Contributors: Jacqueline Bannwart (University of Zurich), Inés Dussailan (University of Zurich), Frank Paul (University of Zurich), Michael Zemp (University of Zurich)
Issued by: UZH / Frank Paul
Date: 09/01/2023
Ref: C3S2_312a_Lot4.WP3-TRGAD-GL-v1_202204_A_TR_GA_i1.1
Official reference number service contract: 2021/C3S2_312a_Lot4_EODC/SC1
History of modifications
List of datasets covered by this document
Related documents
Acronyms
General definitions
Brokered data set: A dataset that is made available in the Climate Data Store (CDS) but freely available (under given license conditions) from external sources. In the case of the glacier distribution service, the Randolph Glacier Inventory (RGI) is brokered for the CDS from https://glims.org/RGI under a CC-BY 4.0 license.
Hypsometry: Area distribution with elevation, usually summarized for 100 m intervals.
Rock glacier: Rock glaciers area landforms composed of ice and rock, which are slowly creeping down slope. They are found in regions with permafrost that are too dry for glacier formation (when talus derived). However, some rock glaciers also originate from debris-covered glaciers or ice-cored moraines and are difficult to be separated from glaciers as the transition us continuous. For several reasons it is important to keep rock glaciers separate from glaciers (e.g. they advance when climate is getting warmer) and they are thus usually not included in a glacier inventory. But as they also constitute locally an important water resource they are sometime included. The best practice here is to at least identify them as rock glaciers (or any other transitional ice-debris land form) so that the related polygons can be excluded from a sample.
Scope of the document
This document is the Target Requirements and Gap Analysis Document (TRGAD) for the glacier distribution service within the Land Hydrology and Cryosphere (LHC) service. A related document for the glacier change service is available from [RD1]. We here provide users with condensed information on product characteristics and data processing, as well as product requirements (technical specifications) and gaps. The latter refer to data availability and spatial coverage or in terms of scientific research, gaps to enable the current glacier products to be evolved (e.g. faster algorithms) to respond to specified user requirements (e.g. further quality improvements). It is noted that the content of this specific document is also included the TRGAD of the LHC service, together with the information on the glacier change service and three other Essential Climate Variables (ECVs).
Executive summary
The document starts with a short introduction into the nature of glaciers and glaciology, their global distribution, how and by whom their extents are mapped and where the created datasets are stored. This should give the background to better understand why certain points of the data processing and distribution is very different from other ECV products. For example, the product format is a shape file that is usually created by a single analyst from manually selected satellite images on a desktop computer by on-screen digitizing. Afterwards we provide an overview on the input and auxiliary data, a short description of the retrieval algorithms and a note on the methodology applied for uncertainty estimation. Second, the target requirements according to international organisations such as the Global Climate Observing System (GCOS) and the scientific community are presented. For the former, we explain how the various entries have to be interpreted, as this is not always fully clear. For the latter (community needs) we explain how they are linked to the former and what the special challenges are when creating the glacier area (or outline) product. Third, the gap analysis provides an overview on the historic development regarding satellite data (input) and glacier outlines (output) to better understand the data gaps and needs as well as the further requirements. The latter have a focus on possible further development of algorithms, uncertainties and CDS products. We close with a description of scientific research needs and opportunities for a further data exploitation from optical sensors.
1. Product overview
1.1. Introduction
Glaciers originate from compressed snow and can thus be found where the annual snow fall survives summer melt. This requires low temperatures, sufficient solid precipitation and a not too steep region where snowfall can accumulate. Glaciers are thus found in mountain regions all over the world, from the Equator to the poles. The warmer it gets, the higher their elevations, the larger the precipitation amounts, the larger they can grow. Thereby, glaciers can flow down to low elevations when the higher mass loss here is compensated by higher precipitation. When temperatures increase, glaciers retreat to higher elevations where it is cooler and mass loss is smaller. If the mountain range is no longer high enough to keep snow over a full year, glaciers will ultimately disappear. The presence of glaciers is thus telling us something about the climatic conditions in a region and the changes of their extent about the changes of climate. If glacier mass loss due to melt of ice and snow during summer is the same as the mass gain due to snowfall in winter, their mass balance over a year is zero. Also in this case they contribute to run-off in summer and thus provide water when it is most needed, e.g. for irrigation. When mass loss over a year is higher than the mass gained, their mass balance is negative and the excess melt contribute to global sea-level rise. For these and several other reasons (e.g. natural hazards, tourism) it is important to know where glaciers are located, how large they are and much water they store (volume) and release (mass balance), and how they change over time.
To answer such questions at the local, regional or global scale, a glacier inventory is required. These days this means that we need to have a geolocated outline of each glacier in the world, i.e. globally complete glacier inventory in a vector and/or raster format that can be handled by special computer software to answer the above questions. Not to forget, for many applications it is also important to know which regions are not covered by glaciers, e.g. to identify stable terrain for uncertainty assessment. When talking about glaciers, we refer to the 215,000 entities that are distributed in mountain and polar regions of the world (with sizes between 0.01 and about 10,000 km2) rather than to the two ice sheets in Greenland and Antarctica which are several orders of magnitude larger. Glaciers can have a wide range of forms and sizes and some of these have special names (such as ice caps, ice fields, valley glaciers or cirques). Owing to their locations in mountains, rock fall from weathering processes might cover a certain part of their surface, hiding the ice underneath. This can make the identification of their boundary very difficult.
Where do the glacier outlines now come from? These days, glacier extents are derived by manual on screen digitizing or semi-automatically from remote sensing data, usually multispectral optical sensors such as Sentinel-2 or Landsat, sometimes also aerial photography or very high-resolution satellite imagery. The multispectral sensors allow mapping of clean ice with a simple spectral band ratio and a threshold applied to the resulting ratio image. The band ratio method is used unchanged since 35 years and usually applied to individual satellite scenes which are selected, downloaded and processed by an analyst on a local computer. This is very different from the processing workflow for other ECVs, which might just process all incoming data automatically. Cloud-based processing with tools such as Google Earth Engine is also possible to get raw (clean ice) glacier outlines, but then the control on the images selected for processing is less good. And this is the essential part of a high quality product as only the scenes with the best glacier mapping conditions should be used.
Accordingly, the workload to create a glacier inventory for a larger region is very high and usually only done as part of science projects for small regions. A key reason for the latter point was that Landsat data were commercially distributed for a long time and covering a larger region was very expensive. To date, data access is free and open, but the workload for the analyst remains. Apart from Copernicus Climate Change Service (C3S), there is no funding available for repeat large-scale glacier mapping and the interested community has organized themselves to get the datasets produced according to some data standards and distributed in a free and open repository, the Global Land Ice Measurements from Space (GLIMS) glacier database (see https://glims.org). This database holds basically everything that has been submitted by the analysts, independent of date, region or number of glaciers. Accordingly, the quality of the outlines varies and the process of filling the database was very slow. Via a special effort for the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC AR5), the scientific community created the first globally near-complete inventory with the name Randolph Glacier Inventory (RGI) in 2012 (Pfeffer et al. 2014). The latest version 6 of the RGI was published in 2017 and created by manually selecting the best datasets from the GLIMS database. The C3S glacier distribution service contributes to the GLIMS database by creating spatio-temporal subsets of glacier outlines (ICDRs) for regions of demand and sending them to GLIMS. The descriptions given here and in other documents mostly describe how this is done in C3S.
1.2. Input and auxiliary data
The dataset that is provided for the Climate Data Store (CDS) by the glacier distribution service is the RGI that is brokered from the GLIMS glacier database at glims.org/RGI and consists of glacier outlines with associated attribute information in shape file format and a csv file providing the hypsometry of each glacier. The dataset covers all glaciers in the world and has been compiled from regional-scale datasets provided by the scientific community to the GLIMS database (glims.org) at the National Snow and Ice Data Center (NSIDC). Whereas GLIMS is multi-temporal and may contain outlines of the same glacier from several points in time, the RGI is a snapshot and includes only one outline per glacier. The datasets in the RGI have been derived by the respective data providers from several sources, for example satellite images, aerial photographs or digitized topographic maps (Pfeffer et al. 2014). We here refer to the methods applied by C3S [RD1] and the European Space Agency (ESA) project Glaciers_cci [RD2].
Glacier outlines are created from (1) high-resolution (10-30 m) multispectral optical satellite data (e.g. Terra Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Landsat Thematic Mapper (TM)/Enhanced Thematic Mapper plus (ETM+)/Operational Land Imager (OLI), Sentinel-2 Multi Spectral Imager (MSI)) and (2) a digital elevation model (DEM) to calculate (a) drainage divides between glaciers and (b) topographic information for each glacier entity (e.g. minimum, mean, median and maximum elevation, mean slope and aspect). As only debris-free ice is mapped by the retrieval method (see Section 1.3) and misclassification of water surfaces and regions in shadow can occur, the raw glacier outlines have to be corrected manually. For this purpose, additional information is used to facilitate interpretation, for example very high-resolution images as available in Google Earth or provided by web map-services or coherence images derived from Synthetic Aperture Radar (SAR)) data. It has to be noted that the required operator editing is always subjective and differences in interpretation can occur that are neither right nor wrong. In effect, all glacier outlines are manually quality improved and a direct product inter-comparison across different analysts essentially provides differences in interpretation rather than an objective algorithm validation. In the same sense, validation data (in most cases very high-resolution satellite or aerial imagery are used) often do not provide a fair comparison as their higher resolution leads to a different interpretation of the glacier boundary and thus to a different outline (see Section 1.4).
1.3. Retrieval algorithm
The core of the glacier mapping algorithm as applied by C3S is the calculation of a spectral band ratio using a visible and shortwave infrared (SWIR) band and a (manually selected) threshold value to classify the resulting ratio image into a class 'glacier' and a class 'other'. The method is applied unchanged since more than 25 years and does not have a version number or a specific name, but we refer to it as the 'band ratio method'. Further details of the largely manual post-processing are provided in Section 3.2.
1.4. Uncertainty estimation
The differences in interpretation described in Section 1.1 also limit the possibilities for a rigorous quality assessment of the generated products. In general, the interpretation differences are much larger than those introduced by the method applied and often also larger than changes over a 5-10 year period. Another problem is that the possibility for quality assessment and validation, depends on the availability of appropriate higher resolution datasets (e.g. aerial photography). However, these are often not appropriate for a comparison as they are generally obtained at a different date. For example, a comparison is of limited value when snow conditions are different in the high-resolution data. Unfortunately, such datasets are in general very expensive and come with a different geolocation (as a different DEM is used for orthorectification). A current workaround is the use of so-called web map-services that might provide, by chance, very high-resolution images that are comparable and can then be used for a quality assessment (e.g. Andreassen et al. 2022).
In principle, all glacier outlines are quality checked by the analyst before submission. This glacier-by-glacier quality control is required as the automated methods do not map glacier ice under debris cover or partly miss ice in shadow (omission errors) but map turbid lakes or ice bergs as glaciers (commission errors). These had to be corrected/removed before an outline has acceptable quality, (Figure 1), i.e. reaching the GCOS uncertainty specifications described in Section 2.1. A critical point for change assessment is to determine if an observed change is significant, i.e. larger than the uncertainty. This requires determination of a quantitative uncertainty measure (see Section 3.3 for related methods).
Figure 1: Overlay of glacier outlines that have been classified automatically (yellow), manually added (red), and removed (white) for the Oberaarglacier (bottom) and the heavily debris-covered tongue of Unteraar-glacier (top). Glacier ice in shadow is correctly mapped in this region. Image width is 14.1 km, North is up.
2. User Requirements
2.1. International documents (GCOS, IGOS)
The requirements for the area product of the ECV Glaciers have been described in documents from international organizations. Two of these documents provide the foundation the work performed in C3S. First, GCOS (2006) identifies the glacier area product under Product T.2.1:
'Product T.2.1: Maps of the areas covered by glaciers other than ice sheets'
This requirement was refined in GCOS (2011) to Product T.3.1:
'2D vector outlines of glaciers and ice caps (delineating glacier area), supplemented by digital elevation models for drainage divides and topographic parameters'
The listed benefits of such a product are:
- Support for the instrumental data record of climate by providing climate-related information, further back in time, in remote areas and at higher altitude than meteorological stations.
- Input to regional climate models and the validation of impact assessment and climate scenarios on a regional scale.
- Computation of glacier melt contribution to regional hydrology and global sea-level rise.
This need for a globally complete glacier inventory (in vector format) was and still is the main driver behind the intense work carried out to create it. The first version of this inventory (the RGI) has been improved several times since 2012, with v5 and v6 being also available from the CDS.
The second document of major importance for C3S work is the Cryosphere Theme Report of the Integrated Global Observing Strategy (IGOS 2007). This document contains in its Appendix B.6 a detailed overview of technical requirements for glacier products. The related table has been widely used in other documents, is still valid and reproduced here as Table 1.
Table 1: Target requirements for the ECV Glaciers according to IGOS (2007). Abbreviations are as follows: C: Current Capability, T: Threshold Requirement (Minimum necessary), O: Objective Requirement (Target), L: Low end of measurement range, U: Unit, H: High end of measurement range, V: Value, mo: month, yr: year.
The entries in Table 1 require some explanation. Firstly, the values listed for measurement range and accuracy are valid and still reflect the current target requirements. Spatial (5 m) and temporal (30 y) resolution for 'Area' in the row airborne refers to the accuracy and repetition frequency of glacier inventories derived from aerial photography (incl. manual digitization of outlines) at that time. This applies similarly to the rows 'Landsat etc.' and 'Hi-res optical'. However, for the latter we are now in the 5 to 10 m range (with Pleiades and Sentinel-2) and the target requirement has already been surpassed. The '1 yr' temporal resolution refers to what would be theoretically possible, but makes little sense in regard to the tiered observing strategy established by the Global Terrestrial Network for Glaciers (GTN-G) as glacier response times are generally a few decades and annual fluctuations are measured as terminus changes in the field. To be significant, glaciers have to change the position of their terminus by about 30 m (with Sentinel-2), which is at current annual glacier retreat rates in the Alps typically achieved after about 5 years.
The row 'Topography' basically refers to the quality requirements for DEMs used to determine glacier volume/mass changes with the geodetic method (DEM differencing). We include this information here as the quality requirements to derive accurate drainage divides for glacier inventories and topographic information for each glacier are about the same.
In the GCOS Implementation Plan from 2016 (GCOS 2016) the requirements for the three products Glacier area, elevation and mass change have been updated to overcome the potential confusion of glacier elevation changes with glacier topography (Table 2). It has also been clarified that uncertainties for glacier area should not exceed 5% to fulfil GCOS requirements. This means that manual editing has to be applied to outlines of glaciers under seasonal snow or debris cover, connected to lakes or to glaciers with parts of their area in shadow (see Figure 1). The latest GCOS report is listing the requirements for the glacier area product in Section 8.2.1 of its Supplement (GCOS 2022) in more detail. This new overview includes a notes section for clarifications along with threshold values that have been adjusted to current sensor capabilities (Table 3). The previous confusion with the temporal resolution has been clarified and product uncertainties are now given in dependence of the spatial resolution of the sensor. Whereas this dependence and the values are largely correct, it has to be noted that uncertainties do more depend on glacier size, image conditions (e.g. seasonal snow) and debris cover (which might cover large parts of a glacier. Overall, threshold values only changed slightly and their description improved over time.
Table 2: Target requirements for glaciers according to GCOS (2016). Please note that the cited IGOS (2009) is actually IGOS (2007). The table has been slightly cropped.
Table 3: Target requirements for glaciers according to GCOS (2016). Please note that the cited IGOS (2009) is actually IGOS (2007).
2.2. Community needs
In agreement with Product T.3.1 of GCOS (2011), the basic need of a very large user community is a globally complete glacier inventory with sufficient quality in a vector format. The related information about glacier location and extents (i.e. the glacier area product) is required for a wide range of applications and calculations serving societal needs. These range from hydro-power production at a local scale, to river-runoff at a regional scale, to determination of their contribution to sea level at a global scale. It includes scientific applications such as ice thickness estimation, modelling of past and future glacier extents, climate change impact assessment and improved process understanding from observed changes. Moreover, all glacier-specific calculations (e.g. geodetic mass balance, flow velocities, end of summer snow cover) need glacier boundaries as a basic input dataset. This is not only required to spatially constrain the calculations and derive a glacier-specific mean value, but also for uncertainty assessment of these products (e.g. the related calculations need to know where the 'stable terrain' off glaciers is) or even basic pre-processing steps such as DEM co-registration (Nuth and Kääb 2011). Finally, early warnings related to hazards (e.g. glacier lake outburst floods) or potential flooding (due to combined snow and ice melt in late spring) from governmental agencies and national hydrological services require precise information of glacier extents with a given time stamp. The dataset in the CDS (the RGI) is serving these needs.
Today, the above users (e.g. hydrologists, glaciologists) convert or directly assimilate shape files in their models and applications so that also the C3S CDS is providing this product in this format (Product T.3.1 in GCOS 2011). Downstream applications might aggregate or rasterize the information to a sampling distance of choice, but this is specific to the respective application / model and has thus so far not been provided by the CDS. We have, however, identified a need for C3S and climate model applications to also have a raster data version of the RGI available in the CDS, e.g. in netCDF format. Such a product will thus be created in the framework of C3S2.
In an ideal case, glacier outlines would be compiled globally each year to combine them with other datasets and get the most precise modelling results. This user need cannot be accomplished as the workload for a single inventory is too high and appropriate satellite imagery with good mapping conditions (e.g. free of clouds and snow off glaciers) are usually only available over much longer time scales, e.g. every 5-10 years. Luckily and as mentioned above, glaciers do (in most regions of the world) not change that fast, i.e. the outlines obtained in a specific year might be used for a time period of ±5 years (depending on the region and area change rates) without introducing too large errors in the derived products. The uncertainty in their extents is often higher than their changes over a 5 to 10-year period, in particular when they are debris covered or frequent seasonal snow is hiding the true glacier perimeter. This might be different when working on a more regional or local scale, but for such applications higher quality datasets are often available or can be created. So the desire for an annual update of the global glacier inventory cannot be fulfilled, but the strategy of an update every one or two decades is in line with GCOS requirements (see Table 3).
Similarly, the community need for a consistent interpretation of glacier extents has not yet been achieved as glacier outlines are created by a globally distributed community and the interpretation of what a glacier is and which parts belong to it can depend on sensor resolution, the application and the experience of the analyst. Most glacier outlines are created by science projects and these might have different needs than other applications. For example, for water resources assessment a science project might not be interested in mapping the steep and thus thin upper part of glaciers, or glaciers smaller than a certain size are not mapped because the region is dominated by huge glaciers and the focus is on the sea level contribution of its glaciers. Moreover, there is room for interpretation and related differences might not be right or wrong, i.e. both can be correct despite being different. Due to the variability in interpretation of glacier extents by different analysts, it is not advisable to use other datasets for change assessment. In principle, the differences can be as large as real glacier changes over a longer period, in particular when satellite scenes with adverse snow conditions were used. Such issues are addressed by C3S in cooperation with data providers.
Finally, for several applications it is beneficial when glacier outlines in a larger region (e.g. a country or mountain range) refer to the same year. This facilitates change assessment and comparisons across glaciers for most of the applications mentioned above (e.g. geodetic mass balances). It can be longer than a year when area change rates are small (e.g. in the high Arctic), but for marine-terminating glaciers, changes can be much stronger than for neighbouring glaciers terminating on land and a frequent update can be required as well. Avoiding a mixture of acquisition years would thus be really beneficial. For some modelling applications (e.g. future glacier evolution), a compensation is possible when at least a date is given for each outline, but also here the temporal differences should not be too large.
From a more technical point of view one can summarize community needs for glacier outlines as:
(1) Globally complete, accuracy better than 5%, consistent in interpretation, well documented.
(2) Acquisition over a short period of time (a month) for large regions.
(3) Globally acquired over a short time period (a few years), matching to existing DEMs.
(4) Free availability, in shape file format and consistent attributes for all entries.
(5) Complete meta-information (analyst, method, satellite, quality flags).
(6) Frequent extension of attributes and quality improvements (as possible); and
(7) Consistent multi-temporal datasets.
All of the above needs are not yet fully met and one goal of the work in C3S2 will be to improve on this situation (cf. also Section 3.4). For example, the community needs (2) and (3) had partly already impacted the previous work of C3S and influenced some of the main goals of the International Association of Cryospheric Sciences (IACS) Working Group 'Randolph Glacier Inventory (RGI) and its role in future glacier monitoring and GLIMS' (https://cryosphericsciences.org/activities/working-groups/rgi-working-group). It was decided for its 2020 to 2023 term that the new version 7 of the RGI should be (a) quality improved over RGI6 and (b) be closer to the year 2000 than RGI6. The latter is important to have a better temporal match of glacier extents to the DEMs available from 2000 (such as from Shuttle Radar Topography Mission (SRTM) or from ASTER stereo data) when calculating elevation/mass changes of glaciers. The C3S team has thus performed a detailed survey of the datasets in RGI6 and identified the regions that needed quality improvements and/or were far away from the year 2000. Together with the community, C3S created or contributed to several of these new datasets, including glacier outlines for Ellesmere and Baffin Island, northern Greenland, the Alps, Kamchatka, Peru/Bolivia, Southern Andes and New Zealand. These new datasets also adhere to the requirements (1), (4) and (6). Due to temporal limitations or missing satellite data, not all regions requiring updates or improvements could be covered (e.g. USA, Svalbard), but substantial progress could be made and open work for C3S2 is already defined.
3. Gap Analysis
3.1. Description of past, current and future satellite coverage
3.1.1. Historic Development
For glacier outlines derived from optical satellite data (Landsat type) the possibilities for data retrieval have constantly improved over the past decades. However, there has been one major break point that changed everything: the opening of the Landsat archive in 2008 (Woodcock et al. 2008, Wulder et al. 2012). Without this step it would have never been possible to utilize the vast archive of images (>3 million scenes) for global scale applications, in our case the global glacier inventory (Pfeffer et al. 2014). However, it has to be mentioned that the glaciological community found a way to have free access to multi-spectral (15 m) ASTER data (a sensor on-board the Terra spacecraft launched in 1999) by registering to the GLIMS (Global Land Ice Measurements from Space) initiative and establishing GLIMS as a major science application of the ASTER data acquisition strategy (Raup et al. 2000). With the GLIMS database in place and algorithms for automated glacier mapping being developed, population of the database with glacier outlines slowly started. There were three main bottlenecks in the earlier days of glacier mapping:
- Debris cover had to be digitized manually (this is still the main bottleneck!)
- The number of ASTER scenes with good snow conditions (i.e. minimum seasonal snow extent, not hiding glacier outlines) were small in the first 5-10 years
- The satellite scenes had to be orthorectified by the analyst, requiring a digital elevation model (DEM) of appropriate resolution and quality, digital image processing software that allows working with the Hierarchical Data Format (HDF) file format, and manual collection of ground control points.
Despite the global network of participating institutions, progress towards global scale coverage was slow as funding for the required mapping had to be taken from science projects and these had their own regional priorities. A first glimpse into a more promising future was established after Landsat scenes collected by the Global Land Cover Facility (GLCF) were made available for free (instead of 475$ per scene) at original resolution and with all spectral bands in geotif format (Tucker et al. 2004). This dataset was also used for mapping glacier extents (e.g. Paul and Kääb 2005), but snow conditions were often not ideal for accurate glacier mapping (i.e. seasonal snow was hiding the glacier perimeter). Already at that time online tools such as earthexplorer.usgs.gov or glovis.usgs.gov revealed that a huge amount of satellite scenes with optimal glacier mapping conditions were available in the archives and numerous scientist were eager to process them.
In 2008, the archive was opened and all scenes were provided as already orthorectified geotifs. This had the enormous advantage that analysts did not have to do this important but very time-consuming step themselves. The quality of the orthorectification was overall sufficient (about ±1 pixel standard error) but regionally variable and mostly dependent on the source data used for the GLS2000 DEM that served as a baseline dataset for this purpose. That DEM was largely based on the SRTM DEM and other national elevation datasets (NED) or military sources outside the SRTM coverage. A first crisis occurred for global glacier mapping when the scan line corrector (SLC) of Landsat 7 ETM+ sensor failed in May 2003. Although scenes were still usable in the middle third, the so-called SLC-off scenes now had strong limitations in global coverage. In consequence, the 20-year old (but still working) Landsat 5 with its Thematic Mapper (TM) sensor was reactivated and helped to complete coverage. When TM failed in 2011, it had more or less continuously acquired calibrated images of the Earth's surface for 27 years (since 1984), the longest time-series for a civilian EO satellite on record (Belward and Skoien 2015).
The regional data gaps in 2012 caused only a small problem in time series continuity as the successor of the Landsat 7 ETM+ sensor, Landsat 8 OLI, was on its way and started acquisitions in 2013 with unprecedented quality (e.g. new bands, revised spectral ranges, and 16-bit quantization). As the free data access policy was continued and the geometric quality of the orthorectification further improved, a second promising phase of global glacier mapping started. This second phase reached new dimensions with the launch of Sentinel-2A in June 2015 (and Sentinel-2B in March 2017) as its much higher spatial resolution (10 m) and larger swath width (290 instead of 180 km) allows glacier mapping (and the still required manual corrections of debris cover) with unprecedented quality (Paul et al. 2016). So compared to the situation in 2012 (where only disturbed ETM+ scenes were available) or in 2005 (where only orthorectified GLCF scenes were freely available), we are now in a glaciologist's paradise. Another important issue was solved recently by space agencies: At the end of 2021 the Landsat and Sentinel-2 scenes archived by National Aeronautics and Space Administration (NASA) (Land Processes Distributed Active Archive Center - LPDAAC) and ESA were reprocessed and orthorectified with the same DEM. This should considerably improve their previously poor co-registration accuracy (Kääb et al. 2016). In effect, this will not only foster the joint use of the satellites, but also improve results of applications that combine the glacier outlines with other geo-spatial datasets such as DEMs or infrastructure data.
3.1.2. Spectral and spatial properties
Automated glacier mapping (clean ice) is largely based on calculating a simple band ratio (e.g. red/SWIR) and applying a threshold to create a binary glacier mask that can be converted to glacier outlines using a raster to vector conversion (e.g. Hall et al. 1988, Bayr et al. 1994, Paul et al. 2002). This works well, as spectral properties of ice and snow are very different in the SWIR (where both have a very low reflectance) compared to the red or Near Infrared (NIR). Owing to the windows of atmospheric transmission and physical principles, the spectral ranges of the required spectral bands are very similar on all optical sensors that can be used for glacier mapping (Table 4). Accordingly, the methods developed for automated glacier mapping with Landsat TM can also be used for Landsat ETM+ and OLI, Terra ASTER, Satellites Pour l'Observation de la Terre (SPOT) High Resolution Visible (HRV), Sentinel-2 MSI and several others. The only requirement is a spectral band in the SWIR, otherwise only manual delineation of outlines can be applied (this is for example required for all of the very-high resolution sensors such as Quickbird, GeoEye, Kartosat or Worldview as well as for aerial images).
A change in the spectral range of the panchromatic band on Landsat 8 OLI (now covering only green and red instead of green to near infrared, see Table 4) now allows using the 15 m band also for glacier mapping with a pan/SWIR ratio (Paul et al. 2016). Accordingly, the resulting outlines are two times sharper. By pan-sharpening the other visible bands with the 15 m band (which is now possible), the quality of the manual editing (debris cover) can also be improved. However, resolving crevasses and thus giving a realistic glacier representation seems to require at least 10 m spatial resolution. On the downside, the ASTER sensor lost its SWIR band in 2008 so that automated glacier mapping only works with scenes acquired before that date.
Table 4: Spectral ranges of individual bands for a range of optical sensors (from Paul et al. 2016). Colours decode spatial resolution (black: 30 m, red: 20 m, blue: 15 m, green: 10 m).
All Landsat TM, ETM+, and OLI sensors have the same spatial resolution in the red, NIR and SWIR bands (30 m) resulting in the spatial consistency of products. For ASTER, SPOT and Sentinel-2 the resolution of the SWIR band is half as good as for the visible and near infrared (VNIR) bands (15 / 30 m for ASTER and 10 / 20 m for SPOT and Sentinel-2). If the higher resolution product should be generated from these sensors, it is required to first resample the SWIR band two times (at best using a simple bilinear interpolation).
All sensors have sun-synchronous orbits with acquisition times around 10:30 am local, a compromise between solar elevation (casting stronger shadows when low) and cloud development (often starting before noon in mountain regions). Apart from the global acquisition strategy, seasonal snow at the end of the ablation (or dry) period and clouds are main obstacles to produce glacier outlines regularly. In some regions it can take more than 10 or 15 years before the next useful acquisition is made (Paul et al. 2011). For this reason, it is required to analyse new acquisitions each year and process them as required.
3.1.3. Global coverage
Apart from clouds and seasonal snow, global coverage of glaciers is also limited by the acquisition strategy. Whereas this has changed today as Landsat 8 and Sentinel-2 acquire images more or less continuously, the limited on-board storage capacity of Landsat TM/ETM+ resulted in a pattern of acquisitions around ground receiving stations (Goward et al. 2006). As the network of these stations was successively extended, more and more regions where covered. During the commercial phase of Landsat acquisitions in the 1990s, images were only acquired upon request in several regions, so that large regions are not covered. In consequence, nearly all glaciers in High Mountain Asia are not covered before 1988 (Figure 2).
A second consequence of the distributed acquisition is that the (since 2008) freely available orthorectified product (L1T) was constrained to the holdings in the United States Geological Survey (USGS) archives at LPDAAC. Scenes outside the US were strongly underrepresented, and it was with some luck to have useful acquisitions over a particular region already covered. The still on-going Landsat Global Archive Consolidation (LGAC) will transfer all Landsat scenes from around the world into the USGS archive and process them to the L1T standard (Wulder et al. 2016). For this reason, global coverage is constantly increasing and new possibilities of product generation emerge. This is also true for cloud processing applications (e.g. in Google Earth Engine) or machine learning as the related data clouds will include a copy of all Landsat and Sentinel-1/2 scenes.
Figure 2: Landsat TM acquisitions from 1982 to 2005 (from Goward et al. 2006). This figure is illustrative of the regions covered and is not meant to be used for detailed analysis.
Spatial coverage with Sentinel-2 will further increase in the coming years and the 5 day (or shorter) repeat cycle with both Sentinels already helps to increase the chance for cloud-free acquisitions over large regions and short time periods. With the recently launched Landsat 9 satellite (Landsat 7 has now been decommissioned) also the Landsat coverage will come back from 16 to just 8 days (more often towards the poles). Also the successors of Sentinel 2A/B and Landsat 9 are already under construction so that for the next 5 to ten years continuity in data acquisition is nominally guaranteed. The problem with seasonal snow cover, however, will remain for the time being and it should not be expected that every region with glaciers will have a useful acquisition within 5 years.
Overall, the outlook for satellite-based glacier mapping and monitoring is very promising. The possibilities in the coming years might revolutionize our understanding of glaciers as the now available sensors can support several glaciological investigations (e.g. flow velocities and snow lines) with unprecedented spatial and temporal resolution (e.g. Paul et al. 2022).
3.2. Development of processing algorithms
The core of the glacier mapping algorithm is the calculation of a spectral band ratio using a visible and shortwave infrared (SWIR) band and a (manually selected) threshold value to classify the resulting ratio image into a class ‘glacier’ and a class ‘other’, usually represented by the values 1 and 0. Threshold selection is an optimization process that aims to reduce the workload for manual corrections of the resulting outlines. As debris cover and water surfaces have to be manually edited anyway, the focus is on obtaining the best possible result in regions with ice in shadow, where the threshold is most sensitive. For specific sensors or conditions, additional thresholds are required to improve the results, e.g. in the blue band (see Paul et al. 2016). The resulting binary raster map might be filtered for noise reduction (e.g. if many small snow patches are present and should be removed) and is then converted to vector format. Further details about the methods applied can be found in the ATBD of Glaciers_cci [RD2]. All editing takes place in the vector domain, using contrast-enhanced band combinations of the original satellite image in the background. Final outlines are intersected by drainage divides (that are obtained by a semi-automated flow-shed algorithm and have to be edited as well) to separate glacier complexes into individual glaciers. Finally, topographic information is derived for each glacier from the DEM using zone statistics (in short, the glacier outlines define a zone for which statistical values from an underlying DEM are calculated).
We follow the further development of processing algorithms closely. They will be performed as part of the Glaciers_CCI+ project and by the science community. In the case improvements appear in the literature, we will test if the method is sufficiently robust for the products we wish to provide for C3S2. Apart from this, the current method to map debris-free glacier ice is so simple and accurate, that a ‘better’ method is not expected to appear any time soon. This is different for the still evolving approaches to map debris-covered glaciers or applications with Google Earth Engine using cloud processing and machine learning, both reducing work load when large regions and/or many scenes are to be covered. As mentioned above, the requirement to manually correct debris-covered glacier parts results in inconsistency in the interpretation among different analysts. This reduces the possibility for change assessment when datasets are provided by different analysts. Harmonizing the interpretation of glacier outlines is, however, more an issue of post-processing rather than the algorithm itself. Thereby, the resulting uncertainty strongly depends on the regional glacier characteristics and the usefulness of glacier outlines for change assessment has to be clarified case-by-case. For the datasets provided by the CDS the impact of the inconsistency in interpretation is small, as the RGI only includes one outline from one point in time for each glacier.
3.3. Methods for estimating uncertainties
As real validation of glacier areas is only performed occasionally (in general due to missing reference datasets), a range of measures for uncertainty assessment have been developed which do not require reference data. The goal is to keep the uncertainty of the derived glacier areas at <5% to comply with GCOS requirements (see Section 2.1). Whereas this is generally possible, one has to consider that uncertainty strongly increases towards smaller glaciers. For glaciers <1 km2 uncertainty is usually larger and vice versa. But uncertainty also depends on the amount of manual corrections to be performed, i.e. is larger when many debris-covered glaciers have to be corrected. Which of the methods is or are applied is thus also depending on the regional glacier characteristics and one method might be favoured over another one. In Table 5, we briefly summarize the main methods sorted by level, which is related to the workload required to perform the related calculations. A full description can be found in the Uncertainty Characterization Report of the Glaciers_cci project (Glaciers_cci, 2016).
Table 5: Overview of the measures to determine accuracy and precision of glacier outlines.
Nr. | Name | Level | Description |
|---|---|---|---|
1 | Outline overlay | L0a | Manual editing, cross-comparison, interpreting differences, visualisation |
2 | Literature value | L0b | Assume accuracy will be as good |
3 | Buffer method | L1 | Buffer outline by ½ or 1 pixel, calculate min & max area, and SD |
4 | Multiple digitizing | L2 | Determine analysts precision (area variability) |
5 | Area difference | L3 | Use of very high-resolution reference data for accuracy assessment |
Based on these methods, common practice is to either:
- adopting a value from more detailed studies (e.g. Paul et al. 2013) to the own dataset (L0b),
- calculate a minimum and maximum extent for all glaciers by adding a buffer related to typical uncertainties (±½ or 1 pixel) and report the standard deviation of the area differences as an estimate of uncertainty (L1),
- digitizing or correcting several (at least 5, better 10) glaciers with different characteristics (e.g. large/small, debris/clean) at least three (better 5) times and use the normalized standard deviation of the derived areas as a measure for accuracy (L2),
- perform a comparison to glacier areas derived from very high-resolution imagery (L3).
The major shortcoming of current studies is that uncertainty estimates are sometimes not applied. In C3S we use measure Nr. 1 (L0a) to identify regions in the RGI that have poor quality and need to be improved (see Section 3.4). For the datasets created within the project we will mostly use measures Nr. 3 and 4 (L1 and L2) to determine product uncertainty. In selected regions we will also apply Nr. 5 (L3) to obtain an estimate of product accuracy. The related values will be reported as part of the Product Quality Assessment Report [RD3].
3.4. Opportunities to improve quality and fitness-for-purpose of the CDRs
As mentioned above, several regions have already been improved by the C3S team for RGI7. In Figure 3 we provide two examples of the improved datasets. In most cases we improved both glacier outlines and drainage divides.
Figure 3: Two examples of improvements by C3S for RGI7. Top: A large icefield on Ellesmere Island with outlines from RGI6 in red and the new ones in blue. Image width is 187.5 km, North is up. Bottom: New ice divides (white) had been created for Penny Ice Cap on Baffin Island as the previous ones from RGI6 (black) crossed the flow field of outlet glaciers, shown here as colour-coded velocities from blue (0 m/y) to red (2000 m/y) (provided by Millan et al. 2022). Image width is 124 km, North is up. Satellite images: earthexplorer.usgs.gov.
We also considered new national datasets (e.g. new inventories for Argentina and Chile) and edited them to make them suitable for the RGI (e.g. by deleting all polygons referring to rock glaciers, see General definitions). There will likely be further work on RGI7 as not all possible improvements have yet been made. However, in C3S2 we will focus on creating new glacier inventories from around the period 2015 to 2021 to feed into a possible RGI8 (or RGI2020). The main reasons being that Sentinel-2 allows creating such inventories with unprecedented spatial resolution (10 m) and quality (Paul et al. 2016), and modellers require a most recent dataset, be it for model initialization, validation or second point in time for change assessment. The possible regions for such new inventories have still to be identified, but as for RGI7 the main driver will be to reach spatial completeness under consideration of then available datasets.
Furthermore, the need for a global glacier inventory in raster format has been identified as an urgent need for the CDS. The technical details (e.g. spatial resolution) of this dataset will be determined in the course of the project.
3.5. Scientific Research needs
To keep the work in C3S2 distinct from Climate Change Initiative (CCI), scientific research needs (e.g. related to algorithm development, uncertainty characterization, or key science questions) will be addressed as part of the CCI. However, assessment of glacier area changes over large regions will primarily be performed in C3S2.
3.6. Opportunities from exploiting the Sentinels and any other relevant satellite
As mentioned above, the time series of Sentinel-2 data will primarily be used to create new glacier inventories over the most recent time period. They will be submitted to the GLIMS glacier database and might potentially be considered by the IACS working group on the RGI for a currently planned new version of the RGI that will contain glacier outlines over a most recent time period, likely the Landsat 8/9 and Sentinel-2A/B period. Data production will be tailored to the requirements for this new RGI version. With the current re-processing of all datasets the former geolocation differences between Landsat and Sentinel should no longer occur. We will analyse this in C3S for regions where both datasets will be processed.
References
Andreassen, L.M., Nagy, T., Kjøllmoen, B., Leigh, J.R. (2022). An inventory of Norway's glaciers and ice-marginal lakes from 2018–19 Sentinel-2 data. Journal of Glaciology 1–22. https://doi.org/10.1017/jog.2022.20 (URL resource last accessed 9th January 2023)
Bayr, K. J., Hall, D. K. and Kovalick, W. M. (1994): Observations on glaciers in the eastern Austrian Alps using satellite data. International Journal of Remote Sensing, 15, 1733-1742.
Belward, A.S., and Skøien, J.O. (2015): Who launched what, when and why; trends in Global Land-Cover Observation capacity from civilian Earth Observation satellites. ISPRS Journal of Photogrammetry and Remote Sensing, 103, 115-128.
GCOS (2011): Systematic Observation Requirements for Satellite-Based Data Products for Climate - 2011 Update, GCOS-154, December 2011.
GCOS (2016): The Global Observing System for Climate: Implementation Needs. GCOS Implementation Plan. GCOS-200, 325 pp. (https://library.wmo.int/viewer/55469?medianame=GCOS-200_OnlineVersion_#page=1&viewer=picture&o=&n=0&q=; URL resource last accessed 9th January 2023)
GCOS (2022): The 2022 GCOS ECV Requirements. GCOS-245, World Meteorological Organisation, 244 pp. Online at: https://library.wmo.int/index.php?lvl=notice_display&id=22135 (URL resource last accessed 9th January 2023)
Glaciers_cci (2016): Uncertainty Characterisation Report v2 (UCRv2): Prepared by the Glaciers_cci consortium, 34 pp (online: http://www.esa-glaciers-cci.org/index.php?q=webfm_send/154; URL resource last accessed 9th January 2023).
Goward, S., Arvidson, T., Williams, D., Faundeen, J., Irons, J. and Franks, S. (2006): Historical Record of Landsat Global Coverage: Mission Operations, NSLRSDA, and International Cooperator Stations. Photogrammetric Engineering and Remote Sensing, 72 (10), 1155-1169.
Hall, D.K., Chang, A.T.C. and Siddalingaiah, H. (1988): Reflectances of glaciers as calculated using Landsat 5 Thematic Mapper data. Remote Sensing of Environment, 25, 311-321.
IGOS (2007): Integrated Global Observing Strategy Cryosphere theme report - For the monitoring of our environment from space and from Earth. Geneva: World Meteorological Organization. WMO/TD-No. 1405, 100 pp.
Kääb, A., Winsvold, S.H., Altena, B, Nuth, C., Nagler, T. and Wuite, J. (2016): Glacier remote sensing using Sentinel-2. Part I: Radiometric and geometric performance, and application to ice velocity. Remote Sensing, 8(7), 598, doi:10.3390/rs8070598.
Nuth, C. and Kääb, A. (2011): Coregistration and bias corrections of satellite elevation data sets for quantifying glacier thickness change. The Cryosphere, 5, 271-290.
Paul, F., Kääb, A., Maisch, M., Kellenberger, T. W. and Haeberli, W. (2002): The new remote-sensing-derived Swiss glacier inventory: I. Methods. Annals of Glaciology, 34, 355-361.
Paul, F. and Kääb, A. (2005): Perspectives on the production of a glacier inventory from multispectral satellite data in the Canadian Arctic: Cumberland Peninsula, Baffin Island. Annals of Glaciology, 42, 59-66.
Paul, F., Andreassen, L.M. and Winsvold, S.H. (2011): A new glacier inventory for the Jostedalsbreen region, Norway, from Landsat TM scenes of 2006 and changes since 1966. Annals of Glaciology, 52 (59), 153-162.
Paul, F., N. Barrand, E. Berthier, T. Bolch, K. Casey, H. Frey, S.P. Joshi, V. Konovalov, R. Le Bris, N. Mölg, G. Nosenko, C. Nuth, A. Pope, A. Racoviteanu, P. Rastner, B. Raup, K. Scharrer, S. Steffen and S. Winsvold (2013): On the accuracy of glacier outlines derived from remote sensing data. Annals of Glaciology, 54 (63), 171-182.
Paul, F., S.H. Winsvold, A. Kaab, T. Nagler and G. Schwaizer (2016): Glacier Remote Sensing Using Sentinel-2. Part II: Mapping Glacier Extents and Surface Facies, and Comparison to Landsat 8. Remote Sensing , 8(7), 575; doi:10.3390/rs8070575.
Paul, F., Piermattei, L., Treichler, D., Gilbert, L., Girod, L., Kääb, A., Libert, L., Nagler, T. Strozzi, T. and Wuite, J. (2022): Three different glacier surges at a spot: What satellites observe and what not. The Cryosphere.16 (6), 2505-2526
Pfeffer, W.T., A.A. Arendt, A. Bliss, T. Bolch, J.G. Cogley, A.S. Gardner, J.-O. Hagen, R. Hock, G. Kaser, C. Kienholz, E.S. Miles, G. Moholdt, N. Mölg, F. Paul, V. Radic, P. Rastner, B.H. Raup, J. Rich, M.J. Sharp and the Randolph Consortium (2014): The Randolph Glacier Inventory: a globally complete inventory of glaciers. Journal of Glaciology, 60 (221), 537-552.
Raup, B. H., Kieffer, H. H., Hare, T. M., and Kargel, J. S. (2000): Generation of data acquisition requests for the ASTER satellite instrument for monitoring a globally distributed target: glaciers. IEEE Trans. Geosci. Remote Sens. 38, 1105-1112.
Tucker, C.J., Grant, D.M. and Dykstra, J.D. (2004): NASA's global orthorectified Landsat data set. Photogrammetric Engineering and Remote Sensing, 70, 313-322.
Woodcock, C.E., Allen, R.G., Anderson, M., Belward, A., Bindschadler, R., Cohen, W.B., …, Wynne, R. (2008): Free access to Landsat imagery. Science, 320(5879), 1011.
Wulder, M.A., Masek, J.G., Cohen, W.B., Loveland, T.R., & Woodcock, C.E. (2012). Opening the archive: How free data has enabled the science and monitoring promise of Landsat. Remote Sensing of Environment, 122, 2-10.
Wulder, M.A., White, J.C., Loveland, T.R., Woodcock, C.E., Belward, A.S., Cohen, W.B., Fosnight, E.A., Shaw, J., Masek, J.G., Roy, D.P. (2016): The global Landsat archive: Status, consolidation, and direction. Remote Sens. Environ. 185, 271-283.
