Glacier outlines were mapped manually using Geographical Information System programs (ArcGIS, QGIS). The AGI3 outlines and, where available, consistent intermediate regional inventories (Table ) were used as a starting point for the new outlines. The primary data type used for the mapping process was high-resolution regional orthoimagery (spatial resolution: 10–25 cm). Where available, high-resolution DEMs (0.5–1 m) and derived products (hillshades from different illumination angles, elevation change rasters) were used as supporting information or instead of orthophotos depending on data coverage. The georeferenced aerial imagery and DEMs were accessed through governmental open data services. The DEMs are derived from airborne laserscanning surveys and provided by Austrian regional authorities. High-resolution satellite imagery (Pleiades images, 0.5 m; Planet Scope images, 3 m ) was used in some cases to compensate for image quality issues in the orthophotos related to snow cover or shading. Table S1 in the supplement provides an overview of the main data sources and acquisition years per subregion. Fig.  shows the different available data types for the examples of Großelend Kees in the Ankogel-Hochalmspitze Group and Nördlicher Schalf Ferner in the Ötztal Alps. Großelend Kees (Fig. a, b) was mapped using mainly information derived from a 2023 DEM, whereas Nördlicher Schalf Ferner (Fig. c, d) was mapped based on multi-temporal orthoimagery.

Data availability and acquisition years vary between subregions and Austrian federal provinces. The target year for AGI5 glacier outlines was 2023. Data acquired in 2021 or 2022 was used if no suitable coverage for 2023 was available. Outlines for 2023 were produced for 60 % of glaciers in Austria (55 % of total glacier area). Outlines for 2022 were produced for 39 % of glaciers (43 % of total area). The remaining 1 % of glaciers (2 % of area) were mapped with data acquired in 2021. For the glaciers in the federal province of Vorarlberg (Fig. ), outlines for 2023 produced for a province-level inventory were incorporated into AGI5.

About 95 % of the total glacier area in Austria was mapped using mainly aerial images. The remaining area was mapped mainly based on DEM derived products (approx. 4 % of area), or high resolution satellite imagery (1 % of area), typically with orthophotos as auxiliary data. In most cases, analysts reported using more than one data type. This mainly refers to using hillshades or elevation change rasters in combination with aerial or satellite imagery, using satellite imagery instead of orthophotos if snow conditions were more favourable in the former, or using imagery from multiple years to assess glacier evolution. As a general guideline and following the approach of AGI3, analysts were asked to use a maximum zoom scale of 1:3000 in both aerial imagery and DEM derivatives as a starting point for mapping. Most analysts mentioned using a higher zoom level for very small features and cases they considered challenging.

Information from field surveys of the glacier margins and relevant local knowledge were incorporated at the discretion of the analysts. Such terrain knowledge was available mainly for glaciers with in situ monitoring programs and a subregion of the Ötztal Alps where an analyst documented debris-covered ice outside of older AGI outlines during site visits over multiple years. In some cases, analysts found that the imagery and data used to delineate AGI5 outlines showed ice where none was mapped in prior outlines due to varying image quality, snow cover, and/or differences in interpretation between analysts. The prior outlines were not modified in these instances. The AGI5 mapping process also revealed several errors in prior inventory attributes, mainly related to wrongly assigned or mistyped ID numbers. Corrections for these cases were issued as updates to the existing AGI3 data publication .

For each glacier ID and corresponding outline, the AGI5 attribute tables (Table ) list information related to the source data (acquisition date, data type, image identifiers). Categorical flags indicate potentially detrimental image characteristics (e.g., snow cover or shadows), a qualitative level of uncertainty of the outline, and a debris cover score. In addition, the presence of visible crevasses was flagged. Crevasses have been used as an indicator of past or present flow e.g., and, hence, a criterion that allows distinguishing “glaciers” from “ice bodies” as defined by . AGI5 does not make such a distinction but including the crevasse flag allows users to filter the data accordingly, for example when comparing AGI5 with datasets that distinguish between these classes, or only include glaciers that show signs of flow (e.g., the SGI). The attribute table also includes a flag indicating overlap of AGI5 with the outlines of the most recent Austrian Rock Glacier Inventory RoGI, to identify potential cases of misclassification or ambiguous landforms.

The AGI5 attribute table further contains the area of each glacier, the number of fragments per glacier, and elevation statistics. Minimum, maximum, and median elevation of all pixels within a given outline were extracted from a 10 m × 10 m resolution DEM provided by the Austrian Federal Office of Metrology and Surveying (BEV). This DEM is a resampled product from original high-resolution DEM data (0.5–1 m) derived from airborne laserscanning produced for the individual provinces of Austria. It includes data from different undisclosed epochs around 2015 and is available through the Austrian open government data platform . Accordingly, the national DEM does not exactly match the outline years in AGI5. It was chosen as the most recent product available across all of Austria to derive consistent per-glacier elevation statistics.

In keeping with the previous AGI, no minimum size threshold for glaciers or fragments of glaciers was applied in AGI5. Analysts were asked to map the remaining glacial ice patches in their subregions to the best of their abilities regardless of feature size. Experience with recent regional inventories showed that very small, highly debris-covered features can be difficult to map even with very high-resolution imagery and auxiliary DEM-derived information . To account for this, analysts had the option to flag glaciers or glacier fragments that they considered impossible to map but assumed may still contain ice as “vanishing features”.

Two possibilities for further handling of such “vanishing features” were incorporated in the processing workflow:

Case 1: A single glacier ID was associated only with “vanishing features”. That is, there were no longer any “mappable” fragments of a given glacier but it was assumed that ice may still be present, for example under debris cover. In this case, the respective glacier was considered a “vanishing glacier” close to disappearance. As the outlines of these features are not mappable with the given source data, they were not included in the main inventory dataset. Centroid coordinates were recorded in a separate file.

The AGI5 region contains nine ski resorts located in glacierized terrain. The resorts apply white reflective coverings to locally reduce melt and preserve particular patches of ice and snow relevant to resort operations (for example, to maintain lift tracks or for snow storage). Such patches of covered ice and snow were included in the outline mapping if they were connected to a remaining uncovered feature. Covered patches not connected to uncovered ice were not included. Figure a and b show examples of coverings and snow management used to locally reduce ablation on Wurten Kees, Carinthia. Figure c and d highlight a narrow connection between a covered remnant of the tongue of Tiefenbach Ferner, Tyrol, and the main glacier.

We maintained the AGI3 ice divides in AGI5 to enable consistent area change assessments and due to the challenges associated with accurately determining the current location of the divides. The ice divides were initially defined for AGI1 and largely applied in the same way in AGI2 and AGI3. We note that the true location of the ice divides may have shifted since AGI1 due to ice losses.

Some Austrian glaciers border on neighboring glaciers in Italy. Country borders along the main chain of the Alps are generally defined based on drainage divides. That is, the country border mostly follows the ice divides in the glacierized regions of the Ötztal, Stubai, and Zillertal Alps. The main exceptions to this are Hochjoch Ferner and Niederjoch Ferner in the Ötztal Alps. Here, the country border locally deviates from the drainage divides and the glaciers extend across the border. For consistency with AGI3, we continue to include the Italian parts of the respective glaciers in AGI5.

In general terms, we follow an epistemic approach to uncertainties. That is, “the best value is that which is most credible, from the current perspective of knowledge” . Uncertainty estimates can be interpreted as a measure of the extent that our knowledge of glacier area – at the glacier-level and regionally – remains inexact. Considering the RR experiment, we assume the most credible value for each glacier is the median of the sample. However, the RR approach is a measure of precision rather than accuracy in the sense that we do not know the “true” area value. Rather, we are assessing how closely the analysts’ interpretations of the source data are aligned (see Sect. ).

The AGI5 uncertainties should be understood as fallible epistemic products that can and should be iteratively improved . Notably, improvement does not necessarily mean a decrease in uncertainty. The AGI5 uncertainties provide an estimate of potential errors introduced by differences in interpretation between analysts, but they do not account for potential systematic biases, for example related to debris cover identification, and are hence fallible for glaciers that may be affected by such biases. The AGI5 uncertainty estimates build upon prior work, adapting the approaches of and , and have been iteratively improved by incorporating feedback from analysts, who are arguably in the best position to estimate how inexact their outlines may be for a given glacier. The uncertainty estimates could be further improved in the future by explicitly including additional sources of uncertainty as related knowledge improves, for example regarding quantitative bias-estimates for debris cover identification.

The “outline quality” and debris attributes assigned by the analysts (Table ), indicate that about a third (259) of all glaciers in the study area have high outline uncertainties due to debris cover or otherwise obscured ice margins. Although the attributes are also affected by subjective interpretations (e.g., partially vs. mostly debris-covered), this clearly reflects the most common challenges in manually mapping glacier outlines from orthoimagery and is aligned with expectations given the numerous very small glaciers in Austria. Findings by other studies similarly indicate that surface conditions, timing of image acquisition, quality and resolution strongly affect the visibility and delineation of glacier margins e.g.,.

Discrepancies in manual mapping between different analysts (or the same analyst mapping a glacier multiple times) are to be expected and it is assumed that they cannot be avoided completely. RR experiments are a standard approach to estimating the general magnitude of the uncertainties introduced by such discrepancies e.g.,. Measures of variability in the area derived from RR outlines are typically used as uncertainty metrics. For example, the 2010 Swiss Glacier Inventory (SGI 2010, ) found uncertainties of ±7.6 % for glaciers smaller than 1 km² and uncertainties between ±3 % and ±5 % for larger glaciers. For SGI 2016, reported a standard deviation of glacier area between multiple analysts of 23.8 % for a “very small glacier in a shadowed, snow-covered north face” and values between 0.3 % and 7.1 % in other cases. In their glacier inventory for Vorarlberg, found uncertainties of over ±50 % in very challenging (small, mostly debris-covered) cases, which is generally in line with studies assessing outlines of debris-covered or otherwise challenging glaciers mapped from satellite imagery e.g.,.

Once uncertainty estimates for individual glaciers are determined (often derived from RR comparisons), they are frequently used to estimate region-wide uncertainties in a second step. For example, applied an area weighted average uncertainty derived from the RR experiment to obtain the total regional uncertainty of SGI 2010. For AGI3, , applied relative area uncertainty estimates to individual glaciers (following , see Sect. ) and summed uncertainties to produce a region-wide uncertainty value.

The regional uncertainties given for AGI3 and SGI2016 are in the range of 2 %–3 % of glacier area in Austria and Switzerland, respectively (Austria: 415.11 ± 11.18 km², Switzerland: 961 ± 22 km², as per the above studies). Uncertainty estimates for AGI5 are slightly higher with ±4.2 % (285.5 ± 12.0 km²). This increase compared to AGI3 can be explained with the higher relative uncertainties applied to highly debris-covered or otherwise very uncertain glaciers. In AGI5, the glaciers with the highest assumed uncertainties account for about 5 % of total area.

The buffer approach derived from the RR outlines yields uncertainty estimates comparable to the uncertainties computed from the tiered system based on glacier size and outline quality scores (Sect. ), with higher values depending on how debris cover is treated. Outlines of clean ice margins in our RR experiment are generally aligned within ±2 m of each other. Extrapolating this to the AGI5 region yields a regional area uncertainty of around ±2 % (i.e., similar values to AGI3 and SGI2016). In debris-covered sections of the glacier margins, relative uncertainties derived from the buffer approach reach up to around ±40 % (i.e., similar values as in Table and as found by , and , for challenging cases).

The results of the AGI5 RR experiment (Fig. , Table ) are broadly in line with the uncertainty magnitudes reported by the studies discussed above and confirm that uncertainties increase with debris cover and decreasing glacier size. Generalizing the comments provided by the analysts regarding the RR experiment (Sect. S3 in the Supplement), differing interpretations of debris cover are the main source of discrepancies. Most of the 15 analysts incorporated geomorphological process understanding in their mapping approach and relied on visible surface features (e.g., changes in surface structure, presence of snow or water) that they considered indicative of subsurface ice to delineate debris-covered ice margins. Other analysts focused more on ice that could be visually identified in the imagery and incorporated less detailed process thinking. The RR experiment showed a general tendency towards smaller glacier area for the latter approach but this is not universally true for all test cases and analysts.

In some cases, analysts provided somewhat contrasting process-based interpretations (e.g., Arvental Kees), or had generally similar interpretations of the geomorphology that nonetheless lead to different conclusions regarding the outline or glacier status (e.g., Madlein Ferner, Sect. S3 in the Supplement). In practice, most of the AGI5 glacier outlines were checked by multiple people and analysts discussed challenging cases among each other. Some of the low outlier cases in the RR (e.g., lowest area estimate for Pasterze, “vanishing” status for Arvental Kees) would likely have been identified as anomalous and revised in such quality control settings. The high outlier cases are all related to interpretations of debris cover and are difficult to refute or verify. Improved mapping consistency might be achieved with agreed upon, detailed guidelines regarding identification of debris-covered areas. However, improved consistency does not necessarily mean improved accuracy, for example if the majority opinion is subject to unknown or unquantifiable systematic biases.

In discussions prior to the main AGI5 mapping effort, it was agreed that all available information, including relevant process understanding, should be used, as opposed to only mapping visible ice. The latter would likely have produced more consistent results in the RR (and the overall AGI outlines), but would have underestimated the debris-covered area. Similarly, setting an initial zoom-scale of 1:3000 was a measure intended to improve consistency and reduce effort spent on very small-scale details. Most analysts reported deviating from this scale to check for evidence of subsurface ice in challenging cases and considered closer zoom levels to result in improved accuracy. Aside from inter-analyst discrepancies, there is also some amount of variability in outlines mapped by the same analyst on different days (sometimes referred to as “digitization uncertainty”). In our experience, this variability is substantially lower than the inter-analyst variability and we assume it to be covered by our general uncertainty estimates (see also Sect. S3 in the Supplement).

As evidenced above, accurately mapping debris-covered ice from remote sensing is challenging even with very high-resolution imagery, and independent ground truth is rarely available e.g.,. Including local terrain knowledge (where available) in the compilation of AGI5 caused area change discrepancies at individual glaciers due to the greater inclusion of debris-covered area compared to AGI3 (Sect. S5 in the Supplement). However, we consider the approach more accurate and, as above, found this more important than maintaining “consistency” with the previous data by omitting the field observations. The same principle (accuracy over consistency with prior outlines) was applied when AGI5 imagery showed evidence of ice beyond the extent of AGI3 (e.g., Eiskar Ferner in the Carnic Alps, Table ; or Arvental Kees in the Venediger Group, Fig. ).

For the subset of glaciers for which detailed local knowledge was available in AGI5 (Sect. S5 in the Supplement), the comparison with outlines mapped exclusively from aerial imagery suggests a systematic underestimation of debris-covered area, similar to previous work in the same region . However, the magnitude of the bias is difficult to quantify at larger scales and it is unknown how often overestimations occur. The tendency towards underestimation of debris-covered area is in line with findings from neighboring South Tyrol (Alto Adige, Northern Italy), where reported an underestimation of glacier area by 2.3 % in their 2017 glacier inventory compared to a subsequent inventory for 2023. They suggest the underestimation is mainly related to debris cover and assume a similar underestimation is present in outlines produced with 2023 imagery. Given the comparable source data and mapping approaches, AGI5 underestimations of this type can be assumed to be of similar magnitudes. We note that the opposite scenario (overestimation of debris-covered area) may also occur but this is hard to verify and generally cannot be determined from optical imagery alone e.g.,.

Incorporating DEM-derived products or other auxiliary information in glacier outline mapping can support the identification of debris-covered ice margins e.g.,. Discussing specifically manual mapping with very high-resolution DEM-derived and optical information at the scale of individual mountain ranges (as applied in this study), and highlighted that surface elevation change and geometric information improve confidence in outline detection under debris cover. Inventory studies from Switzerland have similarly noted that surface elevation change information reduces debris-related uncertainties .

However, even with very high-resolution data (sub-meter pixels), challenges remain. Snow cover during DEM acquisition, avalanches, or rock fall can create locally positive elevation change signals even if glacier ice was lost over the same time period, making such data difficult to interpret especially for very small glaciers close to disappearance (examples in ). In AGI5, analysts used surface elevation change data where available, usually in addition to optical imagery. Feedback on this was mixed – some found the additional information helpful, but others noted that data were inconclusive in their regions, for example because it was not possible to distinguish loss of ground ice from loss of glacier ice, or due to ambiguities in areas where ice cover was lost completely during the elevation change epoch (i.e., an elevation change signal is present but no ice remains at the end of the epoch). In practical terms, on-site inspections of glacier margins would often be beneficial for accurate mapping but are not feasible at larger scales. Lake-terminating glaciers such as Pasterze (Sect. S3, Fig. S10 in the Supplement) can pose an additional mapping challenge if ice extends below the water level .

In summary, it seems likely that debris-covered area is systematically underestimated in glacier inventories compiled from optical imagery, including AGI5, but this remains hard to quantify exactly. Contextual information (surface elevation change, other data types) has the potential to improve delineation of debris-covered glacier margins but needs to be assessed within the specific geomorphological setting. Geophysical investigations would be needed to clearly identify the ice margins beneath thick debris cover.

In addition to differences in interpretation and approach between analysts, “classification errors” and “conceptual errors” are two main sources of uncertainties in remote sensing of glacier outlines . The former refers to “misidentified features” and the latter is used for a variety of challenges, e.g., “glacier definition issues such as ice divides, perennial snowfields, minimum size, and fragmentation” .

Applying these ideas to AGI5, the conceptual issues related to glacier definitions were addressed in the same way as in prior AGIs. That is, there is no explicit distinction between glaciers, ice patches or perennial snow patches; ice divides were maintained since AGI1; and glacier fragments retain their initial ID number.

The imagery and source data used for mapping of the AGI was largely acquired during favourable, mostly snow-free conditions. However, it cannot be ruled out that off-glacier seasonal snow was erroneously included in glacier outlines in some instances across the AGI time series. If these snow patches then disappear and are not included in the subsequent inventory, derived loss rates are overestimated. Due to the mostly good quality of imagery in AGI5 and based on the available information on source data of older AGI, we do not consider this to be a systematic issue in the dataset but acknowledge that the effect may be present in individual cases.

A further relevant classification error may be found in the distinction between glaciers, rock glaciers, and other periglacial landforms. Comparing the AGI5 outlines with the Austrian Rock Glacier inventory shows 29 cases where AGI5 and the rock glacier inventory outlines overlap. About half of these have only minimal overlap indicative of different landform types existing in close spatial proximity (i.e., the outlines “touch”). The remaining cases have more than 20 % area overlap and may represent classification errors of rock glaciers misidentified as glaciers or vice versa. The distinction between debris-covered glaciers and rock glaciers can be challenging, especially for landform-sequences where both occur in close proximity e.g.,. Nonetheless, the low number of overlapping cases between AGI5 and the rock glacier inventory suggests that this is a relatively minor error source for both data sets.

Whether a glacier can and should be classified as “vanished” or “vanishing” might be considered a classification issue as well as a conceptual one. Classification requires consistent inventories or other forms of record keeping that allow a comparison over time to determine if a glacier that was previously present has disappeared. A conceptual definition of a “vanished glacier” requires a common understanding of the term “glacier” that is consistent at least across the inventories used for the comparisons e.g.,.

In their recent compilation of vanished glaciers in Switzerland, and consider glaciers that were larger than 0.01 km² in the SGI of 1979 and were no longer included in the SGI2016, or the forthcoming SGI2023 . That is, they classify glaciers as ”vanished” if they no longer meet the size requirements for inclusion in the SGIs. In AGI5, we allow a ”fuzzy” classification as ”vanishing” for cases that may still contain ice but cannot reasonably be mapped with the AGI methodology. This approach was based on feedback by analysts, who indicated that uncertainties in mapping the outlines of very small, debris-covered features can effectively be as large as the remaining area of said features. Of the 95 vanishing glaciers in AGI5, 85 were larger than 0.01 km² in AGI3.

Our RR experiment showed that analysts mostly but not always agree whether a glacier has disappeared. Considering the example of Madlein Ferner (Table , Sect. S3 in the Supplement), we suggest that phrasing such as “14 of 15 analysts agree this glacier is vanishing” or “this glacier has disappeared with over 90 % confidence” can serve as an adequate expression of uncertainty depending on the application. Such an approach would benefit from more extensive RR-type experiments designed specifically to assess vanishing glaciers and the limits of their detection.

In addition to the minimum size requirement, the recent SGIs apply criteria related to “evidence of flow” such as crevasses and deformation features to exclude ice bodies that do not meet the definition of a glacier as per . Similar approaches are described in, e.g., and , although not all types of crevasses necessarily indicate ice flow . In AGI5, filtering by size (>0.01 km²) and the “visible crevasses” attribute, 396 glaciers covering an area of 265 km² remain. That is, almost half of the features included in AGI5 would not meet the criteria for inclusion in the SGI. Although the impact of the filter on glacier count is large, the impact on total area is limited with excluded features accounting for only 7 % of total AGI5 area. Despite the differing criteria, AGI3 and AGI5 and SGI2016 and SGI2023 (forthcoming, ) both indicate the disappearance of about 11 % of the total glacier count in the respective inventories.

How to treat disappeared or vanishing glaciers in inventories, how to define when a glacier has disappeared, and what to call glacial remnants that may no longer meet common definitions of a “glacier” are matters of current discussion within the glaciological community e.g.,. Following the approach of previous AGI, AGI5 includes all glacier ice identified in Austria, regardless of size or other criteria. Comparability with other approaches, such as that of the SGI, is achieved with the various data attributes that allow users to filter depending on their use-cases. We concur with , who points out that when a glacier has disappeared depends on “who is asking and why”, and hope that AGI5 can contribute to answering this question in Austria for different types of users and applications.