Svalbard is a Norwegian archipelago located in the high Arctic, between 74 and 81° N latitude, and 10 and 34° E longitude. About 57 % of the archipelago is covered by ice; the Randolph Glacier Inventory version 7 RGI; inventoried 1666 glaciers on the archipelago, among which 196 terminate in the sea, all of which seem to be grounded . Many Svalbard glaciers have a history of surging, meaning that their flow velocities can dramatically increase for a short period of time (typically 1–10 years). The climate, basal rheology, and glacier geometry of typical Svalbard glaciers seem especially favorable for surging, and the archipelago is one of the world's hotspots for surge activity . Among Svalbard tidewater glaciers, 139 are of surge-type .

Globally, glaciers are an important component of the climate system. They are a freshwater storage, and their mass loss can dramatically influence sea-level rise (SLR). Worldwide, glaciers are expected to lose 26 %–41 % (for mean temperature increase projections of +1.5 to +4 °C, respectively) of their mass by 2100, compared to 2015 . In the Arctic, glacier mass loss contributes to changing the surface albedo, which amplifies global warming through positive local and regional feedbacks . A better understanding of Svalbard glacier change is therefore crucial to assess their future evolution and their impact on the climate system.

Several aerial photograph mapping campaigns have been carried out over Svalbard in the past century. used historical photographs and written observations to produce an inventory of past glacier surges, and identified possible future surges. took advantage of aerial images from 1936–1938 to generate a DEM of the archipelago, used to investigate historical mass balance changes of Svalbard glaciers in a 70-year time span. They discovered a strong correlation between the temperature and the mass loss rate: a 1 °C increase in mean summer temperature corresponds to an area-normalized mass balance of -0.28mw.e.yr-1. They also predict a doubling of mass loss by 2100, using a space-for-time substitution. compared a variety of DEMs produced from different aerial campaigns to calculate Svalbard glacier elevation changes and contribution to SLR between the 1960s and the 2010s. They calculated a volume change of -0.36±0.02mw.e.yr-1, corresponding to a SLR of 0.026mmyr-1.

Since 1995 and the declassification of 860 000 US military intelligence satellite photographs from the 1960s and 1970s, it has been possible to extend the time series of satellite remotely sensed glacier data. KH-4, 4A and 4B CORONA (cf. Sect. ) photographs have been used for historical studies, especially in archaeology and more recently in glaciology . used these images along with other sources of historical data for the study of surface mass balance (SMB) and frontal advance/retreat of outlet glaciers in the eastern margin of the Greenland ice sheet and their responses to air and ocean temperature changes. used orthocorrected CORONA stereo-images to produce digital elevation models (DEMs) of the Mansalu region of Nepal, and compared them to recent sources of data to study glacier SMB.

KH-5 ARGON (cf. Sect. ) images have also been used for historical studies. used KH-5 ARGON images to map the ice shelf margins along Queen Maud Land, Antarctica, and compared them to more recent observations. created an orthocorrected version of ARGON images over the Larsen Ice Shelf (Antarctica), and discovered that the acceleration that led to the collapse of Larsen B happened earlier than previously thought. used these images to produce a regional DEM of the Antarctic ice sheet. There is no published study using declassified intelligence satellite photographs (DISP) on Svalbard, despite the availability of several CORONA and ARGON images.

The MultiSpectral Scanner (MSS) instrument equipped Landsat 1–5, and was in use between 1972–1999 (and was briefly turned back on in 2012–2013), enabling retrieval of glacier area changes. For example, successfully used two Landsat MSS images to measure the area of Iceland's largest ice cap (Vatnajökull) and to inventory 38 of its outlet glaciers. used Landsat MSS, along with other types of multispectral satellite data, to calculate the area change of glaciers in Turkey. Landsat MSS images have also been used in Svalbard for glacier studies. used Landsat MSS images to map the ice cap margins of Nordaustlandet and compared them to maps from that time. He discovered that revisions of these maps were necessary, particularly on the East coast and at certain outlet glacier margins. He also mapped the snow line position and the basal melting through the observation of sediment plumes offshore. finally, used Landsat 3 images, among other sources of data, to produce an inventory of Svalbard glaciers and report all known surge events in Svalbard. They documented the type, area, activity, and geometry of every Svalbard glacier larger than 1km2.

Automated glacier mapping techniques using band ratios () and the normalized difference snow index NDSI; are not applicable to Landsat MSS images because of the lack of a SWIR band, which therefore need to be processed manually. Deep-learning approaches allow overcoming this limitation e.g.. used Landsat, Terra-ASTER and Sentinel-2A/B multispectral images, as well as Sentinel-1A/B images to train a deep learning model that produced a dataset of 124 919 glacier calving fronts for 149 marine-terminating glaciers in Svalbard between 1985–2023. analyzed this dataset and observed seasonal cycles in front positions, with a several-month lag between peak retreat on Spitsbergen East and West coasts. Their observations are consistent with regional ocean warming patterns. After removing the seasonal variability, they calculated marine-terminating glaciers have globally retreated during the study period, with a rate of 23.78km2yr-1. and provide a detailed historical monitoring of Svalbard glacier calving fronts. However, early Landsat images (Landsat 1–3) were not used in the production of the final dataset.

The objective of this study is to bridge the gaps between past observations of Svalbard glaciers, by retrieving glacier calving fronts from DISP (1962–1963) and Landsat (1976–1978) images, and comparing the observations to historical aerial image calving front data 1936–1938,. In particular, this is the first study using DISP over Svalbard. It aims at comparing the front positions to assess their cumulative retreat or advance between the late 1930s and 1970s, which is a good indicator of glacier area change; and to document possible surge events that occurred during this period.

In 1936 and 1938, two aerial expeditions were led by the Norwegian geologist Adolf Hoel over Svalbard. During these expeditions, 5507 images were acquired, in order to produce topographic maps of Svalbard. These images have been used more recently to monitor elevation changes on Svalbard glaciers. processed these images with a modern photogrammetric workflow, and produced an orthorectified mosaic (Fig. ) and a DEM. They also produced an updated 1936–1938 glacier outline inventory, which is used in this study together with the new 1960s and 1970s measurements.

Landsat is an ongoing scientific Earth-observation program that started in 1972. Instruments equipped on the Landsat satellites have evolved over time to provide observations in more spectral bands and with higher spatial resolution. In this study, we focus on the early stages of the Landsat program, when the MSS was the main instrument onboard. provides an overview of the Landsat MSS data products. We identified sixteen usable Landsat 2 and 3 (Level 1 Geometric-Systematic, L1GS) images of Svalbard from 1976 and 1978. Each image covers a 180km×180km area with a ground sample distance of 80 m, re-projected to a 60 m pixel size. These images are freely available on the U.S. Geological Survey (USGS) EarthExplorer website (https://earthexplorer.usgs.gov/, last access: 8 July 2026) and can be downloaded in GeoTIFF format. Figure  gives an example of two of these images.

We produced false color images from bands 7, 5, and 4 using the rioxarray Python package. The original geolocation was off by multiple kilometers, which required correction. To do this, we selected about 20 GCPs for each image (cf. Table ) in QGIS using the Georeferencer tool and applied a 2nd degree polynomial transformation (cf. Appendix  for GCP maps and RMSE).

Once we had georeferenced the KH and Landsat images, we digitized tidewater glacier fronts. Given the small number of usable images, a single operator carried out this process manually in QGIS. Therefore, the accuracy of the front positions is, at best, the image resolution (60–140 m). On several Landsat images and on 1962 KH images, sea ice is present in the fjords, resulting in a lower accuracy. We delineated the fronts on every image on which they were visible (Figs.  and , right panel). Therefore, for several glaciers, multiple mappings from the same year are available and can be used to assess the positioning uncertainty. In total, we made 467 Landsat and 526 KH observations, for 185 calving fronts on 171 glaciers. For the aerial photographs, had already digitized glacier outlines, and we simplified the shapefile containing the glacier polygons to extract only the tidewater glacier fronts. 169 front observations are included in this dataset (Fig. , right panel), covering all Svalbard tidewater glaciers, except the easternmost part of Nordaustlandet. Figure  shows the extent of the complete dataset.

We then enriched the datasets with the image names, dates, missions, RGI (a collaborative global glacier outlines dataset, designed to represent the glacier extents around year 2000, based on various sources of data [airborne, satellite, etc.]) index versions 6 and 7, and flow directions to facilitate data analysis. Since some glaciers in the dataset have several calving fronts (e.g. Rijpbreen, cf. Fig. ), we created an index to precisely identify each front. In some places, several glaciers merge to form a unique front. In these cases, we split front observations so that one portion of the front is associated with each glacier, following the RGI and 1936–1938 partitions. An example is given in Hornsund (Fig. ). We used the regular WGS84 ellipsoid projection for the Landsat and KH fronts, while we kept the original CRS (EPSG:32633, WGS84 – UTM zone 33N) for the aerial image fronts.

This study focuses on quantifying glacier frontal changes. Since calving fronts may have a complex geometry and may not move uniformly, we focus on reducing the problem to quantifying their position change along the glacier flow axis. We used the extrapolated centerline method, introduced by , as an improvement of the centerline method. In the regular centerline method, the distance between two front observations is simply calculated as the distance between the intersection of the fronts and the glacier centerline, resulting in a loss of all the information about the front geometry. The extrapolated centerline method uses inverse distance weighting to define the position of every front point on the glacier centerline. The final position of the front is then defined as the mean value of the positions of its points cf.for implementation details; the method of this study is identical to theirs. We digitized centerlines by hand, due to the lack of concomitant topography and flow information. We compared the results obtained through this method to those calculated with the box method when possible. We found a strong agreement between both methods. Since the extrapolated centerline method is designed to improve tracking of glaciers with changes in front width and non-rectilinear flows , we chose it to calculate all the results.

We retrieved area change by multiplying glacier front displacements (calculated with the previously presented methods) and the average lengths of the fronts . In this study, we calculated cumulative area change by adding the contributions of all glaciers together.

For the fronts that were obtained from KH-5 ARGON and Landsat MSS images, we assessed the positioning uncertainty by comparing fronts that were digitized from images dating from the same year. We consider this method to be reliable as Landsat MSS images were all taken during summer, and KH-5 ARGON images were taken in series on the same day. Therefore, the true position of a front should be the same on two photographs from the same year. identified three marine-terminating glaciers that were possibly in active surge phase when the Landsat images were taken: Allfarvegen (1970–1980), Rijpbreen (1938–1992) and Bodleybreen (1976–1980). We therefore excluded these glaciers from the Landsat uncertainty calculations. The typical positioning uncertainties for each year are summarized in Table . We also recombined these values to calculate uncertainties on front displacements and on cumulative area change. Since the qualities of the georeferencing and the front delineation depend mainly on the image quality, we calculated one value for each uncertainty type per year or possible combination of years (cf. Appendix  for calculation details).

RMSE values calculated for GCP positions are provided in Appendix . These RMSE values are high: up to 389 m for KH-5 ARGON (frame 1962-05-16-b) and 200 m for Landsat (frame 1976-07-18-b). However, these values are lower than those we obtained with our uncertainty assessment based on front comparisons (Table ). Indeed, our uncertainty assessment accounts for georeferencing errors, delineation precision and human mistakes at glacier positions. Therefore, our uncertainties are conservatively bigger than these RMSE values. Moreover, since all fronts were digitized within the bounds of the GCPs, the risk of extensive extrapolation due to polynomial georeferencing transformations is limited. If unidentified systematic biases exist, they are included in our uncertainty assessment.

High positioning uncertainties and RMSE may lead to inaccurate front displacement calculations, which impact both average frontal change assessment and individual advance/retreat identification. To mitigate this risk, we calculated uncertainties on frontal changes (cf. Sect. ) and filtered all retreats/advances that were not in the 95 % interval (cf. Sect. ).

Figure a displays time series of front positions (corrected to set their values to zero in 1936–1938). The mean glacier front position is also shown, as well as its associated uncertainties. Using a weighted least squares (WLS) method, we calculated that Svalbard tidewater glaciers retreated by 27.3±3.0myr-1 between 1936–1978. This average glacier linear retreat represents 1140±136m in 1978, compared to 1936–1938. The graphs show a faster retreat rate between 1936–1938 and 1962–1963, followed by a plateau until 1976–1978. The retreat rate seems to increase again between 1976–1978. However, because of the lack of data points between 1936–1938 and 1962–1963 and the overall high uncertainties (especially in 1962–1963), our observations could also reflect a linear retreat throughout the entire study period. Therefore, differences between 1936–1962 and 1963–1978 are difficult to assess robustly. One can also see the diversity of behaviors of Svalbard tidewater glaciers (gray lines on Fig. a): most are retreating, but some are advancing, and they have a wide variety of retreat/advance rates.

Figure b shows the cumulative area change due to calving front displacements for all measured glaciers between 1936–1978. The area loss between 1936–1978 was 642±167km2, with an average area loss rate of 17.0±2.5km2yr-1.

To identify significant retreat and advance events, we compared all front observations together, for each glacier. Then, we identified pairs of fronts whose distances were higher than the recombined uncertainties presented in Appendix . These events correspond to front advances or retreats with at least a 95 % confidence. Figure  shows the distance distributions and highlights the significant events. Table  in the Appendix summarizes the number of observations used for each distribution, and shows their mean values and standard deviations. Several observations can be made about these distributions. First, the mean values are all negative, meaning that on average, the fronts are observed to retreat. Second, dispersions are higher when the time span between front observations is greater. Finally, more significant observations are made when the uncertainties on front positions for both years are low, and when the time span between the observations is long.

We identify a total of 15 advances and 122 retreats between 1936–1978. A 16th advance of Storisstraumen was detected between 1976–1978, but this is likely due to the georeferencing uncertainty of the Landsat images. The uncertainty is high in this region due to the lack of stable terrain to be used as georeference points. Therefore, this glacier has been removed from the results. On Fig. , maps for all filtered advance events are displayed. One can observe on these maps that a specific front can show both advance and retreat behaviors. For instance, Hilopenbreen retreated between 1938 and 1962–1963 and then advanced between 1962–1963 and 1976–1978, and Freemanbreen advanced between 1936–1963 and then retreated between 1963–1976. Also, an advance/retreat might not be uniform along the front (e.g. Stonebreen), which we bring up further in the discussion.

Because of the high uncertainties (especially in 1962–1963), it is necessary to calculate the distance between fronts for every possible pair of years, and not only between consecutive observations. For example, a visual inspection of the Recherchebreen's fronts (Fig. ) supports an advance of the front between 1936–1962. However, the uncertainties on the 1962 delineations are much higher compared to those of 1978. Therefore, the distance between the 1936 and 1962 fronts is not significant to conclude to an advance, whereas the distance between the 1936 and 1978 fronts is lower but significant when compared to the displacement uncertainty. Thus, the advance is only detected between 1936–1978.

We calculated an average retreat rate of 27.3±3.0myr-1 for the 171 tracked glaciers between 1936–1978. found a wide variety of front change rates for 1985–2023, although most are within the range of 25–50myr-1, which is of the same order of magnitude as the one we calculated for a prior time frame. We found a cumulative area loss rate of 17.0±2.5km2yr-1 for all the tracked glaciers between 1936–1978, which is consistent, though lower, compared to the value of 23.78km2yr-1 calculated by for the 1985–2023 period, and the value of 26.79±0.54km2yr-1 calculated by for 2000–2020. This could indicate an acceleration of the area loss of Svalbard marine-terminating glaciers due to calving front retreats over the past century.

Figure  displays the time series of 2-meter temperature (T2m), sea surface temperature (SST) and sea ice concentration (SIC) from the ERA5 reanalysis , yearly averaged over Svalbard (74–81° N, 7–35° E) between 1940–2023. The graph clearly shows an overall increase in T2m and SST, and a decrease in SIC over the whole period. In fact, the average T2m is -6.42 °C for 1940–1978 and -4.05 °C for 1979–2023, the average SST is 0.42 °C for 1940–1978 and 1.23 °C for 1979–2023, and the SIC is 41.93 % for 1953–1978 (no SIC data is available prior to 1953) and 25.58 % for 1979–2023. Therefore, the evolutions of all three variables support an acceleration of the calving fronts' retreat after 1978 and are consistent with our results.

However, several differences in the methods to calculate these numbers remain: like , we calculate the change in area as the product of the retreat length and the glacier width, but we keep surge-type glaciers in the dataset for this calculation, and do not take into account the seasonal variability. The number given by includes Svalbard and Jan Mayen, and surge-type glaciers are kept in this dataset. However, they manually digitized the outline of the area that changed from glacier-covered to not glacier-covered, instead of calculating it as a product of retreats' lengths and glacier's widths. assessed the possible error sources arising from the comparison of their results with those of . They concluded that it remained valid to assess a large scale trend rather than to give a precise estimate. This discussion being limited to the former, we believe the comparison remains valid.

We detected a total of 13 glacier surges and 2 other advances (cf. Sect. ). Table  lists all the surge/advance events we detected, and reports the studies that mention them, as well as the time windows they identified. For two glaciers (Allfarvegen and Stonebreen, cf. Sect. ), we narrowed the time window in which the surge was known to happen, and we discovered one new undocumented surge of Emmabreen (cf. Sect. ).

In this study, we used archived satellite imagery to determine glacier calving front positions on Svalbard. However, it is likely that the dataset also contains geophysical information that is outside the scope of our study. A natural example of such information is the extent of land-terminating glaciers. It is possible that the images show previously undocumented surges of land-terminating glaciers, or at least can be used to better constrain the times of past surges in the same manner as we have done in our study for marine terminating ones. To facilitate future research, we have made all the data available, including the georeferenced satellite images. Furthermore, this study has shown that the KH-5 ARGON images are useful for the study of the Svalbard glaciers. Thus, we suggest that their use for other remote geographic areas be explored.