Multi-Sensor Analysis of Predicted and Observed Glacier Instabilities in the Hissar–Alay of Central Asia

Highlights

What are the main findings?

• Dedicated processing of multi-sensor satellite data finds a much higher prevalence of unstable glacier flow than previously known in the Hissar–Alay mountain range.
• A rediscovered statistical model by Soviet glaciologists from 1980 correctly predicted the high prevalence but not the identification of individual unstable glaciers.

What are the implications of the main findings?

• Dedicated reprocessing of multi-sensor raw satellite data should be further explored to improve historical reconstructions of glacier dynamics compared with commonly used existing products.
• Historical investigations from the Soviet period, not always known to the international community, are an invaluable but often overlooked resource for glaciological information in Central Asia.

Abstract

Surge-like glacier instabilities in Central Asia remain underexplored, particularly in regions of mild instability or smaller glaciers. In 1980, two leading Soviet glaciologists proposed a classification method (GS1980) to calculate the spatial distribution of “pulsating” glaciers in the Hissar–Alay range, predicting a 20% prevalence of unstable flow and claiming highly accurate detection. These findings were unconfirmed in subsequent studies, which typically reported fewer than 10 surge-type glaciers in the region. Here, we address this discrepancy by reassessing the GS1980 predictions using a newly compiled multi-sensor satellite dataset covering nearly six decades. We systematically examine glacier dynamics in the region, assessing ice flow instabilities from changes in terminus position, ice thickness, and surface morphology. We identify 171 glaciers that exhibit pulsating behavior, corresponding to 25% of the sample—in broad agreement with GS1980. Flow instabilities tend to be modest in scale, with slow advances and long active phases (mean duration of 14 years). We find that the GS1980 model shows some ability to distinguish pulsating from stable-flowing glaciers; however, its predictive power is lower than claimed due to the simplifying assumptions of its morphology-based approach and the uncertainties in the input data. Our results indicate that pulsations in the region are more widespread than previously reported, but fall at the weaker end of the spectrum of glacier instability, which may not be well represented by a sharp binary classification (surge-type versus stable). As more detailed satellite records become available, we suggest that a more nuanced framework may be useful to recognize and interpret subtler instabilities of small glaciers.

1. Introduction

Glacier surges are loosely defined as quasi-periodic episodes of accelerated ice flow, driven by internal instability, which quickly redistribute ice from an upper reservoir to a lower receiving area [1,2,3]. Such events alternate with longer phases of quiescent recovery, where the glacier flows slower than balance velocity and reverses the geometric changes of the active phase [4]. Traditional definitions of glacier surging invoke a 10× speed-up threshold (e.g., [2]), a binary classification of surge mechanisms into Alaskan (hydrology-driven) and Svalbard (thermal-driven) type [5,6,7,8], a strict distinction between surge-type and stable glaciers (e.g., [9,10]), and an often-quoted figure of 1% for the overall global prevalence of surge-type glaciers [11]. However, following the earliest descriptions of the phenomenon, subsequent observations indicate a more complex picture, including smaller speed-up magnitudes [12,13,14,15], spatio-temporal patterns not fitting either Alaskan or Svalbard models [13,16,17], and more widespread prevalence of unstable flow than expected (e.g., [18]). It is generally accepted that a continuous spectrum of flow instabilities exists between the stable and surge-type end-members [3,19,20,21,22]. Moreover, climatic forcing and mass balance can affect surge occurrence and distribution [22,23,24], shape the characteristics and duration of flow instabilities [25,26,27,28,29], and possibly promote surge initiation [30,31]. Overall, it is non-trivial to achieve a robust definition of glacier surges and a clear distinction from mass balance-driven advances [3,17,22].

However, the indicators used to detect such glacier instabilities are quite well established: commonly mentioned are the following (e.g., [2,4,19,32]):

• Contrasting patterns of ice thickness change between the reservoir and receiving areas.
• An increase in ice velocity due to enhanced sliding, often leading to plug-like flow.
• Rapid terminus advance.
• Geomorphological signs such as distorted or looped moraines, sheared margins, increased and expanded crevassing, visible fronts, and kinematic waves traveling along the glacier surface.

These indicators have been increasingly used in remote sensing studies to compile inventories of surge-type glaciers at the regional and global scale (e.g., [8,16,33,34,35,36,37]). Such inventories are often derived by analysis of a few satellite datasets (most commonly Landsat and the Advanced Spaceborne Thermal Emission and Reflection Radiometer—ASTER), thanks to their multi-decadal global coverage and standardized data products. Other Earth-observing satellites such as the Satellite Pour l’Observation de la Terre (SPOT) and RapidEye notably provide higher resolution, but due to their raw data format, heterogeneous spatio-temporal coverage, or other access barriers, they have so far seen more limited usage. However, they were found to significantly outperform the Landsat and ASTER products to monitor glacier dynamics, for example, at the relatively small and slow Abramov glacier [29].

Moreover, while glacier surges have been tabulated at the regional and global level via manual inspection and automated classification, comparably few studies have formulated predictive models of their spatial occurrence. In [38], the authors presented a logistic distribution model of surge-type glaciers in Svalbard based on geological boundaries, mass-balance conditions, and thermal regime. In [24], the authors examined and predicted the global prevalence of surging activity with Maxent (also logistic) based on a compilation of published surge inventories. In [39], the authors experimented with several machine learning models to classify surge-type glaciers in Svalbard based on geometry, climatic mass balance, and runoff.

Here, we describe and evaluate an earlier attempt at predicting the spatial distribution of glacier flow instabilities in Central Asia. The study ([40], hereafter GS1980) constitutes the earliest published effort at this kind of spatial modeling; however, to this day, it was not well known outside Soviet glaciology. The authors compiled a dataset of “pulsating” and stable-flowing glaciers, enumerating their morphological properties according to inventory data. Then, they used these properties to train a non-parametric Bayesian classification model ([41]; Section 3.1.1, Appendix A). Through leave-one-out validation, they claimed a highly accurate classification skill, with up to 87% and 69% of correct predictions for stable and pulsating glaciers, respectively. Finally, they used the classifier to calculate the probability of pulsation for glaciers in the Hissar–Alay mountain range (Section 2), predicting the occurrence of 194 pulsating glaciers over the study area.

We conduct a systematic re-examination of glacier dynamics in the Hissar–Alay in order to validate the 45-year-old predictions of GS1980. We first inspect existing inventories of unstable glaciers in the region, both from the Soviet period and from recent studies, finding highly incomplete coverage and multiple mismatches (Section 2). Then, we compile a new multi-sensor satellite dataset covering the region at higher spatio-temporal resolution than provided by Landsat and ASTER products (Section 3.1). On this dataset, we investigate changes in glacier length, thickness, and geomorphology over the past 60 years, detecting the signs of glacier pulsation (Section 3.2). Based on commonly used identification criteria [18,34,35], we classify glaciers as pulsating or stable flowing. We then use our observations to verify the predictions of GS1980 and assess the overall performance of that early classification model (Section 4). Of note, the authors adopted a broad definition of “glacier pulsation”, specifically accommodating a wide spectrum of glacier instabilities—including “soft” modes with low peak velocities and limited mass redistribution, and forgoing strict thresholds on the observed glacier changes. To reflect this, here we choose to maintain the “pulsation” terminology instead of “surge”, while acknowledging the significant overlap between the two terms and the challenges of a rigorous definition.

2. Study Region

The Hissar–Alay mountain range (Figure 1) is located at the extreme south-western edge of Tien Shan, close to the northern margin of the Pamirs, in the headwaters of the Syr Darya, Amu Darya, and Zeravshan rivers. Shared between Uzbekistan, Tajikistan, and Kyrgyzstan, the mountain range stretches west to east over about 600 km, with an average north–south extent of 80 to 150 km. It is bounded by the Fergana valley to the north, the Alay valley to the south, the Karshi steppe to the west, and the Fergana range to the east. Three main ridges (Turkestan, Zeravshan, and Hissar) rise from the plains at the western end of the range, converging into the Alay ridge at the so-called Matcha glaciation node. The highest peaks, located in the central area, reach an altitude of 5600 m asl.

The Hissar–Alay is located in the broader arid–semiarid region of Central Asia, characterized by a generally continental climate [42,43]. However, in winter and spring, the Siberian anticyclonic and southwest cyclonic circulations bring relatively warm and moist air masses to the area. Therefore, winter and annual precipitation are slightly more abundant than in the surrounding regions [43,44]. Precipitation increases from 350–700 mm yr⁻¹ in the lowlands to more than 1500–2500 mm yr⁻¹ above 3000 m; the given ranges correspond to a marked west–east decreasing gradient [43,45]. Accordingly, the mean annual snowline on the western Turkestan ridge was identified in the 1970s at a height of about 3600 m asl, while in the east and south-east direction, it rises to more than 4300 m asl [46].

The Hissar–Alay range features widespread, unevenly distributed glacierization ([47]; Figure 1). The first complete glacier inventory of the region was compiled from aerial photographs in the late 1950s, indicating a total extent of 2184 km² [47]. A second mapping effort in the late 1970s, based on space imagery, revealed a decrease to 1839 km² and a 15% increase in moraine cover [48,49]. According to the Randolph Glacier Inventory version 7.0 (RGI7.0, corresponding to about year 2000; [50]), the area encompasses 5058 glacier polygons, for a total surface of 1912 km². The apparent area increase since the 1970s does not reflect an actual glacier expansion and could be attributed to different interpretation during the compilation of the inventories, especially concerning debris-covered glaciers (also see Section 3.2.1 and Figure 2). Total ice volume was estimated at 105 km³ in 1957 and 88 km³ in 1980, using the average result of four empirical formulas [49,51].

Several individual glaciers of the Hissar–Alay were monitored in situ over the second half of the 20th century, including Abramov glacier with a resident program over 1967–1999 [52]. At present, annual glaciological observations have been re-established at several sites, and Abramov is a reference glacier with the World Glacier Monitoring Service (WGMS; [53,54]). Those in situ measurements indicate a consistent pattern of negative mass balance since at least the 1960s, with almost no year of mass gain. At Abramov glacier, total mass loss exceeded 20 m w.e. between 1973 and 2023 [29,55,56,57]. Such a negative trend was confirmed at the regional scale by multiple studies, including geodetic assessments [58,59,60], snowline-constrained modeling [61], and geo-statistical down-scaling of multi-source observations [62].

Besides mass balance, glacier dynamics in the region have long attracted attention, first with the recording of glacier terminus positions during early 20th-century explorations [63,64,65,66], and later (since the 1960s) with the compilation of glacier inventories and the discovery of intense glacier pulsations in the nearby Pamir mountains [67,68]. As such, several early compilations of pulsating glaciers exist covering (in some cases partially) the Hissar–Alay [68,69,70]. They are based on observed active phases of pulsation, and/or inferred geomorphological evidence from ground surveys and aerial imagery. More recently, several remote sensing studies have tried to identify surge-type glaciers at the global or High Mountain Asia (HMA) scale, also including the Hissar–Alay area. These studies have applied a variety of manual and automated recognition methods on dense global datasets such as Landsat imagery, ASTER Digital Elevation Models (DEMs), ITS_LIVE ice velocity grids, or Sentinel-1 radar backscatter products [8,31,34,35,71]. On average, recent studies have detected fewer than five glacier pulsations in the Hissar–Alay over the past few decades (Section 5.4). The results usually do not match across the studies; overall, the region is considered to have a low prevalence of unstable ice flow. We note that data about pulsation of about five glaciers were integrated from the early compilations into the RGI7.0 and the Fluctuations of Glaciers (FoG) datasets; however, we found several index mismatches resulting in incorrect identification (Section 5.4).

3. Materials and Methods

3.1. Data

3.1.1. List of Predicted Pulsations by GS1980

The GS1980 study presents the predicted distribution of pulsating glaciers in the Hissar–Alay, based on a Bayesian classification model [41]. The model (detailed in Appendix A) was trained on a sample of 32 pulsating and 31 steady-flowing glaciers. Then, it was applied to calculate the probability of pulsation of all glaciers larger than 1 km² (N = 935), based on their morphological parameters. The authors tested several combinations of such morphological parameters [19], finally identifying the following three as the most effective to reproduce the observed training sample:

• K, the glacial coefficient (ratio of accumulation to ablation area extents);
• C, the ratio of accumulation area extent to mean width of the tongue (the latter being defined as the area between the firn line and the glacier terminus);
• $I_T$, the average surface slope of the tongue.

These parameters were derived from inventory data (Section 3.1.2), manually compiled from maps and aerial surveys according to a set of guidelines [72]. Based on the morphological parameters, GS1980 provides a list of 194 Hissar–Alay glaciers that are predicted to be of the pulsating type (probability > 50% according to the classification model). Their probability of pulsation was also calculated a second time with a simpler, two-parameter version of the model, forgoing surface slope $I_T$. Additional topographic information on each glacier, such as area and length, is also reported in the list.

3.1.2. Glacier Inventories

The following three independent glacier inventories, systematically describing glacierization of the region, are relevant for our study:

• The Union of Soviet Socialist Republics (USSR) glacier inventory, covering the whole area of the former Soviet Union [73,74,75,76];
• The database of A.S. Shchetinnikov, focused on the Pamir and Hissar–Alay regions [47,77,78];
• RGI7.0 [50].

The USSR glacier inventory was published between 1965 and 1982. It consists of 108 sections within 69 books, and upon completion, it was claimed to be the first comprehensive country-level inventory of glaciers [74]. It is organized in a three-level geographical subdivision (volume, issue, and section), made according to drainage basins. The Hissar–Alay region is covered in issues 1 (“Syr Darya”) and 3 (“Amu Darya”) of volume 14 (“Central Asia”). The source data for the compilation of those issues consist of aerial photographs acquired between the 1940s and the 1960s (mostly in 1957: [78]). Each section of the USSR inventory contains a table of individually numbered glaciers, described with their attributes: name, catchment, morphological type, and geometrical parameters such as aspect, length, surface area, estimated ice volume, and minimum/maximum elevation. Also included are schematic maps with the numbered outlines of all listed glaciers, the main rivers, and the watershed lines, enabling unambiguous (albeit manual and time-consuming) localization of all glaciers.

Between 1978 and 1980, Soviet glaciologist A.S. Shchetinnikov compiled a new inventory of glaciers covering both the Pamir and the Hissar–Alay. This database (hereafter referred to as Sh1980) was based on analog satellite images acquired over the same years [78]. It includes a set of tables with glacier indices and morphological attributes, but no figures or maps. A look-up table attempting to match the indices of Sh1980 to the USSR glacier inventory was also compiled, albeit with a large number of discrepancies and mismatches [78].

Finally, RGI7.0 is a database describing the state of global glacierization around year 2000. The Hissar–Alay is included in region 13 (“Central Asia”); there, all glacier outlines are derived from the Glacier Area Mapping for Discharge from the Asian Mountains (GAMDAM) glacier inventory version 2 [79]. In our study region, the source data are Landsat images mostly acquired in 2002. Glacier outlines were delineated manually due to the high prevalence of debris-covered glaciers. In RGI7.0, each glacier polygon is assigned a surge_type index, belonging to the categories “observed”, “probable”, “possible”, and “no evidence”. The main data sources for surging information are the inventory compiled by [24] and the multi-factor remote sensing study of [34].

The three mentioned inventories are independent in the identification of individual glaciers and enumeration of their morphological properties; correspondence between the entries is in general not one-to-one (see Section 3.2.1).

Glaciers on the GS1980 list are derived from the USSR glacier inventory, except for three sections of the latter (numbers 1, 2, and 5 of issue 3), which were not yet published by 1980; thus, those were replaced by those from the Sh1980 database [40]. The GS1980 list also provides the geometric parameters of each glacier and the corresponding values of the predictors (Section 3.1.1) used by the classification model.

3.1.3. Optical Satellite Archives

In our study, we use optical satellite imagery and DEMs to look for the indicators of glacier pulsation. Such an analysis is usually carried out on the dense, analysis-ready datasets of Landsat and ASTER. However, ref. [29] found that such datasets struggle to capture the dynamics of the reference Abramov glacier, which is the fifth largest of the entire Hissar–Alay. Thus, we gathered raw optical imagery from several alternative satellite archives; we assembled a new dataset with full spatial coverage of the study region and a repeat interval of a few years, improving on the pixel count of existing Landsat products by one to two orders of magnitude. Specifically, we acquired a total of 68 monoscopic scenes from the SPOT 1 to 5 satellites [80], distributed via the Spot World Heritage (SWH) program [81] and covering the period 1986–2007, and 11 scenes from the RapidEye constellation [82], made available by Planet Labs’ Education and Research Program [83] and covering 2009–2015. SPOT 5 and RapidEye provide a ground sampling distance of up to 5 m at nadir, while the earlier SPOT 1–4 achieved 10 m from 1986 to 2003. Moreover, we selected 14 image pairs from the High Resolution Stereoscopic (HRS) instrument on board SPOT 5, providing stereo coverage at a resolution of 5 × 10 m (along-track supersampling) between 2003 and 2006. To extend coverage before the digital sensor era, we included 13 declassified scans of stereo pairs from film-based reconnaissance satellites: KH-4 CORONA Panoramic Camera (PC) and KH-9 Hexagon Mapping and Panoramic Camera (MC and PC, respectively) (

Table A1. Declassified scenes from reconnaissance satellites used in our study, listed by chronological order of acquisition.

Date	Sensor	Identifier
18 August 1968	KH-4 PC	DS1104-2169DF090–093
		DS1104-2169DA096–099
15 September 1971	KH-4 PC	DS1115-1072DF185
		DS1115-1072DA191
16 July 1973	KH-9 MC	DZB1206-500007L015001
		DZB1206-500007L016001
3 August 1973	KH-9 MC	DZB1206-500080L016001
		DZB1206-500080L017001
		DZB1206-500080L018001
		DZB1206-500080L019001
22 November 1973	KH-9 MC	DZB1207-500041L001001
		DZB1207-500041L002001
		DZB1207-500041L003001
25 July 1980	KH-9 PC	D3C1216-200279F036–039
		D3C1216-200279A037–040
20 August 1980	KH-9 MC	DZB1216-500273L006001
		DZB1216-500273L007001
		DZB1216-500273L008001

), obtained from the United States Geological Survey [84,85,86]. Further technical details of these sensors are provided in [29], where we already processed a subset of data from the same satellite archives in the region of Abramov glacier.

3.1.4. Analysis-Ready Data

To further improve spatio-temporal coverage of the study region, we integrated in our dataset several analysis-ready or previously processed satellite products. Specifically, we included orthoimages from the Sentinel-2 Multi-Spectral Instrument (MSI) at 10 m resolution [87]; we downloaded all tiles acquired over the Hissar–Alay at the end of the summer seasons of 2018, 2021, and 2024, achieving full cloud-free coverage of the study area in those years. We also incorporated the dataset of orthoimages and DEMs produced by [29], covering the central region of Abramov glacier (Figure 1) at annual intervals between 1996 and 2020 from SPOT, RapidEye, Pléiades, and the Indian Remote Sensing (IRS-1C/D) program [88,89]. In order to better estimate ice thickness changes, we downloaded several existing global DEMs at 1 arc-second resolution: NASADEM [90], ALOS World AW3D30 [91,92], and Copernicus GLO-30 [93]. They provide full coverage of the study region, but suffer from radar penetration in accumulation areas (NASADEM, GLO-30; [94,95,96]) or from mosaicing over multi-year acquisition periods (AW3D30, GLO-30). Finally, we included nine DEMs from the High Mountain Asia DEM (along-track and cross-track pairs: HMA DEM AT/CT, [97,98]). These optical DEMs cover limited regions (encompassing only a few glaciers), but provide a high spatial resolution of 8 m and well-defined acquisition times (between 2007 and 2015 in our subset).

3.2. Methods

Our study started with the identification of glaciers of interest by combining and matching Soviet-era glacier inventories with RGI7.0 (Section 3.2.1). Then, we processed multi-sensor raw optical satellite data (Section 3.2.2), compiling a standardized glacier-wise dataset of orthoimages and ice thickness change maps (Section 3.2.3). Subsequently, we detected the occurrence of glacier pulsations (Section 3.2.4) and calculated the performance of the GS1980 classifier (Section 3.2.5). Finally, we performed a spatio-temporal analysis of the identified pulsations (Section 3.2.6).

3.2.1. Identification of Glaciers of Interest

To verify the predictions of GS1980, it was necessary to identify the glaciers appearing in their published table. We decided to match the GS1980 entries with the contents of RGI7.0 and to use the latter as the base for our analysis. The RGI is used by virtually all regional and global studies of glacier surges; its standardized, quality-controlled dataset of digital outlines is an effective starting point for further geo-processing and analysis.

The GS1980 list of pulsating glaciers is based on two different inventories (Section 3.1.1 and Section 3.1.2); thus, we first matched the entries of the Sh1980 list to the USSR glacier inventory, whose maps enable localization of the glaciers of interest. We found several mismatches in the look-up table between the Sh1980 list and the USSR inventory. We corrected such indexing errors using the supplementary data found in both lists—namely, glacier names and topographical parameters, which allowed us to achieve an unambiguous, one-to-one match for all 63 Sh1980 indices appearing in the GS1980 list.

After retrieving all indices of the GS1980 list within the USSR inventory, we matched those glaciers to the polygons of RGI7.0. It should be noted that the USSR glacier inventory was already digitized and integrated into the National Snow and Ice Data Center’s Eurasian Glacier Inventory (NSIDC EGI) and subsequently into the World Glacier Inventory [99,100]. However, given the aforementioned indexing inaccuracies, as well as inconsistencies found during the compilation of the EGI, we rather opted to unambiguously match glaciers based on their actual geometry, using the shape and overlap of each outline. To this end, we first geo-referenced the sheets of the USSR inventory encompassing the Hissar–Alay (

Table 1. Geo-referenced maps from the USSR glacier inventory.

Volume	Issue	Section	Map Numbers	Notes
14	1	9	22, 23, 26–28, 30, 33, 35	Northern slope of Alay ridge
14	1	10	14–18	Northern slope of Turkestan ridge
14	3	1/2	10–14, 16, 18, 20	Two sections in a single book; Zeravshan ridge
14	3	4	14	Southern slope of Hissar ridge: Western half
14	3	5	10, 12	Southern slope of Hissar ridge: Eastern half
14	3	6	30–32, 34–37	Southern slope of Alay ridge: Western half
14	3	7	18, 20, 21	Southern slope of Alay ridge: Eastern half

). We geo-referenced a total of 34 maps from 8 sections, using manual ground control points (GCPs) and a Thin-Plate Spline transform. Then, we inspected each of the 194 entries of GS1980, matching them to the RGI outlines based on spatial overlap (Figure 2). To our knowledge, this is the first systematic comparison at the regional scale between RGI7.0 and the USSR inventory, which was compiled from higher-resolution data and with the contribution of several dedicated ground surveys. Because of the different aggregation choices between the USSR inventory and RGI7.0, the relationship between the inventories is not one-to-one: we found either one or several RGI7.0 polygons corresponding to each GS1980 entry, and conversely, instances of multiple USSR inventory entries corresponding to the same RGI7.0 polygon (Figure 2). Such a many-to-many relationship would significantly complicate analysis of the GS1980 predictions. Thus, to enable further processing, we converted the matches into a one-to-many relationship: we aggregated the spatially contiguous glacier polygons of RGI7.0 into larger complexes, until each GS1980 entry corresponded to at most one such complex. As a result, we obtained 169 aggregated polygons out of 226 original RGI7.0 ones. In the following, we label those 169 polygons as “predicted pulsating complexes” (PPCs): each corresponds to one or more glaciers predicted to pulsate by GS1980. We note that this aggregation affects only a minority of our dataset: 80% of the 169 PPCs are composed of a single RGI7.0 polygon, and just six include more than three polygons.

These RGI-derived PPCs constitute the basis for our inspection of recent glacier dynamics (Section 3.2.3). According to GS1980, all other glaciers in the region larger than 1 km² are classified as stable-flowing. In order to check that prediction, we also inspected all such glaciers. The RGI7.0 (GAMDAM) outlines are split according to ice divides [79,101], potentially into polygons that do not meet the 1 km² size threshold. As such, we first aggregated all spatially contiguous RGI7.0 polygons not already identified within the PPCs; then we selected the resulting polygons larger than 1 km². We term those “non-pulsating complexes” (NPCs). Such an aggregation also improves the effectiveness of our analysis by reducing instances of repeated inspection of the same area. Overall, between the PPCs and the NPCs, we compiled a list of 411 items to be inspected, aggregated from 678 individual RGI polygons.

3.2.2. Processing of Raw Satellite Archives

To inspect flow dynamics of the selected glaciers, we compiled a new dataset of orthoimages and DEMs. We processed the raw satellite archives (Section 3.1.3) using specific pipelines adapted to each data product. Our methods to produce orthoimages out of monoscopic SPOT and RapidEye scenes are a further development of the workflow introduced in [29]. Here, we present the main steps and the new advances that we have since implemented; additional technical details are discussed in depth in that previous study.

For SPOT, we orthorectified the raw scenes over the NASADEM terrain using the rigorous sensor model (RSM) of [102,103], as implemented in the open-source geoCosiCorr3D package. The method involves refinement of image geometry using GCPs to correct inaccuracies of the parameters reported in the scene metadata. Claimed orthorectification accuracy is up to 1/20th of pixel size [104]. Compared with the study of [29] over Abramov glacier, we did not have a high-resolution Pléiades orthoimage to use as reference over the entire region of interest. Thus, for that purpose, we assembled a cloud-free Sentinel-2 mosaic from satellite passes on 1, 6, 7, and 9 September 2018. We co-registered the individual Sentinel-2 tiles, applied a simple radiometric matching (linear least square fit over the respective overlap regions), and finally stitched the images into a mosaic covering the full region. Then, for each raw SPOT scene, we first refined the footprint reported in the metadata by running interest point matching (image_align tool within the NASA Ames Stereo Pipeline (ASP; [105]). This step found and corrected inaccuracies up to 100 km in the location reported for SPOT 1-4 HRV/HRVIR scenes, and up to 10 km for SPOT 5 HRG. Compared with [29], we bilinearly reprojected to Universal Transverse Mercator (UTM) the raw SPOT scenes (originally provided with rotated longitude/latitude coordinates), which significantly improved the matching step. After computing the refined scene footprint, we produced an ad hoc Sentinel-2 reference for tie point detection of each SPOT scene, which again improved the matching. As such, we overcame the orthorectification issues of [29] and could generate SPOT orthoimages without having to optimize parameters of the ASP tools for each scene. As such, the updated pipeline is on average 50 times more computationally efficient than the previous version. We also found it more robust to changes in image resolution, cloud cover, and terrain surface, successfully processing SPOT 1 scenes more than 30 years older than the Sentinel-2 reference. Finally, the improved result also allowed us to shorten the loop used by geoCosiCorr3D to refine image geometry: we implemented termination of the optimization loop upon hitting both thresholds of 0.01 and 0.5 pixel for mean signed error and root mean square error, respectively.

RapidEye scenes were orthorectified by means of the mapproject tool from NASA ASP based on the rational polynomial coefficient (RPC) information provided within each product. We used geoCosiCorr3D to apply fine registration of each orthoimage to the Sentinel-2 reference by subtracting the median residual displacement along the X and Y coordinates over stable terrain (defined via the RGI7.0 polygons). We found this registration method to be more robust and computationally efficient than our previous approach (interest point matching with singular value decomposition), which would usually require fine-tuning of the parameters to achieve accurate results [29].

Stereo scenes from reconnaissance satellites were processed into orthoimages and DEMs according to the methods of [106,107,108,109]. Technical details of those, as well as the stereo processing of SPOT 5 HRS pairs, are unchanged compared with [29].

Finally, we produced mosaics with all the orthoimages acquired on the same day by the same sensor: after fine registration with geoCosiCorr3D, we stitched them both along-track (for longer acquisitions not fitting a single scene) and cross-track (for tandem scenes acquired by the twin instruments on board SPOT satellites). Mosaicing provides uninterrupted coverage of glaciers located along the scene borders. Overall, our dataset contains orthoimages at 68 different dates between 1968 and 2024. We applied a similar mosaicing to the dataset of DEMs, the only major difference being in the co-registration step that we applied with xdem [110,111].

3.2.3. Glacier-Wise Standardization

In order to examine regional glacier dynamics effectively and consistently, we implemented a new glacier-wise standardization procedure for our heterogeneous and multi-sensor satellite dataset. For each PPC and NPC (Section 3.2.1), we first found the orthoimages providing coverage of the glacier terminus (defined as the lowest glacierized point on the GLO-30 DEM). Then, from all such images we extracted a standard-sized crop surrounding the glacier outline, with a 1 km margin on all sides (but constrained between 1.5 and 4 times the glacier extent, to accommodate the largest and smallest glaciers). We performed fine geoCosiCorr3D registration of all such crops to the same reference (Sentinel-2 mosaic from 2018), then bilinearly resampled them to 5 m resolution, and finally applied Contrast-Limited Adaptive Histogram Equalization (CLAHE; [112,113]) to normalize and standardize pixel intensities. We saved each resulting image with an overlay of the respective RGI7.0 outline(s) to provide a fixed reference towards evaluation of glacier terminus changes. The result is a dataset of images whose coverage, nominal resolution, and visual appearance are consistent and uniform for each glacier, enabling effective inspection of the dynamics.

We applied a similar standardization pipeline to the dataset of DEMs in order to assess ice thickness changes for each PPC and NPC. We bilinearly resampled all DEMs to the same resolution of 30 m, striking a balance between sufficient detail on small glaciers and reduced noise (thanks to downsampling) on data from declassified imagery [29]. To ensure robust processing, we discarded DEM pairs having valid data coverage below 10% of stable terrain and 50% of glacierized area; then we co-registered all remaining pairs [111] and computed the respective DEM differences and their uncertainties [114]. We removed outliers from all differences with the local hypsometric filtering of [115]; while the filter is designed to preserve the elevation change signal of an altitude-dependent mass balance profile, we found it to be effective also the in case of the observed glacier pulsations, since they redistribute mass longitudinally down-glacier—thus generally preserving spatial coherence within relatively small (50 m) elevation bands.

3.2.4. Detection of Pulsation Events

We systematically inspected the standardized datasets to detect glacier pulsations. We focused specifically on direct observation of active phases of unstable flow (and the recovery phases in between), excluding relict signs that might be only suggestive of past pulsations. We simultaneously examined three indicators of ice flow instability (Figure 3): (1) terminus behavior, (2) surface morphology on the orthoimages, and (3) ice thickness changes on the DEM differences. The criteria applied for the identification of active phases across the three indicators are detailed in

Table 2. Criteria for the identification of the active phase of surge-like instabilities.

Indicator	Identification Criteria
Terminus behavior	Sudden switch from terminus retreat to advance, larger than pixel uncertainty
	Multi-year advance, asynchronous compared with similar neighboring glaciers
	Downwasting terminus overrun by an active front
Surface morphology	Actively deforming moraines
	Expansion of crevassed areas by multiple times
	Appearance of sheared margins
Thickness changes	Downstream mass redistribution, spatially coherent and exceeding the uncertainty of thickness change by a factor of two or more

.

For thickness changes, we identified downstream mass redistribution qualitatively rather than with a quantitative classification, because the actual spatial patterns observed in DEM differences are highly dependent on glacier shape, the presence and flow of surface features (such as séracs or depressions), and the surface mass balance over the time interval of the DEM difference [18,35,116]. In all the retained cases, the mass transfer signal was readily apparent, with thickness change exceeding (by a factor of two or more) the respective uncertainty. We classified glaciers as pulsating based on simultaneous evidence from at least two out of the three indicators (e.g., [34]).

We additionally ranked pulsating glaciers into tiers according to the confidence of the detection. We used the “confirmed” class for glaciers identified as pulsating by all three indicators simultaneously (rather than only two). We also used that class for glaciers where we observed the recurring manifestation of two or more active phases (rather than a single one), separated by multi-year intervals of quiescence. We used the “probable” class for all other identified glaciers; we did not use a lower class such as “possible”, since it is usually reserved for glaciers whose unstable flow is deduced by indirect observation of relict surface features [24].

While our processing is based on PPCs and NPCs (Section 3.2.1), we still highlighted the dynamics of each RGI7.0 component individually. During this work, we also inspected an additional 18 RGI7.0 outlines, fortuitously located within the selected areas of interest: these outlines were not part of any PPC or NPC, but showed indications of unstable flow according to the criteria outlined above. This brought the total number of examined RGI7.0 glacier polygons to 696, for a total glacierized area of 1263 km²—about two-thirds of the RGI7.0 surface in the Hissar–Alay.

We assessed the reliability of our workflow for identification of pulsations by running a second independent examination on a random sample of 50 glaciers out of the 696 considered ones. We determined the agreement between the two examinations by calculating Cohen’s $\kappa$ statistic [117], as follows:

$$\kappa ={\displaystyle \frac{p_o-p_e}{1-p_e}}$$

(1)

where $p_o$ is the agreement rate between the two examinations, and $p_e$ is the hypothetical probability of agreement by chance. The coefficient value is 1 in case of perfect agreement between the examinations; it is 0 if the agreement can be fully explained by chance.

3.2.5. Performance of the Predictive Model

We calculated the performance of the GS1980 predictive model by comparing the counts of correct and incorrect classifications (true/false positives and negatives: TP/FP/TN/FN). Such counts are not trivial to define because of (1) the one-to-many relationship between the USSR glacier inventory and the RGI, (2) the occurrence of pulsations on only some areas or branches of a glacier, and (3) the choice by GS1980 to report only glaciers predicted to pulsate, which somewhat obfuscates the identification of predicted stable glaciers. As such, we tested two different methods of counting the classified glaciers, reflecting glacier aggregation according to (1) RGI7.0 and (2) the USSR inventory, respectively. In version (1), we count as TP an RGI7.0 polygon that respects the following: we found it to be pulsating, and it belongs to one of the PPCs (Section 3.2.1). An RGI7.0 polygon is counted as FP if it belongs to a PPC, but we did not find it to pulsate. By contrast, in version (2), an RGI7.0 polygon is counted as TP even if we did not specifically observe it to pulsate, as long as we found pulsation of at least one other polygon belonging to the same PPC. Under this definition, a non-pulsating RGI7.0 polygon is an FP only if no glacier within the PPC is pulsating. Definition (1) always yields a lower count of TPs than version (2), thus a lower calculated performance of the predictive model. With both counting methods, FN and TN are defined in the same way: an FN is an RGI polygon that we found to pulsate while it is not included in any PPC, while TN is applied to all RGI7.0 polygons (of the 696 examined) not falling into another category. We note that the majority of PPCs are actually composed of a single RGI polygon; thus, the difference between the two counting methods is relatively minor (Section 4.3).

We applied these two counting methods to both versions of the GS1980 predictions—with two and with three predictors (Section 3.1.1). Thus, we obtained four independent counts of TP/FP/TN/FN. We measured the performance of these results with standard metrics commonly used for classifiers [118], including recall, specificity, and standard accuracy, as well as metrics specifically accounting for class imbalance in the dataset: balanced accuracy, precision, and F1 score (Appendix C). The latter metrics are especially meaningful in our case: even though the GS1980 study performed leave-one-out validation on the almost balanced training dataset (Section 1), it later found a significant difference in prevalence between pulsating and stable glaciers in the region (194 and 741 out of 935, respectively). Formulas for all computed metrics are given in Appendix C. Finally, we also compared the performance of GS1980 with that of uninformed random guessing (assuming equiprobable classes) as well as stratified random guessing (assuming rates of pulsation prevalence as found by GS1980 and by our study).

3.2.6. Spatio-Temporal Analysis

We examined the spatio-temporal distribution of the identified glacier pulsations to compare with the GS1980 predictions as well as with results from other studies in the region and in the vicinity. We defined the temporal bounds of single pulsation events by selecting the periods corresponding to positive identification criteria for unstable flow (Section 3.2.4).

As such, in several cases, the start/termination of a pulsation is coincident with the first/last evidence of terminus advance. However, to accommodate multi-phase pulsations and seasonal dynamics (due to the inconsistent orthoimage dates), we relaxed the requirement of continuous terminus advance over individual years: we considered as a single pulsation all instances of interrupted advance resuming within a year.

Despite the use of multiple satellite sources, our dataset still has limited temporal coverage, especially in the earliest period (until 1985). This results in multi-year intervals between consecutive observations, complicating the accurate definition of start and end dates of pulsations. Thus, we calculated three estimates for the duration of each pulsation—low, mean, and high. Given a pulsation observed over a few consecutive intervals, the “mean” estimate of duration assumes that the pulsation starts on the same date as the satellite pass at the start of the first interval, and terminates exactly at the satellite pass marking the end of the last interval. The “low” estimate assumes a pulsation starts halfway through the first interval of observed instability and ends halfway through the last interval. Finally, the “high” estimate assumes a pulsation starts halfway through the interval preceding the first unstable one, and similarly extends halfway into the next interval after the last observed signs of unstable flow.

4. Results

4.1. Overview

We found evidence of pulsating behavior at 171 of the 696 examined RGI7.0 glaciers, corresponding to a prevalence of 25%. For the majority of glaciers, we could confirm a single pulsation event within their observed time span; however, we also detected two pulsations at 31 glaciers, and three pulsations at three glaciers, for a total of 196 recorded pulsation events. Of those, 128 (65%) are fully contained within our dataset, that is, not already (or still) ongoing at the start (or end) of the covered period. Moreover, 145 events (74%) last longer than a single observation interval (Section 3.2.6); on average, each pulsation can be tracked over three intervals. A total of 103 pulsations are both fully contained and longer than one interval, enabling robust calculation of the event duration. According to the criteria for confidence ranking (Section 3.2.4), we classified 47 glaciers as “confirmed” pulsating and the remaining 124 as “probable”. The reliability assessment of the classification (Section 3.2.4) yielded an agreement (Cohen’s $\kappa$) of 0.86, confirming the high repeatability of the detections and indicating an expected low prevalence of misclassifications.

4.2. Pulsations Distribution and Characteristics

The observed glacier pulsations are distributed very unevenly across the study region (Figure 4). The highest prevalence is found between the central Matcha region and Abramov glacier, and a strong decrease can be seen towards the western and eastern margins of the Hissar–Alay. The largest pulsating glacier is Archa-Bashy (27 km²), while the smallest is a 0.17 km² unnamed glacier in the Turkestan ridge. The median extent of pulsating glaciers is 1.56 km², and 23% are smaller than 1 km². The total extent of the pulsating glaciers is 464 km², or 37% of the surface of all examined glaciers (Section 3.2.4). The largest 26 pulsating glaciers account for half of this extent. There is a significant positive correlation (Spearman’s rho = 0.92) between pulsation prevalence and glacier area, with the former increasing from 1% to 54% between the smallest and largest area classes (<0.25 km² and >5 km²: Figure 5). However, the most extensive glacier complex of the study region (Zeravshan, with 93 km²) is apparently stable in all its branches, as are the largest valley glaciers in the vicinity (Rama and Farakhnau). Among the pulsating glaciers, 159 exhibit terminus advance, while 12 only have internal pulsation with a stable or retreating terminus.

Over the almost 60 years of observations, two periods clearly stand out with an increased occurrence of pulsations: around year 1970 and 2004 (Figure 6). During those years, as many as 17% and 11% of observed glaciers are simultaneously pulsating. However, the sample of examined glaciers is relatively small in the earlier period, with coverage available for only one in three glaciers. As such, the largest number of simultaneously pulsating glaciers occurs in 2004, with a total of 79. The occurrence of pulsations is strongly reduced in the years around 1982 and towards the present, with only about 2.4% and 4.3% of glaciers showing evidence of pulsation, respectively. The mean duration of the active phase is 14 years over 103 pulsations with good observations (Section 4.1, with a standard deviation of 7.5 years (Figure 7a). Low and high estimates (Section 3.2.6) are 9.0 and 18.4 years, respectively. For glaciers with multiple pulsations, we observed a mean interval between terminations of 31.7 years (Figure 7b) and a mean duration of 20.1 years for the quiescent phase. Low and high estimates differ by up to 11 years due to limited temporal coverage over the early period.

4.3. Assessment of the GS1980 Predictive Model

By all computed performance metrics, all versions of the GS1980 predictions are less accurate than claimed by the authors via leave-one-out validation (Figure 8). The discrepancy is especially visible in the metrics that account for class imbalance in the dataset, such as precision and F1 score. However, by the same metrics, the GS1980 model performs significantly better than all versions of random guessing, demonstrating concrete predictive power at discriminating pulsating and stable glaciers. As expected, model performance increases both with the addition of a third predictor (mean slope of the tongue) and when using the more permissive counting of glacier pulsations (based on the USSR inventory rather than RGI7.0: Section 3.2.4).

Overall, the observed prevalence of pulsating glaciers (171/696 = 24.6%) is similar to the predicted one (194/935 = 20.7%). There is no clear pattern in the spatial distribution of model performance across the study region (Figure 9a). However, the model predictions appear to become increasingly accurate with increasing glacier area (Figure 9b), exceeding what would be expected from the correlation between area and pulsation prevalence (Figure 5). The GS1980 model correctly anticipated the high concentration of pulsating glaciers in the central region of the domain, even though the largest glacier complexes (around Zeravshan glacier) are classified as false positives. Finally, the correlation between the predicted probability of pulsation and the observed prevalence (Figure 9c) is quite weak. Spearman’s rho metric is just 0.15; however, it is heavily affected by the absence of observed pulsation at the two glaciers with the highest computed probability (>95%: rightmost class in the figure); excluding those, the same measure of correlation would rise to 0.58.

5. Discussion

5.1. Evaluation of GS1980

Our findings indicate that the GS1980 predictive model has lower accuracy than claimed by the authors, but still some distinct predictive power compared with random guessing (Figure 8). The method is simple and computationally inexpensive (Appendix A); however, we note some serious shortcomings within its application to pulsating glaciers. Specifically, the morphological parameters underpinning the calculation can be highly uncertain and somewhat subjective, depending on the determination of accumulation area, firn line, and glacier terminus from aerial imagery (Section 3.1.1). Measurement uncertainty and subjective interpretation arise in the manual compilation of inventories from just one or few optical acquisitions [40], which may not reflect long-term steady-state conditions over the glaciers. Moreover, the combination of data from different inventories, with heterogeneous reference times (Section 3.1.2), can introduce a mismatch both in the training dataset and in the parameters of the examined glaciers, producing inconsistent results across the study region. Even postulating a rigorous definition of the predictors, additional ambiguities arise with the identification of debris-covered glaciers and the dynamical separation of ice masses at ice divides (Figure 2), again potentially affecting the computed values of the morphological parameters. The problem of dynamical separation appears to have been at least contemplated by the GS1980 authors, who tested the method at the dendritic Egorov glacier (Figure 2) both on individual tributaries and on the single, contiguous complex. While the authors do not discuss the result, the computed probabilities are indeed quite sensitive to the partition of tributaries, ranging between 52% and 80% (compared with 75% for the whole glacier complex). Such a high sensitivity is also confirmed by our analysis of the GS1980 table (Figure A1d). These weaknesses underscore a key shortcoming of the method, namely, the assumption that pulsating behavior is mostly an expression of glacier morphology. This notion was supported by early conceptual models of glacier surging (e.g., [119,120,121]). However, subsequent theoretical developments [7,24] rather attribute surge-like instabilities to a combination of climatic, topographic, and geologic factors, leading to imbalances of glacier enthalpy and thus cyclic instabilities on diverse time scales [21]. While the GS1980 predictors do have an indirect link to climate (via the relative location of the firn line on the glacier), they are too crude to meaningfully represent its relationship with glacier instabilities. Overall, a lesser importance of glacier geometry is consistent with the lower predictive power found for the GS1980 model. Concretely, it is not straightforward to establish a direct link between the model predictors and modern theories of flow instability [7,21,22]; going beyond our study, an analysis of the propensity to enthalpy imbalance for a realistic glacier morphology could yield some insights in that sense.

Our findings indicate that the GS1980 predictive model has lower accuracy than claimed by the authors, but still some distinct predictive power compared with random guessing (Figure 8). The method is simple and computationally inexpensive (Appendix A); however, we note some serious shortcomings within its application to pulsating glaciers. Specifically, the morphological parameters underpinning the calculation can be highly uncertain and somewhat subjective, depending on the determination of accumulation area, firn line, and glacier terminus from aerial imagery (Section 3.1.1). Measurement uncertainty and subjective interpretation arise in the manual compilation of inventories from just one or few optical acquisitions [40], which may not reflect long-term steady-state conditions over the glaciers. Moreover, the combination of data from different inventories, with heterogeneous reference times (Section 3.1.2), can introduce a mismatch both in the training dataset and in the parameters of the examined glaciers, producing inconsistent results across the study region. Even postulating a rigorous definition of the predictors, additional ambiguities arise with the identification of debris-covered glaciers and the dynamical separation of ice masses at ice divides (Figure 2), again potentially affecting the computed values of the morphological parameters. The problem of dynamical separation appears to have been at least contemplated by the GS1980 authors, who tested the method at the dendritic Egorov glacier (Figure 2) both on individual tributaries and on the single, contiguous complex. While the authors do not discuss the result, the computed probabilities are indeed quite sensitive to the partition of tributaries, ranging between 52% and 80% (compared with 75% for the whole glacier complex). Such a high sensitivity is also confirmed by our analysis of the GS1980 table (Figure A1d). These weaknesses underscore a key shortcoming of the method, namely, the assumption that pulsating behavior is mostly an expression of glacier morphology. This notion was supported by early conceptual models of glacier surging (e.g., [119,120,121]). However, subsequent theoretical developments [7,24] rather attribute surge-like instabilities to a combination of climatic, topographic, and geologic factors, leading to imbalances of glacier enthalpy and thus cyclic instabilities on diverse time scales [21]. While the GS1980 predictors do have an indirect link to climate (via the relative location of the firn line on the glacier), they are too crude to meaningfully represent its relationship with glacier instabilities. Overall, a lesser importance of glacier geometry is consistent with the lower predictive power found for the GS1980 model. Concretely, it is not straightforward to establish a direct link between the model predictors and modern theories of flow instability [7,21,22]; going beyond our study, an analysis of the propensity to enthalpy imbalance for a realistic glacier morphology could yield some insights in that sense.

A relatively clear change could be expected for the GS1980 predictor parameters in response to recent accelerated climatic warming, with a marked reduction in accumulation area extent. Such a change would generally lead to a decrease in the computed probabilities (Figure A1d), resulting in a smaller prevalence of predicted glacier pulsations. Continued observations of glacier dynamics in the region would be needed to support this hypothesis.

5.2. Pulsating Glacier Dynamics

Overall, the observed pulsations appear to confirm the “quiet” or “soft” nature of ice flow instabilities in the region [19,20,29,40]; this is characterized by relatively slow and minor terminus advances (tens to few hundreds of meters over several years, even for the largest glaciers), and usually lacks separation of the receiving area into a static, downwasting mass of dead ice.

The relatively widespread occurrence of glacier pulsations is consistent with the findings of [18] over the northern area of the Pamirs. Moreover, that study reported a peak in surging activity taking place between 2000 and 2008, which also matches our results over the Hissar–Alay (Figure 6). Other similarities include the observation of pulsations on glaciers smaller than 0.5 km² and a mean duration of the active phase slightly exceeding one decade. However, the surge-type glaciers described by [18] are in general much larger (mean area > 10 km²) and have more intense events compared with the pulsations observed in the Hissar–Alay.

The variable number of pulsating glaciers over the years, peaking around 1970 and 2004 (Figure 6), could express the influence of a climatic driver over the glacier dynamics. For instance, occurrence of enhanced melting—potentially triggering widespread frictional instabilities at the mountain range scale—was also discussed by [22,30,31]. However, a deeper analysis of the climatic history of the region would be needed to support this consideration and rule out the possible impact of sampling bias on this result [18].

A high prevalence of glacier pulsations has implications for mass balance studies in the region. Specifically, geodetic calibration or validation of regional-scale modeling studies is delicate and requires frequent updates to the used outlines, to accommodate evolving glacier geometries [57,61,122]. Moreover, [29] previously confirmed that the Abramov glacier (reference glacier in the Pamir–Alay for the WGMS) is pulsating, casting doubt on its regional representativity. Our study expands that finding with the observation of widespread pulsation across the region, further complicating the interpretation of mass balance trends at the single glacier and regional scale. While in situ measurements remain invaluable for the calibration and validation of mass balance studies, information on the status of the pulsation cycle should be researched and included for a meaningful interpretation of their results.

5.3. Study Uncertainties and Limitations

A key challenge in remote sensing studies of glacier pulsation is the correct identification of surging activity. The issue is particularly critical in our case, since the study region is characterized by glaciers at the lower end of the spectrum of instability, closer in appearance to stable flow or to mass balance-driven advances than most other surge-type glaciers. As such, individual misclassifications are possible; our results (including the accuracy evaluation of GS1980) could be affected by a certain degree of interpretation, only partially mitigated by the combined use of multiple classification criteria. This characteristic is shared by all studies compiling inventories of glacier instabilities, which routinely find discrepancies in the evaluation of the behavior of single glaciers (e.g., [10,18]). However, such divergences are usually minor and do not significantly alter the conclusions of statistical and distribution analyses.

Coverage of the study region within our dataset is incomplete at the start of the period (Figure 6, Appendix B). As such, it is possible that we did not detect some pulsations taking place between the late 1960s and the late 1980s. In fact, some pulsating glaciers may be missed altogether due to their long recurrence interval, potentially exceeding the full duration of our dataset (Figure 7b). Because annual imagery is not available for most glaciers (Section 3.2.6), the reported durations and intervals are also spread out as reported in Section 4.2. Moreover, the exact timing of a pulsation is subject to interpretation, whenever the active phase starts within the glacier body and then leads to terminus advance: this could lead to discrepancies with studies analyzing pulsations from flow velocity [18]. To this respect, we did not attempt measurement of ice surface velocity from the orthoimages, because the small extent of most glaciers and the multi-annual intervals in the dataset hamper the successful application of window-based correlators [29]. However, it should be noted that changes in ice velocity are indirectly reflected in all three indicators of flow instability that we considered (Section 3.2.4). Indeed, glacier acceleration can be qualitatively inferred both in a sharp change from terminus retreat to advance and (via the continuity equation) in the patterns of ice thickness change (Figure 3c). Changes in flow direction and in relative velocity of contiguous ice masses drive the development of many surge-like geomorphological features such as looped and deformed moraines (Figure 3d).

Additional uncertainties in our study arise from the processing of remote sensing data archives, especially in the case of declassified reconnaissance data affected by various types of artifacts (e.g., [106]). A systematic uncertainty analysis was performed by [29], where the accuracy and suitability of both orthoimages and DEMs were established against in situ measurements and independent remote sensing observations during two full pulsation cycles of Abramov glacier. Given the similar data sources and methods used here, we expect those conclusions to also hold at the mountain range scale.

One last source of uncertainty with glacier identification is given by the recurrent confusion in the indices of pulsating glaciers that are reported or compiled by previous studies. We discovered several mismatches between the Sh1980 and USSR inventories (Section 3.2.1), and also in the surge events compiled into RGI7.0 (Section 2, Figure 10). We found and corrected such mismatches via manual examination of the individual glaciers: we suggest that studies comparing their results with previous compilations should carefully check glacier identification in order to rule out potential errors of this kind.

5.4. Significance and Comparison with Previous Studies

Our study has found 171 pulsating glaciers in the Hissar–Alay. After correction of indexing mismatches, only 15 were already known from past studies (including those assumed from indirect evidence), with major discrepancies in the identification of individual glaciers (Figure 10). In particular, the HMA-wide remote sensing analyses of [34,35] found just two and five glacier surges over 2000–2018 and 1970–2020, respectively. Rigorous inter-comparisons of surge-type glacier inventories are challenging due to the inconsistent data, methods and criteria employed by different studies (e.g., [34]). Still, the large number of newly identified pulsations could be explained by several factors:

• Our selection of multi-source satellite data provides finer spatial resolution and longer temporal coverage than previous studies, enabling better detection of the low-intensity events that characterize the region.
• Larger-area studies usually set thresholds for automated detection of glacier surges, and those are probably not matched by the observed low-intensity pulsations.
• Manual identification of the highest-quality satellite scenes, later processed via a specifically tuned, glacier-wise pipeline (Section 3.2.2), can likely extract and preserve detail better than the standardized data aggregation and automated selection methods used within HMA-wide or global studies.

The observation of pulsations at several very small glaciers (<1 km²; Figure 5) supports the assumption that existing inventories of surge-type glaciers can be biased towards larger glaciers, especially due to preferential sampling of the latter (e.g., [18]). Such a bias could potentially persist into the published distribution analyses of morphological parameters (such as slope) for surge-type versus stable glaciers—analyses that are frequently used to pinpoint characteristics conducive to surging activity (e.g., [34,35]). This bias at the lower end of the spectrum of instability may not be sufficient to overturn the conclusions of such analyses. However, the relevance of a sharp classification (surge-type versus stable) could be called into question for such small, slightly unstable glaciers. As higher-resolution data keep revealing more numerous and less intense pulsations on smaller glaciers, it may become necessary to adapt such distribution analyses, for example, (1) restricting the sample to a given detection threshold for glacier area and magnitude of instability or (2) explicitly incorporating into the analysis a quantitative metric of flow instability, such as relative magnitude of speed-up and mass redistribution.

The methods we have implemented to compile the multi-source remote sensing dataset are mostly open-source and automated, enabling efficient up-scaling and potentially the detection of small glacier pulsations in other regions. In particular, cloud-free satellite mosaics and global DEMs are readily available to serve as reference for the orthorectification and stereo processing of raw scenes. The most time-consuming parts of our workflow are (1) the selection of good-quality initial raw scenes, often lacking usable cloud or snow masks, and (2) the manual inspection of orthoimages and DEM differences to assess the occurrence of pulsations. For (1), several methods are being developed to calculate quality masks even on monoscopic, panchromatic scenes (e.g., [124]). Full automation of our multi-criterion inspection is more challenging: existing studies have used algorithmic evaluation of ice thickness changes and ice surface velocities (e.g., [34,35,36,37]), but effective machine recognition of terminus advances and especially of surface geomorphology is still in the nascent phase [125,126].

6. Conclusions

This study presents a comprehensive assessment of glacier dynamics in the Hissar–Alay mountain range. The assessment enabled the verification of the predictions of glacier pulsation made by leading Soviet glaciologists Glazirin and Shchetinnikov in 1980. By compiling a high-resolution, multi-sensor satellite dataset spanning nearly six decades, we systematically re-examined the dynamics of 696 glaciers and found evidence of pulsation in 171 of them, revealing a prevalence of 25%—a significantly higher figure than previously reported for the region. Systematic comparison between historical compilations and modern inventories of glacier instabilities found major discrepancies and inaccuracies in the latter. Archived Soviet publications, based on historical knowledge and undisclosed aerial surveys, were found to still hold the most complete information on the topic.

Our analysis confirms that glacier pulsations in the Hissar–Alay are widespread but typically subdued in nature, with modest terminus advances and relatively long active phases (mean duration of 14 years). These instabilities affect glaciers of all sizes, including many smaller than 1 km², suggesting that prior inventories may have underrepresented the full extent of pulsation activity due to limitations in spatial resolution, temporal coverage, or detection criteria. An increased occurrence of pulsations in the early 2000s matches the results of previous studies in the neighboring Pamir region.

The GS1980 classification model, while less accurate than originally claimed, exhibits meaningful predictive ability in distinguishing pulsating from stable-flowing glaciers. In particular, it predicts an overall prevalence of glacier pulsation very close to the one observed in our study. However, the model reliance on geometric parameters—which are only indirectly related to local climate—introduces important limitations, as the predictor values are subject to uncertainty and interpretation. While the model yields some insight into the distribution of glacier instabilities, it cannot fully account for the occurrence of pulsations as these are influenced by factors beyond morphology.

Finally, the frequent detection of low-intensity pulsations confirms the challenges of a sharp binary classification between stable and pulsating behavior, particularly in small glaciers. Many of the observed events belong to the lower-intensity end of a continuous spectrum of glacier flow instabilities. This suggests the need for more flexible classification frameworks that can reflect the full range of dynamical behaviors, now observable thanks to increasingly detailed satellite records.