. This dataset is presented at a spatial resolution of 2°/2° on a monthly basis. The observational analyses both contain missing values, mainly concentrated near the poles due to a lack of observations. By simply setting these to zero, we found that our CNN still performed well for these observations. However, for future investigations, more sophisticated interpolation techniques could be used to improve the quality of the data.

Data pre-processing

All data were provided in NetCDF4 (Network Common Data Form) and pre-processed using the climate data operators 57 and Freva, a free evaluation system 58 . For the EVA-Ens dataset, we calculated anomalies from a reference period of 1985-1990 from the MPI-GE ensemble mean. This was subtracted from each of the EVA-Ens members. We chose this reference period, because it contains only small volcanic forcings, that would not interfere with the signals from the EVA simulations. For each resulting ensemble member anomaly, we calculated the boreal summer mean from 1992, one year after the simulated eruption. These were then each labelled as NHE, TR, SHE, and UF, forming our input-label pairs for training.

To address the potential influence of the global warming trend, anomalies in the MPI-GE, reanalyses, and observational data were calculated dynamically during the evaluation. In specific, for each year Y N , we selected the preceding years Y N −3 and Y N −2 as reference periods:

From these anomalies, we again calculated the boreal summer mean for each year, respectively, forming the input data for our evaluation. The short reference periods of only two years prevent the compensating effects of eruptions occurring in quick succession. However, if the eruptions are very close to each other, this might still pose a problem to the classifier. The gap year in between the reference period and the year under consideration is intended to deal with eruptions that occur in the early part of the year and have already had an impact on the same year’s boreal summer temperatures. This would reduce the anomalies of the following year and thus challenge our machine learning model. All datasets have been conservatively remapped to 1.8°/1.8° (192 × 92), the original resolution of the EVA-Ens dataset.

Deep-learning model

We used a CNN, a machine learning approach specifically designed to process gridded data by extracting spatial features. Our CNN consists of three convolutional layers with 5 × 5 kernels, a stride of 2 and global padding, which applies circular padding to the boundaries of the longitudes to account for the round shape of the Earth. Each convolutional layer is followed by a subsequent max-pooling operation (2 × 2) and rectified linear unit activation, and finally two fully connected (FC) layers. The implementation was done in Python 3.9, using the PyTorch deep learning framework. We trained each CNN over 10,000 iterations, using a batch size of 4 and a learning rate of 5e-5 with the Adam optimiser. For the loss function, we used the standard classification loss, which is the cross-entropy loss.

LRP is a technique for bringing accountability to complex deep-learning models. The trained CNN is propagated backwards after making a prediction, using a set of propagation rules. In this study, we used the interpretability library for PyTorch by Captum 59 . We applied the γ -propagation rule 60 , which aims to reduce noise and improve stability by favouring the effect of regions that contributed positively to the prediction. However, other propagation rules and neural network explainability techniques gave very similar results. In Supplementary Fig.  1 , we show a comparison of gradient-based interpretability techniques, including three different LRP propagation rules (Gamma, Alpha1Beta0, Epsilon) and Integrated Gradients 61 . In addition, we show the results of the post-hoc interpretability techniques SHAP 62 and Occlusion 63 . It is important to note that SHAP and Occlusion are fundamentally different from LRP-based methods and Integrated Gradients as they do not rely on the internal structure of the model.

Model tuning

The initial model was taken from the PyTorch documentation for an image classifier 64 designed to process images of size 3 × 32 × 32, denoting 3-channel colour images of 32 × 32 pixels. Our adaptation involved modifying the network architecture to handle images of size 1 × 192 × 192. Initially, using the original architecture with two convolutional layers and three FC layers yielded an accuracy of approximately 80% on the simulated dataset. Subsequently, through experimentation, we optimised the model by introducing an additional convolutional layer while removing an FC layer, resulting in a significant improvement to 92% accuracy. In addition, fine-tuning the learning rate to 5 e  − 5, as opposed to the original 1 e  − 3, significantly improved the training process. We also investigated the effect of the number of epochs and batch size on the effectiveness of the training. After extensive evaluation, we found that about 10,000 iterations and a batch size of four produced optimal results on our evaluation set. Deviating from this threshold resulted in either reduced performance due to underfitting or exacerbated overfitting due to excessive iterations.

Model validation

Typically, CNNs require extensive training data to effectively capture complex patterns and establish reliable classification capabilities. However, expanding our training dataset is challenging and constrained by the 100 members of the MPI-GE. When considering adding additional scenarios to the EVA-Ens for future exploration, we encountered potential redundancy issues, exemplified by the 2.5 Tg S ensemble (Supplementary Table  2 ), which we discuss further in the sensitivity experiments section.

To overcome this limitation, we developed an approach using leave-one-out cross-validation 65 . This method involves isolating a piece of data, using the remaining data for CNN training, and using the isolated data for validation. By iterating this process several times, each time isolating a different split, it is possible to train several different CNNs, allowing comprehensive validation without excessive data generation. While generating 100 ensemble members can be computationally demanding, our strategy mitigates this burden while ensuring robust training and validation, thereby increasing the accuracy and efficiency of our analysis. We observed a marginal improvement in results by reducing the validation data to a single member, thus maximising the sample availability for training the CNN model. This provided us with 100 trained CNN models, allowing us to estimate uncertainty when evaluating observational data without multiple members. By using the same data across our 100 trained CNNs, we can derive a probability distribution of predictions.

The validation metrics across all tables were computed by averaging the results of the 100 trained CNNs. Each CNN produced a single output for each scenario (e.g. TR—5 Tg S), resulting in 16 outputs per network. Aggregation of these outputs produced 100 results for each scenario and a cumulative total of 1600 validated results.

Sensitivity experiments

The CNN model exhibited sensitivity to data normalization when utilizing reanalysis and observational inputs. Specifically, the prediction accuracy varied depending on the reference period used for data normalization. Extending the reference period beyond three years into the past resulted in certain volcanoes being undetected. For instance, the 1982 El Chichón eruption was consistently misclassified as UF across multiple datasets. Conversely, selecting only a single year as a reference caused the model to overly prioritize NHE, TR, or SHE classifications.

We tried extending our training data with 2.5 Tg S ensembles (Supplementary Table  2 ). We retrained a new CNN model on this extended dataset, which increased the accuracy of the 5 Tg S ensembles. However, it significantly decreased the accuracy of the UF ensemble and the overall detection accuracy, even when excluding the 2.5 Tg S ensembles from the evaluation. This suggests that the weak 2.5 Tg S events are barely distinguishable from the internal variability in their surface temperature anomaly patterns and we do not include them in our CNN in order to maintain the high detection accuracy on UF events.

We further investigated how other variables performed compared to the surface temperature anomalies: precipitation (Supplementary Table  3 ), sea-level pressure (Supplementary Table  4 ) and all three variables combined (Supplementary Table  5 ). The CNN trained on precipitation anomalies achieved a total accuracy rate of over 83%, falling slightly short of the CNN trained solely with surface temperature anomalies. In contrast, the CNN trained with sea-level pressure anomalies exhibited the lowest accuracy, with a score of 75% correct classifications. This implies that volcanic eruptions leave a more identifiable signal in terms of their effect on global temperatures compared to precipitation and sea-level pressure. Interestingly, the CNN trained with all three variables—precipitation, sea-level pressure, and surface temperature anomalies—achieved slightly lower accuracy when compared to the CNN utilizing only surface temperature anomalies. This suggests that similar to including the ensembles of 2.5 Tg S, including additional variables alongside surface temperature anomalies may introduce redundancy rather than enhance the performance.

The observational datasets we examined posed challenges due to varying resolutions. Reanalysis datasets offered higher resolution compared to the simulations used for training our machine learning model. Conversely, the observational analyses HadCRUT5 and GISTEMP4 required conservative downscaling to align with the higher resolution. In Supplementary Fig.  1 , we observe a misclassification by the CNN, labelling the 1992 Pinatubo eruption as NHE in these datasets, potentially due to the resolution limitations posing a challenge to our CNN model. However, the majority of classifications align with the higher spatial resolution reanalysis datasets despite the considerable resolution gap.

Data availability

The primary data and scripts used in the analysis, along with other supplementary materials that are useful for reproducing the model simulations, have been archived by the Max Planck Institute for Meteorology. These materials can be accessed via the following links: • http://hdl.handle.net/21.11116/0000-0007-8B38-E • https://hdl.handle.net/21.11116/0000-000D-4B1F-E . Additionally, the following open-source reanalyses and analyses were used for this study, and are available through public repositories: •ERA5: https://doi.org/10.24381/cds.adbb2d47 •20CR-v3: https://doi.org/10.5065/H93G-WS83 •JRA55: https://doi.org/10.2151/jmsj.2015-001 •NCEP-1: https://downloads.psl.noaa.gov/Datasets/ncep.reanalysis/ •HadCRUT5: https://www.metoffice.gov.uk/hadobs/hadcrut5/data/HadCRUT.5.0.2.0/download.html •GISTEMP: https://data.giss.nasa.gov/gistemp/ These resources provide the necessary datasets for reproducing our analyses.

Code availability

The code used to perform the analysis and simulations, along with other supporting software to reproduce the results, is available at: https://doi.org/10.5281/zenodo.12755134 . This includes the scripts and documentation necessary to replicate our study’s results. It facilitates the validation and further exploration of the methods used in our research, as well as access to the trained models that generated the proposed results.

Hegerl, G. C., Crowley, T. J., Hyde, W. T. & Frame, D. J. Climate sensitivity constrained by temperature reconstructions over the past seven centuries. Nature 440 , 1029–1032 (2006).

Article   CAS   Google Scholar  

Schurer, A. P., Tett, S. F. & Hegerl, G. C. Small influence of solar variability on climate over the past millennium. Nat. Geosci. 7 , 104–108 (2014).

Neukom, R. et al. Consistent multidecadal variability in global temperature reconstructions and simulations over the common era. Nat. Geosci. 12 , 643–649 (2019).

Article   Google Scholar  

Timmreck, C. Modeling the climatic effects of large explosive volcanic eruptions. Wiley Interdiscip. Rev. Clim. Change 3 , 545–564 (2012).

Marshall, L. R. et al. Volcanic effects on climate: recent advances and future avenues. Bull. Volcanol. 84 , 54 (2022).

Timmreck, C., Pohlmann, H., Illing, S. & Kadow, C. The impact of stratospheric volcanic aerosol on decadal-scale climate predictions. Geophys. Res. Lett. 43 , 834–842 (2016).

Illing, S., Kadow, C., Pohlmann, H. & Timmreck, C. Assessing the impact of a future volcanic eruption on decadal predictions. Earth Syst. Dyn. 9 , 701–715 (2018).

Raible, C. C. et al. Tambora 1815 as a test case for high impact volcanic eruptions: earth system effects. Wiley Interdiscip. Rev. Clim. Change 7 , 569–589 (2016).

Sigl, M. et al. Timing and climate forcing of volcanic eruptions for the past 2,500 years. Nature 523 , 543–549 (2015).

Burke, A. et al. Stratospheric eruptions from tropical and extra-tropical volcanoes constrained using high-resolution sulfur isotopes in ice cores. Earth Planet. Sci. Lett. 521 , 113–119 (2019).

Toohey, M. & Sigl, M. Volcanic stratospheric sulfur injections and aerosol optical depth from 500 BCE to 1900 CE. Earth Syst. Sci. Data 9 , 809–831 (2017).

Dai, J., Mosley-Thompson, E. & Thompson, L. G. Ice core evidence for an explosive tropical volcanic eruption 6 years preceding tambora. J. Geophys. Res. Atmos. 96 , 17361–17366 (1991).

Brohan, P. et al. Constraining the temperature history of the past millennium using early instrumental observations. Climate 8 , 1551–1563 (2012).

Google Scholar  

Brönnimann, S. et al. Unlocking pre-1850 instrumental meteorological records: A global inventory. Bull. Am. Meteorol. Soc. 100 , ES389–ES413 (2019).

Wilson, R. et al. Last millennium northern hemisphere summer temperatures from tree rings: part I: the long term context. Quat. Sci. Rev. 134 , 1–18 (2016).

Anchukaitis, K. J. et al. Last millennium northern hemisphere summer temperatures from tree rings: Part ii, spatially resolved reconstructions. Quat. Sci. Rev. 163 , 1–22 (2017).

Guillet, S. et al. Climate response to the samalas volcanic eruption in 1257 revealed by proxy records. Nat. Geosci. 10 , 123–128 (2017).

Brönnimann, S. et al. Last phase of the little ice age forced by volcanic eruptions. Nat. Geosci. 12 , 650–656 (2019).

Timmreck, C. et al. The unidentified eruption of 1809: a climatic cold case. Climate 17 , 1455–1482 (2021).

Barnes, E. A. et al. Indicator patterns of forced change learned by an artificial neural network. J. Adv. Model. Earth Syst. 12 , e2020MS002195 (2020).

Toms, B. A., Barnes, E. A. & Hurrell, J. W. Assessing decadal predictability in an earth-system model using explainable neural networks. Geophys. Res. Lett. 48 , e2021GL093842 (2021).

Kadow, C., Hall, D. M. & Ulbrich, U. Artificial intelligence reconstructs missing climate information. Nat. Geosci. 13 , 408–413 (2020).

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. nature 521 , 436–444 (2015).

Giorgetta, M. et al. Climate change from 1850 to 2100 in mpi-esm simulations for the coupled model intercomparison project 5. J. Adv. Model. Earth Syst. https://doi.org/10.1002/jame.20038 (2012).

Azoulay, A., Schmidt, H. & Timmreck, C. The arctic polar vortex response to volcanic forcing of different strengths. J. Geophys. Res. Atmos. 126 , e2020JD034450 (2021).

Freychet, N., Schurer, A. P., Ballinger, A. P., Suarez-Gutierrez, L. & Timmreck, C. Assessing the impact of very large volcanic eruptions on the risk of extreme climate events. Environ. Res. Clim. 2 , 035015 (2023).

Maher, N. et al. The max planck institute grand ensemble: enabling the exploration of climate system variability. J. Adv. Modeling Earth Syst. 11 , 2050–2069 (2019).

D’Agostino, R. & Timmreck, C. Sensitivity of regional monsoons to idealised equatorial volcanic eruption of different sulfur emission strengths. Environ. Res. Lett. 17 , 054001 (2022).

Timmreck, C. et al. Linearity of the climate response to increasingly strong tropical volcanic eruptions in a large ensemble framework. J. Clim. 37 , 2455–2470 (2024).

Schmidt, H. et al. Response of the middle atmosphere to anthropogenic and natural forcings in the cmip5 simulations with the max planck institute earth system model. J. Adv. Model. Earth Syst. 5 , 98–116 (2013).

Stenchikov, G. L. et al. Radiative forcing from the 1991 mount pinatubo volcanic eruption. J. Geophys. Res. Atmos. 103 , 13837–13857 (1998).

Hersbach, H. et al. ERA5 hourly data on single levels from 1940 to present [dataset]. Copernicus climate change service (C3S) climate data store (CDS) https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=overview (2023).

Luo, B. Stratospheric aerosol data for use in cmip6 models https://iacftp.ethz.ch/pub_read/luo/CMIP6/Readme_Data_Description.pdf (2016).

Luo, B. Release notes stratospheric aerosol radiative forcing and sad version v4. 0.01850–2016 https://iacftp.ethz.ch/pub_read/luo/CMIP6_SAD_radForcing_v4.0.0/Release_note_v4.0.0.pdf (2018).

Kovilakam, M. et al. The global space-based stratospheric aerosol climatology (version 2.0): 1979–2018. Earth Syst. Sci. Data 12 , 2607–2634 (2020).

Zielinski, G. A. Use of paleo-records in determining variability within the volcanism–climate system. Quat. Sci. Rev. 19 , 417–438 (2000).

Gleixner, S. Southern Annular Mode response to volcanic eruptions in the MPI-ESM . Master’s thesis, Christian-Albrechts-Universität https://oceanrep.geomar.de/id/eprint/14714/ (2012).

Fujiwara, M., Martineau, P. & Wright, J. S. Surface temperature response to the major volcanic eruptions in multiple reanalysis data sets. Atmos. Chem. Phys. 20 , 345–374 (2020).

Bach, S. et al. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS One 10 , e0130140 (2015).

Chen, D., Cane, M. A., Kaplan, A., Zebiak, S. E. & Huang, D. Predictability of el niño over the past 148 years. Nature 428 , 733–736 (2004).

Christiansen, B. & Ljungqvist, F. C. Challenges and perspectives for large-scale temperature reconstructions of the past two millennia. Rev. Geophys. 55 , 40–96 (2017).

Slivinski, L. C. et al. Noaa-cires-doe twentieth century reanalysis version 3 (2019).

Sigl, M. et al. A new bipolar ice core record of volcanism from wais divide and neem and implications for climate forcing of the last 2000 years. J. Geophys. Res. Atmos. 118 , 1151–1169 (2013).

Garrison, C. S., Kilburn, C. R. & Edwards, S. J. The 1831 eruption of babuyan claro that never happened: has the source of one of the largest volcanic climate forcing events of the nineteenth century been misattributed? J. Appl. Volcanol. 7 , 1–21 (2018).

Garrison, C., Kilburn, C., Smart, D. & Edwards, S. The blue suns of 1831: Was the eruption of ferdinandea, near sicily, one of the largest volcanic climate forcing events of the nineteenth century? Climate 17 , 2607–2632 (2021).

Reichen, L. et al. A decade of cold eurasian winters reconstructed for the early 19th century. Nat. Commun. 13 , 2116 (2022).

Fang, S.-W., Timmreck, C., Jungclaus, J., Krüger, K. & Schmidt, H. On the additivity of climate responses to the volcanic and solar forcing in the early 19th century. Earth Syst. Dyn. 13 , 1535–1555 (2022).

King, J. M. et al. A data assimilation approach to last millennium temperature field reconstruction using a limited high-sensitivity proxy network. J. Clim. 34 , 7091–7111 (2021).

Mauritsen, T. et al. Developments in the mpi-m earth system model version 1.2 (mpi-esm1. 2) and its response to increasing co2. J. Adv. Model. Earth Syst. 11 , 998–1038 (2019).

Timmreck, C. et al. The interactive stratospheric aerosol model intercomparison project (isa-mip): motivation and experimental design. Geosci. Model Dev. 11 , 2581–2608 (2018).

Toohey, M., Stevens, B., Schmidt, H. & Timmreck, C. Easy volcanic aerosol (eva v1. 0): an idealized forcing generator for climate simulations. Geosci. Model Dev. 9 , 4049–4070 (2016).

Kobayashi, S. et al. The jra-55 reanalysis: general specifications and basic characteristics. J. Meteorol. Soc. Jpn. Ser. II 93 , 5–48 (2015).

Kalnay, E. et al. The NCEP/NCAR 40-year reanalysis project. In Renewable Energy , 146–194; https://downloads.psl.noaa.gov/Datasets/ncep.reanalysis/ (2018).

Morice, C. P. et al. An updated assessment of near-surface temperature change from 1850: the HadCRUT5 data set. J. Geophys. Res. Atmos. 126 , e2019JD032361 (2021).

for Space Studies, N. G. I. Giss Surface Temperature Analysis (gistemp), version 4 (accessed 26 October 2023); https://data.giss.nasa.gov/gistemp/

Lenssen, N. J. et al. Improvements in the gistemp uncertainty model. J. Geophys. Res. Atmos. 124 , 6307–6326 (2019).

Schulzweida, U. Cdo user guide https://doi.org/10.5281/zenodo.10020800 (2023).

Kadow, C. et al. Introduction to Freva—a free evaluation system framework for earth system modeling. J. Open Res. Softw. 9 , 13 (2021).

Kokhlikyan, N. et al. Captum: a unified and generic model interpretability library for PyTorch. Preprint at https://doi.org/10.48550/arXiv.2009.07896 (2020).

Montavon, G., Binder, A., Lapuschkin, S., Samek, W. & Müller, K.-R. Layer-wise relevance propagation: an overview. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning 193–209 (Springer, 2019).

Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. In International Conference on Machine Learning 3319–3328 (PMLR, 2017).

Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems 4765–4774 (Curran Associates, Inc., 2017).

Zeiler, M. D. & Fergus, R. Visualizing and understanding convolutional networks. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6–12, 2014, Proceedings, Part I 13 818–833 (Springer, 2014).

Training a Classifier (accessed 23 January 2023); https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html

Quenouille, M. H. Approximate tests of correlation in time-series 3. In Mathematical Proceedings of the Cambridge Philosophical Society 483–484 (Cambridge University Press, 1949).

Gidden, M. J. et al. Global emissions pathways under different socioeconomic scenarios for use in cmip6: a dataset of harmonized emissions trajectories through the end of the century. Geosci. Model Dev. 12 , 1443–1475 (2019).

Download references

Acknowledgements

The authors are grateful to the German Climate Computing Center (DKRZ) for providing the hardware for the calculations and all the data that were used for this study. Johannes Meuer and Claudia Timmreck are funded by the German National Funding Agency (DFG), provided by the research unit FOR 2820, titled “Revisiting The Volcanic Impact on Atmosphere and Climate-Preparations for the Next Big Volcanic Eruption" (VolImpact), with the project number 398006378. Shih-Wei Fang acknowledges support from the German Federal Ministry of Education and Research (BMBF), research programme “ROMIC-II, ISOVIC” (FKZ: 01LG1909B) and was supported by the Institute for Basic Science (IBS), Republic of Korea, under IBS-R028-D1. Support for the Twentieth Century Reanalysis Project version three dataset is provided by the U.S. Department of Energy, Office of Science Biological and Environmental Research (BER), the National Oceanic and Atmospheric Administration Climate Programme Office, and the NOAA Physical Sciences Laboratory. Thanks to ICDC, CEN, University of Hamburg for data support.

Open Access funding enabled and organized by Projekt DEAL.

Author information

Authors and affiliations.

Data-Analysis, German Climate Computing Center (DKRZ), Hamburg, Germany

Johannes Meuer & Christopher Kadow

Max Planck Institute for Meteorology, Hamburg, Germany

Claudia Timmreck & Shih-Wei Fang

Center for Climate Physics, Institute for Basic Science (IBS), Busan, Republic of Korea

Shih-Wei Fang

Pusan National University, Busan, Republic of Korea

You can also search for this author in PubMed   Google Scholar

Contributions

Johannes Meuer conceived the study, designed the research methodology, conducted the data analysis, and collected and processed the data. Claudia Timmreck performed the simulations for the EVA-Ens dataset. Christopher Kadow contributed crucial methodological insights. Shih-Wei Fang provided valuable ideas for validation methods. All authors contributed to interpreting the results and offered critical feedback throughout the manuscript’s preparation. Additionally, all authors participated in drafting and editing the manuscript.

Corresponding author

Correspondence to Johannes Meuer .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Communications Earth and Environment thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Rachael Rhodes, Aliénor Lavergne, and Carolina Ortiz Guerrero. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Peer review file, supplementary information, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Meuer, J., Timmreck, C., Fang, SW. et al. Fingerprints of past volcanic eruptions can be detected in historical climate records using machine learning. Commun Earth Environ 5 , 455 (2024). https://doi.org/10.1038/s43247-024-01617-y

Download citation

Received : 05 February 2024

Accepted : 08 August 2024

Published : 23 August 2024

DOI : https://doi.org/10.1038/s43247-024-01617-y

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.