Repeating long-period (lp) earthquakes are commonly observed in volcanic regions worldwide. They are usually explained in terms of a volcanic source effect or anomalous propagation through the volcano. Recently, large lp events have also been associated with the motion of massive ice streams. Our joint analysis of climatic and new seismic data shows that small lp events observed at Katla volcano, Iceland, are in fact related to ice movement in a steep outlet glacier and not, as previously thought, to volcanic intrusive activity. The over 13000 lp events recorded since 2000 are consistent in character and magnitude with seasonal changes of the glacier. As the current global warming trend could cause similar earthquake sequences at other glacier covered volcanoes, identifying them as glacial rather than eruption precursors is vital.
 Katla is one of the biggest and most active volcanoes in Iceland and it is overlain by Mýrdalsjökull, the fourth largest glacier on the island (Figure 1a). The volcano has not erupted since 1918, despite quite regular semi centennial eruptions throughout historic times. Katla eruptions are explosive and accompanied by tephra fall and jökulhlaups [Larsen, 2000].
 The seismic activity can be divided into two groups; volcano tectonic (vt) earthquakes with hypocenters within the volcano's caldera down to 20 km depth, and shallow lp-seismicity in Goðabunga, the western part of Katla [Einarsson and Brandsdóttir, 2000]. The vt-earthquakes typically have dominant frequencies of 7–15 Hz while the lp events have 1–4 Hz. A few short lived vt-seismicity episodes in the caldera during the last century were accompanied by glacier floods but no apparent eruption. Despite much activity, no events with magnitudes above 3.3 have been observed in Goðabunga [Jónsdóttir et al., 2007] while five events with magnitude 4 have been recorded in the less active caldera.
 The seasonal lp-seismicity in Goðabunga, observed first in the 1950's [Einarsson and Brandsdóttir, 2000], started to increase dramatically in 2001. At the same time (October 2001) the observation capability of the seismic network increased (by moving a noisy station south of Katla closer to the seismicity south–west of Katla), and the magnitude of completeness was lowered from Mc 1.5 to Mc 1.2 (Figure 1a). However, this improvement cannot explain the increased seismicity since we observe increased seismic activity of all magnitudes. Between Oct 2001 and Sep 2002 we recorded over 3800 events, instead of the 1250 expected based on the b-value curve of the year before.
 In October 2002 over 900 events of magnitude below 3 were recorded, demonstrating a very high deformation rate and leading to fears of an imminent Katla eruption. In December that year the Icelandic National Civil Defence Authority requested a hazard assessment including the volcanic glacier flood threat. It was concluded that earthquake data were central in planning possible evacuations [Guðmundsson and Gylfason, 2005]. The report was made public in 2004 and was followed by a total evacuation drill for the area.
 An explanation for shallow lp events observed at volcanoes has usually been sought in terms of a volcanic source effect, commonly in terms of fluid motion [Chouet, 1996; Neuberg et al., 2000] or anomalous propagation through the complex structure of a volcano [Bean et al., 2008]. However, Métaxian et al.  report lp events at Cotopaxi volcano which they interpret as being of ice origin and propose a fluid-driven crack model. Recently, lp events have been associated with the motion of ice streams [Ekström et al., 2006; O'Neel and Pfeffer, 2007; Wiens et al., 2008; Tsai et al., 2008]. Ekström et al.  report much larger lp events (M ∼ 4.5–5.2) of very low frequency (0.01–0.04 Hz) and find that they are seasonally modulated similarly to the Godabunga earthquakes. O'Neel and Pfeffer  report calving events with extraordinary long seismic codas with frequencies focussed between 1–3 Hz and conclude that these features are a source effect which they model in terms of a fluid-driven crack. Roux et al.  report lp events (1–4 Hz) generated by free falling iceblocks in an Alpine glacier and refer to it as serac fall. Importantly, their observations have been confirmed by eyewitnesses.
2. Lp Events in Goðabunga
 We investigate the seismicity in Goðabunga using data from the permanent Icelandic seismological network [Böðvarsson et al., 1999] complemented by new seismological data collected in the spring of 2007 on 9 broadband Guralp (ESP3-compact, 60 sec.) instruments deployed on nunataks (bedrock) and one short band (Lennartz, 5 sec.) installed in the ice at western Mýrdalsjökull. Four of the instruments were installed as a mini-array (Figure 1b). Since the lp events are often emergent and the onset of different phases difficult to detect, the event locations were significantly improved by particle motion analysis and frequency-wave number analysis (Figure S1 in auxiliary material). The temporary network recorded 30 lp events of which 25 could be divided into four groups of similar waveforms and amplitudes (Figure 2). This observation suggests that events within individual groups have a common source.
 Our new locations of stacked events (Figure 1b) representing the four groups observed in spring 2007 point however to a very different origin than has been suggested before; a steep and partly discontinuous outlet glacier, Tungnakvíslarjökull, where a steep escarpment creates an ice fall that (dry) calves 80 m thick glacier ice in a fall of approx 100 m. (Figure S2). We believe this is our seismic source. The different groups probably represent falls at slightly different places introducing variations in the source (variations in ice thickness, fall height etc.) as well as the propagation path.
 Generally the waveforms have an unclear onset followed by low amplitude coda of 0.3–8 seconds when a higher amplitude P-phase arrives (Figure 2). The shallow origin is confirmed by following surface waves.
 Signals from large lp events were deconvolved using the signals small lp events (with a hypocenter close to the large event) as an approximation of the Green's function. The resulting source time functions (stf) (Figure S3) are consistent in character with a time extended source (7–9 sec.), indicating that the unusual low frequencies as well as the extended coda are indeed a source phenomena although path effects as well as reverberations from the glacier are likely to contribute.
 This observation is consistent for different instruments, in different settings (one station is deployed in the ice), and at different azimuths, supporting the conclusion that the obtained stf is reliable.
3. Seasonal Seismicity, Environmental Data, and Glacial Model
 The temporal distribution of the Goðabunga seismicity shows other peculiarities. The activity is highly seasonal and largely lacking in aftershock sequences [Jónsdóttir et al., 2006, 2007]. The seismicity within the caldera and Goðabunga both show annual cycles of increased activity in the autumn, but the behaviour of the two clusters differs significantly.
 The less pronounced annual variations within the caldera are probably caused by water influx from the summer melt increasing pore-fluid pressure in the bedrock, reducing effective normal stress and allowing shear stress to cause slip on fractures in the seismogenic crust. The pore pressure diffusion time is consistent with the annual seismicity within the caldera [Jónsdóttir et al., 2007].
 However, this mechanism cannot explain the shallow lp-seismicity at Goðabunga which reaches its maximum two and a half months later in the autumn [Jónsdóttir et al., 2007]. Annual glacier unloading above a pressure sensitive volume, i.e. a magma chamber, at shallow depths is not likely to explain the Goðabunga data since the elastic response is instantaneous and the inferred ice load has a minimum one month before the Goðabunga seismicity peak [Jónsdóttir et al., 2007]. In addition there is neither geothermal melting beneath nor geothermal drainage via Tungnakvislarjökull [Scharrer et al., 2008; Guðmundsson et al., 2007].
 Both seismic and climatic data show a change in character from the year 2000 (Figure 3). While the number of earthquakes, temperature and rain (seen clearly in the river flow data) increased, the glacier net mass balance, which can be estimated from the long term trend of the snow budget index (estimated according to Sigurðsson and Jónsson ) decreased [Jónsdóttir et al., 2007] (Figure 3). The earthquake activity decreased again after year 2004, as did maximum earthquake magnitude and the temperature and rainfall. Clearly, the earthquake record is strongly correlated with weather data, implying that long term changes in earthquake activity may be related to changes in local climate.
 At Mýrdalsjökull glacier the input of melt water from the summer melt peaks in August (see river flow in Figure 3). Since the outlet glaciers of western Mýrdalsjökull are very steep, it is likely that the glacial drainage system collapses in September/October when the melt water discharge diminishes [Röthlisberger, 1972]. This reduced permeability implies that the late autumn rain surges will increase basal water pressure and thus reduce bed friction and enhance ice movements [Röthlisberger, 1972; Bartholomaus et al., 2008]. Electrical conductivity data from Gígjökull glacial river also correlates well with earthquake rate (Figure 3). Similar annual conductivity variations of glacial-fed rivers in Iceland have been interpreted as being caused by varying ion content due to changing interactions with the bedrock material beneath the glacier as the hydrological situation changes over the year [Kristmannsdóttir et al., 2002]. Thus, high conductivity means that water flow is slow and vice versa. We conclude that the raised conductivity in October suggests that water flow is slow, i.e. we have a linked water-pocket drainage system at the base of the ice sheet, enhancing glacier motion in terms of basal sliding.
 During winter, precipitation as snow rather than rain reduces water inflow and thus water pressure, implying less movement. The Goðabunga seismicity record correlates strongly (R. = 0.6) with rainfall information, from river flow data as well as nearby weather stations, with the peak of the cross correlation function indicating that seismicity lags rainfall by ∼10 days.
 We suggest that the lp events occur when we have an ice fall event. These events occur throughout the year. However, more events are observed when the ice motion is faster. The rainy season in October induces basal sliding of the outlet glaciers thus speeding up their motion and in our case more ice blocks fall of the cliff. All of the data in Figure 3 is consistent with this model.
 A natural question is whether the energy from the ice movements is sufficient to generate the large and repeated seismic signals. The largest events (M 2.5) imply seismic energy of 0.35 GJ [Gutenberg and Richter, 1956]. A realistic example of a big ice slice, 100 m × 50 m × 1 m and of density 910 kgm−3, falling down the 100 m high escarpment will release up to 4.5 GJ of energy, some as seismic waves (ice thickness from Björnsson et al. ). This is sufficient energy to explain the largest events. An estimation of annual ice velocity suggests that enough ice flows over the escarpment to explain the data (see glacial velocity in auxiliary material).
 Our model, including icefall episodes of limited magnitude controlled by glacier thickness and height of which the ice blocks fall, implies significant deviation from the normally observed earthquake magnitude-frequency behaviour. This is confirmed in our data since the record has a clear upper limit of magnitude and the resulting b-value curve is by no means a straight line, instead it is convex throughout most of the magnitude range [Jónsdóttir et al., 2007].
 The non-volcanic origin of the Goðabunga seismicity is corroborated by recent geodetic measurements of Mýrdalsjökull with GPS (measured on nunataks) and InSAR [Sturkell et al., 2008; Árnadóttir et al., 2009; A. Hooper and R. Pedersen, Deformation due to magma movement and ice unloading at Katla volcano, Iceland, detected by persistent scatterer InSAR, paper presented at ENVISAT Symposium, Montreux, Switzerland, 2007], which indicate that ongoing vertical deformation can be explained by isostatic adjustment due to the observed melting of the glacier. The observed local horizontal GPS rates are, however, higher than expected from glacial rebound which has been interpreted as demonstrating a very shallow magma intrusion into the north eastern part of the Katla caldera [Sturkell et al., 2008]. In Goðabunga, however, there is no geodetic evidence for magmatic activity [Sturkell et al., 2008; Árnadóttir et al., 2009; Hooper and Pedersen, presented paper, 2007]. Since the Goðabunga seismicity started to increase there has been no visual eruption, and neither jökulhlaups nor significant changes in the ice cauldron depths (manifestations of subglacial geothermal activity) have been observed [Guðmundsson et al., 2007].
 We conclude that the lp-seismicity recorded in Goðabunga is caused by glacial movements rather than, as previously presumed [Sturkell et al., 2008; Soosalu et al., 2006; Einarsson et al., 2005], a volcanic intrusion or other direct volcanic effects for the following reasons: (1) The seismicity has been continuous for decades and despite the very high deformation rate, manifested in the extreme number of events, there has been no sign of volcanic activity coinciding with the hypocentral locations. (2) Neither InSAR nor GPS measurements reveal any uplift in the area for the last decade that cannot be explained by glacial rebound. (3) The earthquake catalogue for western Katla lacks volcano tectonic events, which can be expected to accompany volcanic intrusion episodes. (4) The seismicity is seasonal pointing to an outside controlling factor. (5) The seismicity rate correlates well with rain and periods of distributed subglacial water channels enhancing glacial motion. (6) Our results of an extended source time function as well as energy models, magnitude and temporal distributions [Jónsdóttir et al., 2007] support our hypothesis. (7) New hypocentral locations coincide with a dramatic ice fall. (8) Far field radiation pattern shows two lobes consistent with a single force source mechanism (Figure S4). (9) Similar ice fall events have been shown to generate low frequency, time extended waveforms [Roux et al., 2008]. While these results are specific for Katla, we suggest that similar phenomena may be relevant elsewhere, and against the background of climate change this possibility should be included in future assessments of volcanic hazards.
 We thank the reviewers for their helpful comments. Thanks to J. Hólmjárn, S. S. Jakobsdóttir, and P. Halldórsson, who assisted with network GODA 2007; Benni and Rína at Arcanum and the guides at Básar for assistance and discussions; H. Björnsson for providing rain and temperature data; Á. Snorrason and Ó. Þórarinsson for the delivery of glacial river data; and H. Björnsson and F. Pálsson for providing us with the surface and bedrock topography of western Mýrdalsjökull.