From Spectral Indices to Artificial Intelligence: A Review of Remote Sensing Methodologies for Glacier Mapping

Highlights

What are the main findings?

• The review emphasizes a significant transition in technology, moving from traditional spectral indices to more sophisticated Machine Learning (ML) and Deep Learning (DL) models. These methods now represent 20% and 56% of recent glacier mapping studies, respectively.
• Data fusion, especially combining optical, Synthetic Aperture Radar (SAR), and geomorphometric data, consistently achieves the highest glacier mapping accuracy compared to single sensor approaches.

What are the implications of the main findings?

• The transition to automated, data-driven systems allows for precise, large-scale monitoring of complex features such as debris-covered glaciers and calving fronts that were previously difficult to characterize.
• Future progress relies on overcoming the “ground truth” bottleneck by fostering community-led efforts to develop standardized, high-resolution benchmark datasets and enhance model transferability across various geographic regions.

Abstract

Glaciers are critical indicators of global climate change, and their accelerated retreat has profound implications for sea-level rise, water resources, and ecosystem stability. Accurate and timely mapping of glacier extent is essential for monitoring these changes. This review provides a comprehensive overview of the evolution of remote sensing techniques for glacier mapping, charting the progression from traditional spectral indices to the current state-of-the-art machine learning (ML) and deep learning (DL) models. We analyze the strengths and limitations of various methods, including the computational efficiency of indices like the Normalized Difference Snow Index (NDSI), the classificatory power of ML algorithms like Random Forest (RF), and the superior performance of DL architectures, particularly U-Net and its variants, for semantic segmentation of glacier mapping. Our analysis highlights a clear trend towards automated, data-driven approaches that have significantly enhanced the accuracy and scale of glacier delineation. However, progress is slowed by key challenges, most importantly the difficulty in getting accurate ‘ground truth’ data due to a lack of standardized, high-resolution training and validation datasets. Other key limitations include an over-reliance on a few model architectures and the need to bridge the gap between research-level accuracy and operational, real-time monitoring systems. Future progress in the field will depend on community-led efforts to create robust benchmark datasets, explore more diverse and efficient model architecture, develop sophisticated data fusion techniques, and improve model transferability and uncertainty quantification. By integrating cutting-edge AI with improved data practices, the remote sensing community can deliver the crucial data needed to understand and respond to the impacts of a changing climate.

1. Introduction

The impact of glaciers on the global environment is profound, particularly in the context of climate change and its widespread effects on global surface temperatures and sea levels [1,2]. As sensitive indicators of a warming climate, glaciers are retreating at an accelerated rate, a trend with significant global consequences [3,4]. Historically, glaciers outside of the major ice sheets contributed approximately 15–20% of the observed sea-level rise during the latter half of the 20th century [5]. This trend has intensified; since the year 2000, the world’s glaciers have lost an average of 273 billion metric tons of ice annually [6]. This global decline has critical implications for both planetary and human systems, as glacial meltwater directly influences weather patterns, the hydrological cycle, and sea level rise [7]. In regions such as the Andes and the Himalayas, the reduction in meltwater threatens freshwater resources vital for sustaining agricultural productivity and food security, potentially leading to water scarcity, resource competition, and significant socio-economic challenges. Furthermore, the alarming rate of glacial retreat disrupts local ecosystems and habitats, causing instability in hydrological regimes that can result in both floods and water shortages [8,9].

Projections from the Intergovernmental Panel on Climate Change (IPCC) indicate that this cryosphere decline will continue throughout the 21st century. Relative to a 1986–2005 baseline, projections for glacier mass loss from 2015 to 2100 indicate a likely reduction of 22–44% under a low-emissions scenario (RCP2.6) and a more substantial loss of 37–57% under a high-emissions scenario (RCP8.5). The consequences are particularly dire in regions with smaller glaciers, such as the European Alps, the Pyrenees, and the tropical Andes, where an excess of 80% of current glacier mass is projected to be lost by 2100 under RCP8.5. This underscores the critical role glaciers play in maintaining global climate stability and regional ecological integrity [10].

To meet the challenges of monitoring these vast, remote, and often inaccessible glacial environments, remote sensing has become an indispensable tool. A variety of satellite platforms equipped with different sensors are employed to acquire data across the electromagnetic spectrum. Optical sensors, such as those on the Landsat and Sentinel-2 missions, capture the spectral reflectance of the Earth’s surface, allowing for the differentiation of snow and ice from surrounding features like rock, vegetation, and water based on their unique signatures. Complementing this are Synthetic Aperture Radar (SAR) sensors, which possess the critical advantage of being able to penetrate cloud cover and operate regardless of daylight, ensuring consistent data acquisition in frequently overcast mountain regions [11,12,13]. Furthermore, these remote sensing technologies are fundamental for generating advanced geospatial products that provide a deeper understanding of the terrain [14]. For instance, Digital Elevation Models (DEMs) are created to provide a 3D view of the surface [15], which in turn serves as the basis for geomorphometry the quantitative analysis of landforms [16,17]. From DEMs, critical parameters like slope and aspect are calculated, which directly influence factors such as snow accumulation, solar radiation, and avalanche risk. Additionally, thermal sensors enable the derivation of Land Surface Temperature (LST), a key variable for monitoring the surface energy balance and modeling snow and glacier melt processes. Together, these derived products transform raw satellite data into actionable information for analyzing complex environmental systems. The data acquired from these sensors form the basis for numerous analytical techniques designed to extract and monitor glacier characteristics. The most fundamental of these are spectral indices, which use mathematical ratios between different spectral bands to enhance the contrast between ice and non-ice surfaces. More advanced approaches leverage classical Machine Learning (ML) algorithms to perform sophisticated image classification, improving accuracy in complex situations involving shadows or debris-covered ice. Most recently, the field has been advanced by deep learning (DL) models, which can automatically learn complex spatial and spectral features from the imagery, offering a pathway toward highly accurate and fully automated large-scale glacier monitoring. According to a survey conducted in Scopus with key words of “Ice Sheet”, “Glacier”, and “Remote sensing”, the number of publications utilizing remote sensing techniques in the field of ice sheets and glaciers has been increasing over the last three decades and exhibited a significant sharp increase since 2014, as shown in Figure 1.

Prior to the widespread availability of satellite remote sensing, glacier mapping relied primarily on systematic field surveys, aerial photography, and photogrammetric techniques. Early pioneering work by [18] provided one of the first comprehensive glacier inventories in the Canadian Arctic, demonstrating the importance of standardized mapping for understanding regional glacier distribution and mass balance. Building on these foundations, major glacier inventory initiatives were launched in the late 1960s and 1970s under the leadership of Fritz Müller and Gunnar Østrem, particularly in Scandinavia, Rhone Glacier, Switzerland, led by Axel Heiberg and North Water projects, and other mountain regions. These projects established consistent methodologies for glacier identification, classification, and cartographic representation using aerial photographs and topographic maps [19,20,21,22].

The Scandinavian glacier inventory programs led by Østrem further advanced quantitative glacier mapping by integrating long-term mass-balance measurements with spatial inventories, laying the groundwork for linking glacier extent changes to climatic variability [23]. Collectively, these early efforts laid the conceptual and methodological foundation for modern glacier inventories such as the World Glacier Inventory and the Randolph Glacier Inventory (RGI), and they strongly influenced the transition toward satellite-based glacier mapping frameworks.

Building on these early inventory efforts, and as global glacier retreat accelerates, the need for automated, repeatable, and scalable glacier mapping methods has become increasingly critical. Early glacier inventories provided static snapshots of glacier extent, but ongoing climate change demands frequent updates across vast, often inaccessible regions. In this context, global glacier databases such as the Global Land Ice Measurements from Space (GLIMS) initiative and the RGI have become indispensable reference frameworks. GLIMS was designed to systematically collect and standardize glacier outlines from satellite imagery while preserving regional expertise and metadata [24,25]. The RGI subsequently consolidated multiple regional inventories into a globally complete, temporally consistent dataset suitable for climate modeling and large-scale change detection [26]. These inventories provide not only glacier outlines but also essential historical and contextual information, including glacier aspect, elevation range, and regional setting, which remain crucial for interpreting observed changes and for training and validating automated mapping algorithms.

Early review papers from the mid-2000s to the early 2010s focused on establishing and consolidating foundational remote sensing techniques. For instance, comprehensive reviews on glacier mapping summarized the widespread adoption of optical satellite data from missions like Landsat, Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Moderate Resolution Imaging Spectroradiometer (MODIS) [27,28,29,30,31,32,33,34]. The primary technique discussed was the use of spectral indices, particularly the Normalized Difference Snow Index (NDSI), which leverages the strong contrast between the visible and short-wave infrared reflectance of snow and ice. The main challenges identified in these reviews were persistent and fundamental: frequent cloud cover obscuring visibility, topographic shadows in high-relief terrain causing data loss, and the spectral similarity between debris-covered glacier tongues and surrounding rock, which rendered simple index-based methods ineffective. To address issues of cloud and shadow, these reviews pointed to solutions like multi-temporal image compositing and the use of digital elevation models (DEMs) for topographic correction.

By the mid-2010s, a new generation of review articles began to document a shift towards multi-sensor data fusion and the application of classical ML. Reviews in this era discussed the cooperative use of optical data with SAR and thermal imagery to tackle the most difficult classification problems [2,35,36,37,38,39]. SAR was noted for its all-weather, day-night imaging capabilities, while thermal data were shown to help distinguish debris-covered ice (which remains cooler) from adjacent sun-warmed moraine. The primary challenge shifted from simple data availability to the complexity of integrating these disparate datasets. The solution, as outlined in reviews of the period, was the application of supervised classification algorithms such as Maximum Likelihood, Support Vector Machines (SVM), and RF. These methods could ingest multiple data layers (e.g., spectral bands, SAR backscatter, texture, and topographic parameters) to perform more robust and accurate classifications of cryosphere features than any single-sensor method could achieve alone.

The accurate identification and delineation of glaciers from remote sensing imagery present a significant and multifaceted challenge due to the spectral and spatial complexity of high-mountain environments. The primary difficulty lies in distinguishing glacier ice from other spectrally similar or visually ambiguous surface features. Clean glacier ice can be confused with seasonal snowpack, especially during accumulation periods, while differentiating it from large ice sheets or perennial snowfields requires contextual and topographic analysis. The problem is further exacerbated by the presence of supraglacial and proglacial water bodies, which have distinct spectral signatures but complicate the precise delineation of a glacier’s boundary.

Three of the most persistent and difficult challenges are the presence of debris cover, snow cover, and topographic shadow. Debris-covered glacier tongues, which are increasingly common as glaciers recede, are often spectrally indistinguishable from the surrounding rubble and exposed rock, making them a primary source of underestimation in glacier inventories. Conversely, fresh snow cover can obscure the true extent of a glacier, mask its margins and make it difficult to differentiate from adjacent snow-covered terrain. Finally, the extreme topographic relief of glacial regions creates extensive areas of shadow, which corrupts the spectral information recorded by optical sensors and can cause shadowed ice or snow to be misclassified as water or bare ground. Separating glaciers from these features, as well as from other standard land cover classes like vegetation and soil [40,41,42,43], requires more than simple spectral analysis.

The importance of this review lies in the fact that overcoming these challenges has driven innovation in remote sensing over the past decades. Early methods’ limitations in handling debris, shadow, and snow led to the development of more advanced techniques. This technological progress, from basic spectral indices that struggled with these issues, to traditional machine learning algorithms that combined multiple data sources for improved classification, and finally to modern deep learning models that can automatically infer contextual relationships, demonstrates clear scientific advancement. This review uniquely adds value to the field by offering a comprehensive, future-oriented synthesis that supports strategic research planning through several key pillars.

Comparative Analysis of Method Classes: Unlike previous surveys, this study compares spectral indices, machine learning algorithms, and deep learning models, evaluating their performance, efficiency, and suitability for different glacial environments.

Documentation of Data Fusion Trends: We examine the shift toward multi-source integration, highlighting how combining optical and SAR data, incorporating Digital Elevation Models (DEMs), and using LiDAR-assisted mapping are essential for overcoming sensing limitations.

Identification of Research Gaps and Future Directions: By assessing the current state of the art, we identify key bottlenecks such as the lack of standardized “ground truth” benchmark datasets and the need for improved model transferability, providing a clear roadmap for future research and monitoring systems.

This review synthesizes advances in remote sensing approaches for glacier extraction, with a particular emphasis on three methodological domains: spectral indices, ML, and DL methods. The paper highlights the strengths, limitations, and comparative effectiveness of these approaches, identifies knowledge gaps, and discusses emerging trends that could shape future research in glacier monitoring.

2. Methodology

This study employs an overview methodology to comprehensively identify, evaluate, and synthesize existing research on the application of glacier mapping from remote sensing data. The research framework, illustrated in Figure 2, is structured into three primary phases: (1) Identification of studies via databases, (2) Screening of peer reviewed article based on keywords, and (3) Analysis and Manuscript Preparation. This structured approach ensures a transparent, replicable, and unbiased review of the current state-of-the-art.

The initial phase involved defining the research scope and conducting comprehensive research of relevant academic literature. The research was performed across major scientific databases, including Scopus, Web of Science, and Google Scholar, to ensure broad coverage of the field. The research strategy was built around a combination of keywords using Boolean operators (AND/OR) to refine the results. The primary keywords used were: “Glacier”, “Ice Sheet”, “Glacier Mapping”, and “Remote Sensing”,”Glacier Spectral Indices”, “Glacial Lake”, “Debris Covered Glacier”, “Snow Covered Glacier”, “Supraglacial glacier”, “Machine Learning”, and “Deep Learning”.

The retrieved literature underwent a rigorous two-stage screening process. Initially, titles and abstracts were evaluated to identify potentially relevant studies. In the subsequent stage, the full texts of the shortlisted articles were assessed against specific criteria. To the best of our knowledge, this review encompasses the most significant contributions to the field since the year 2000, with a primary focus on: (a) remote sensing methodologies for mapping diverse glacier types, including clean-ice, debris-covered, and snow-covered glaciers, as well as glacier calving; (b) peer-reviewed articles and conference proceedings published in English; and (c) high-impact studies published within the last two decades to ensure the inclusion of modern technological advancements.

2.1. Glacier Surfaces

2.1.1. Glaciological Classification of Ice-Related Landforms

Glacierized terrains often include a continuum of ice-related landforms that differ in structure, dynamics, and degree of activity. Clean and debris-covered glaciers represent active glacier systems with continuous ice cores and dynamic flow. In contrast, buried or residual ice occurs within debris-dominated terrains where ice is discontinuous and often partially insulated by surface material. Dead ice refers to stagnant remnants of former glaciers that no longer exhibit flow but may persist beneath debris cover. Rock glaciers, while often containing significant ice, are characterized by creep-driven deformation and may represent transitional forms between glacial and periglacial processes. Although these landforms are genetically related in many mountain environments, they differ in ice continuity, dynamics, and geomorphological expression, making their distinction critical for both glaciological interpretation and remote sensing analysis. Figure 3 clarifies the differences between different types of surface glaciers and their surrounding features from terrestrial images.

2.1.2. Remote Sensing Characteristics of Glacier Surfaces

The ability to detect glaciers from satellite observations is directly influenced by their diverse physical and spectral characteristics. To address this variability, the field has generally concentrated on identifying four primary glacier surface types. Accordingly, this review categorizes the literature into four features: clean glacier, debris-covered glacier, snow-covered glacier, and supraglacial lake.

Clean or open glaciers consist of bare ice with minimal debris cover, showing high reflectance in the visible/near-infrared (VNIR) bands (e.g., Landsat/Sentinel Band 2–4) and strong absorption in the shortwave infrared (SWIR; Bands 5–7), making them readily distinguishable from surrounding terrain at higher elevations where temperature remains low. Debris-covered glaciers, however, are spectrally similar to surrounding moraines in VNIR and Short-Wave Infrared (SWIR) bands, reducing their separability in optical imagery. Their detection is more effective when incorporating thermal infrared (TIR; Landsat Band 10–11) to exploit surface temperature contrasts, and synthetic aperture radar (SAR; Sentinel-1, ALOS PALSAR) to capture surface roughness and structure, especially at mid-elevations (3000–5000 m) where melt rates are high. Snow-covered glaciers exhibit very high reflectance in the visible wavelengths (particularly the blue and green bands), with strong absorption in the SWIR, making indices such as NDSI effective for mapping. Yet seasonal snow cover at lower elevations can obscure glacier boundaries, requiring multi-temporal optical and DEM-based analyses. Mixed-surface glaciers, which include supraglacial lakes, vegetation patches, and debris layers, present complex spectral signatures. Meltwater bodies show strong absorption in near-infrared (NIR) bands and low reflectance in SWIR, while vegetation exhibits characteristic peaks in NIR reflectance; thus, a combination of VNIR, SWIR, and thermal bands is necessary. These complex glacier types often occur at lower (<3000 m) to mid (3000–5000 m) elevations, where higher temperatures accelerate melting. In summary, clean ice and snow-covered glaciers are best differentiated in optical VNIR/SWIR bands at high elevations (>5000 m), while debris-covered and mixed-surface glaciers require integrated approaches using optical, thermal, SAR, and elevation data to account for heterogeneous surface properties and temperature-driven dynamics.

2.2. Satellite Data for Glacier Mapping

The automated delineation of glaciers is fundamentally dependent on data acquired by satellite remote sensing. Over recent decades, a suite of spaceborne sensors operating across the electromagnetic spectrum has been pivotal in advancing cryosphere science. The literature analyzed in this review predominantly utilizes imagery from a core set of satellite missions, frequently augmented with geomorphometric data to refine glacier extraction. To structure our analysis of the diverse datasets employed glacier mapping, we have categorized them into five distinct groups: optical imagery, SAR imagery, a hybrid of SAR and optical, a combination of geomorphometric and optical, and a complete fusion of geomorphometric, optical, and SAR data.

Table 1. Different types of satellite imagery utilized in glacier mapping.

Satellite/Mission	Sensor	Type	Spatial Resolution	# Bands	Operational Period	Status
IKONOS	-	Optical	1 m (Panchromatic), 4 m (Multispectral)	5	1999–2015	Discontinued
SPOT	Various	Optical	1.5 m–20 m	4–5	1986–Present	Partially Active
RASAT	-	Optical	7.5 m (Panchromatic), 15 m	4	2011–Present	Active
Sentinel-2	MSI	Optical	10 m, 20 m, 60 m	13	2015–Present	Active
ASTER	-	Optical	15 m (VNIR), 30 m (SWIR), 90 m (TIR)	14	1999–Present	Active
Landsat-9	OLI-2/TIRS-2	Optical	30 m (MS), 15 m (Pan)	11	2021–Present	Active
Landsat-8	OLI/TIRS	Optical	30 m (MS), 15 m (Pan)	11	2013–Present	Active
Landsat-7	ETM+	Optical	30 m (MS), 15 m (Pan)	8	1999–Present	Active (degraded)
Landsat-5	TM	Optical	30 m	7	1984–2013	Discontinued
MODIS	-	Optical	250 m, 500 m, 1k m	36	1999–Present	Active
Sentinel-1	C-SAR	Radar	10 m–40 m	1 (C-Band)	2014–Present	Active
TerraSAR-X	-	Radar	1 m–16 m	1 (X-Band)	2007–Present	Active
Envisat	ASAR	Radar	30 m–1k m	1 (C-Band)	2002–2012	Discontinued
ALOS	PALSAR	Radar	10 m–100 m	1 (L-Band)	2006–2011	Discontinued
ERS-1/ERS-2	AMI	Radar	~30 m	1 (C-Band)	1991–2011	Discontinued
SRTM	SIR-C/X-SAR	Radar	~30 m	2 (C & X-Band)	2000	Completed

shows the common types of satellite imagery utilized in glacier mapping.

2.2.1. Optical Satellite Imagery

Optical sensors, which capture solar radiation reflected from the Earth’s surface, are the most widely used data source for glacier mapping. The long-running Landsat program, a joint initiative of NASA and the U.S. Geological Survey, stands out as the most critical resource, providing an unparalleled historical archive for monitoring long-term glacier changes. The reviewed studies frequently employ data from the Thematic Mapper (TM) on Landsat-5, the Enhanced Thematic Mapper Plus (ETM+) on Landsat-7, and the Operational Land Imager (OLI) on Landsat-8. The multispectral bands provided by these sensors, particularly in the visible, NIR, and short-wave infrared (SWIR) ranges, are essential for differentiating ice and snow from surrounding features like rock, soil, and vegetation. Complementing the Landsat archive is the European Space Agency’s (ESA) Sentinel-2 mission, which offers similar multispectral data at a higher spatial resolution and with a more frequent revisit time, making it invaluable for contemporary glacier monitoring. Other high-resolution optical sensors such as SPOT (Satellite pour l’Observation de la Terre) have also been utilized in specific regional studies.

2.2.2. Synthetic Aperture Radar (SAR) Imagery

To overcome the limitations of optical imagery, such as cloud cover and the lack of illumination during polar winters, many studies incorporate Synthetic Aperture Radar (SAR) data. SAR is an active remote sensing system that transmits microwave pulses and measures the backscattered signal, allowing it to acquire images regardless of weather conditions or time of day. The reviewed literature shows extensive use of data from ESA’s Sentinel-1 mission. Other SAR sensors cited include TerraSAR-X, ERS-1/2, Envisat, and ALOS PALSAR. SAR data are particularly effective for delineating calving fronts, identifying grounding lines, and distinguishing between different ice facies due to its sensitivity to surface roughness, structure, and the presence of liquid water.

2.2.3. Geomorphometric Data

Geomorphometric data provide the essential topographic context needed to distinguish different glacier surfaces by quantifying the shape and form of the ice. A Digital Elevation Model (DEM) serves as the foundational dataset, providing the raw height information from which all other metrics are derived. From the DEM, slope and aspect are calculated to reveal the steepness and orientation of the glacier’s surface, which are critical for identifying flat accumulation zones versus steep icefalls and for modelling solar radiation that drives melting. Furthermore, other metrics like curvature (both profile and planform) and roughness help to highlight detailed textural features. By combining these quantitative measures of the glacier’s shape, researchers can effectively classify and map distinct zones such as the accumulation area, ablation zone, crevasse fields, and icefalls based on their unique topographic signatures.

3. Glacier Mapping Techniques

Glacier mapping is a critical component of cryosphere research, providing essential data for understanding climate change, sea-level rise, regional hydrology, particularly in understanding glacier dynamics, mass balance, and the implications of climate change. Over the past few decades, remote sensing has become the primary tool for this purpose, leading to the development of a wide range of methodologies. These techniques can be broadly categorized into spectral indices, spectral unmixing, and increasingly, ML and DL approaches. This study synthesizes contemporary approaches in glacier mapping, highlighting their advantages and the contexts in which they are applied.

3.1. Spectral Indices in Glacier Mapping

Spectral indices are the most common and computationally efficient methods for mapping glaciers. These techniques leverage the distinct spectral properties of ice and snow compared to other land cover types. The core principle involves applying mathematical formulas, typically band ratios (BRs), to multispectral imagery. Figure 4 shows the steps in how the spectral indices have been utilized for glacier mapping.

A classic example is the NDSI, which exploits the high reflectance of snow and ice in the visible green band and strong absorption in the shortwave infrared (SWIR) band [44,45]. Researchers have subsequently developed numerous specialized indices to address challenges. Several studies developed new spectral indices and reported higher accuracy compared to NDSI index [46,47,48,49,50,51,52,53]. For instance, Ref. [52] developed the Adjusted Normalized Difference Snow Index (ANDSI), to overcome the shortness of NDSI and differentiate glacier from water accurately. The traditional NDSI struggles to differentiate between glaciers and water, as they appear similar in the green and shortwave-infrared bands. The ANDSI solves this problem by including additional red and NIR bands, making it a more reliable index for mapping glaciers in regions with many lakes. The results demonstrated the clear advantage of ANDSI, which achieved an average overall accuracy of approximately 95% and a Kappa coefficient of 0.92, compared to NDSI’s 91% accuracy and 0.85 Kappa. Similarly, Ref. [47] conducted a study across four test areas in China, from Tibet to Xinjiang, and proposed a new index called the Automated Glacier Extraction Index (AGEI) to enhance the accuracy of mapping debris-free glaciers. Using Landsat and Sentinel-2 imagery, the research aimed to improve glacier delineation, especially in regions with shadows and water bodies. The AGEI is a weighted ratio index, calculated as $AGEI = (\alpha \times Red + (1-\alpha ) \times NIR)/SWIR$, where α is set to 0.5. a is an empirical parameter that needs to be calibrated for different study sites. This index was developed to overcome the limitations of existing methods like the NDSI, and other simple BRs such as Red/SWIR and NIR/SWIR. The methodology involved selecting pure glacier and non-glacier pixels to develop the index, followed by glacier mapping using a threshold of 2 ± 0.5. The effectiveness of the AGEI was validated through various methods, including a confusion matrix, sub-pixel analysis, and comparison with Google Earth imagery and manual delineations. The results demonstrated that AGEI consistently outperformed other indices, achieving a high overall accuracy ranging from 86% to 90% across the test regions, with a strong correlation (r² = 0.878) to reference data.

Table 2. Summary of studies highlighting the spectral indices utilized for glacier mapping, 1995–2025.

Study	Study Area	Purpose of Study	Spectral Index	Threshold Rules	Satellite Data	Validation	Performance Metrics
[53]	Koxkar, Yengisogat, Karakoram, China	developed new BR using NIR, SWIR and TIR to extract debris covered glacier	BR = TIR/(NIR + SWIR_01)	Density slicing and slope < 12°	▪ Landsat-5,7 ▪ SRTM	RGI_v04	The method showed 0.34–2% difference compared to manual delineation.
[46]	Qinghai Lake, China	To develop a new index, differentiate between snow, ice and ice-snow	$NDISI=(Band\_01-Band\_02)/(Ban{d}_{01}+Band\_02)$, band 1 (range: 1080–1120 nm) and band 2 (range: 1760–1800 nm)	NDISI for snow, 0.6 to 0.7, NDISI for ice-snow > 0.9 and NDISI for ice > 0.4	▪ BRDF observations	Field data using multi angle tool and PSR-3500 field spectroradiometer	NDISI was validated with in situ spectral data from ALT, BBR, and CGL, showing differences below 0.08.
[54]	Qinghai-Xizang, China	To propose a new index for mapping of snow/ice cover using VGT sensor data on SPOT	$NDSII\_TM=(TM3-TM5)/(TM3+TM5)$ $NDSII\_VGT=(B2-MIR)/(B2+MIR)$ $NDSI=(TM2-TM5)/(TM2+TM5)$	NDSII_TM > 0.4 NDSII_VGT > 0.4	▪ Landsat-5, ▪ Vegetation sensor on SPOT_4	Compared glaciered area of NDSI with NDSII_TM	The difference recorded was 0.1%
[55]	The central Southern Alps, New Zealand	To measure glacier area changes between 1978 and 2002	$NDSI=(AST{E}_{01}-ASTER\_04)/(ASTE\_01+ASTER\_04)$ $BR=(ASTER\_03)/(ASTER\_04)$	NDSI > 0.5 ASTER Ratio > 2	▪ ASTER ▪ aerial photographs from 1978	GPS Boundary	Overall accuracy for ASTER Ratio 97%, NDSI = 96.74% and supervised classification = 91.64%
[56]	Pindary, India	map the change in the frontal length or snout region of a glacier	$NDSI=(Green-SWIR)/(Green+SWIR)$ $NDSI=(Green-NIR)/(Green+NIR)$	NDSI > 0.85 for snow, NDSI −0.15 for rocky body	▪ Landsat ▪ Sentinel-2
[57]	The Chenab basin in the Himalayas	discrimination and mapping of supraglacial cover types using ASTER data	$NDGI=(Green\_B1-Red\_B2)/(Green\_B1+Red\_B2)$ $NDSII=(Green\_B1-NIR\_B3)/(Green\_B1+NIR\_B3)$	NDSI > 0.61 NDSII > 0.025	▪ ASTER	ASTER VNIR image was used as a reference for validation	Overall accuracy of 91%
[47]	Tibet and Xinjiang in China	To develop new Index for debris-free glaciers in shadowed and water-affected areas	$AGEI={\displaystyle \frac{\alpha \cdot D{N}_{Red}+\left(1-\alpha \right)\cdot D{N}_{NIR}}{D{N}_{SWIR}}}$ A = 0.5 $NDSI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{SWIR}}{{\rho }_{Green}+{\rho }_{SWIR}}}$ $Red/SWIR={\displaystyle \frac{D{N}_{Red}}{D{N}_{SWIR}}}$ $NIR/SWIR={\displaystyle \frac{D{N}_{NIR}}{D{N}_{SWIR}}}$	AGEI > 2 ± 0.5	▪ Landsat TM, ▪ Landsat-8 ▪ Sentinel-2	Inventory data, Google Earth image and a cross validation between Landsat and Sentinel	Across three regions’ AGEI with OA of 88.68% and Kappa of 0.75 AGEI was the best
[58]	Shisper Glacier, Indus Basin, Pakistan	To carry out a detailed analysis of the Shisper Lake breach of 7 May 2022	$NDDI=(MODIS\_B7-MODIS\_B3)/(MODIS\_B7-MODIS\_B3)$ $NDWI=(MODIS\_B3-MODIS\_B5)/(MODIS\_B3-MODIS\_B5)$ Debris Cover $=(Sentinel2\_B4)/(Sentinel2\_B11)$	NDSI > 0.4 for snow; NDDI > 0.28 for snow features, <0.28 for dust; Debris Cover Index ≥ 1.8 for clean vs. debris-covered glaciers.	▪ Landsat-8,9, ▪ Sentinel-2, ▪ MODIS, ▪ Sentinel-1
[59]	Cilo mountain, Turkey	To compare the effectiveness of different indices for Snow Cover Glacier mapping	$NDSI=(\mathrm{T}\mathrm{M}2-\mathrm{T}\mathrm{M}5)/(\mathrm{T}\mathrm{M}2+\mathrm{T}\mathrm{M}5)$ $NDSII=(\mathrm{T}\mathrm{M}3-\mathrm{T}\mathrm{M}5)/(\mathrm{T}\mathrm{M}3+\mathrm{T}\mathrm{M}5)$ $\mathrm{N}\mathrm{D}\mathrm{P}\mathrm{C}\mathrm{S}\mathrm{I}=(1\mathrm{s}\mathrm{t}\mathrm{P}\mathrm{C}-2\mathrm{n}\mathrm{d}\mathrm{P}\mathrm{C})/(1\mathrm{s}\mathrm{t}\mathrm{P}\mathrm{C}+2\mathrm{n}\mathrm{d}\mathrm{P}\mathrm{C})$ WET_Ladsat8 = 0.1511 × OLI2 + 0.1973 × OLI3 + 0.3283 × OLI4 + 0.3407 × OLI5 − 0.7117 × OLI6 − 0.4559 × OLI7	WET = 0.15 NDSII = 0.2 NDSI = 0.2 NDPCSI is not reliable for snow detection	▪ Landsat-5,8, ▪ RASAT	RASAT image and ground truth	Kappa of 0.74, 0.76, 0.4, 0.77 for NDSI, NDSII, NDPCSI and WET, respectively. WET is the most accurate threshold
[48]	The Beas River basin, India	To develop a new spectral index to differentiate snow/ice from others	$SWI=Green(NIR-SWIR)/(Green+NIR)(NIR+SWIR)$	SWI > 0.21	▪ Landsat-8, ▪ Sentinel-2	Ground truth data from field surveys using a spectroradiometer and DGPS were used for validation.	SWI achieved an overall accuracy of 93% and Kappa of 0.95
[60]	Gangotri Glacier, Himalayas India	To develop new index for differentiation of snow/ice from water bodies	$NDSTI=VIS\left(Blue\right)-{\displaystyle \frac{TIR}{VIS\left(Blue\right)}}+TIR$		▪ ASTER, ▪ Landsat TM	ASTER Image of 2001 VNIR and thermal bands was used as reference	Overall accuracy of 93%
[61]	Aletschgletscher region in the Swiss Alps and the Jostedalsbreen Ice Cap in Norway	Extract glacier extents and surface facies from Sentinel-2 and compare with Landsat-8.	$NDSI=(Green-SWIR)/(Green+SWIR)$ $BR=\left(Red\right)/\left(SWIR11\right)$ shaded ice $BR=\left(Red\right)/\left(SWIR12\right)$ $BR=(NIR\_MSI8)/\left(MSI11\right)$	NDSI ≥ 0.20 0 ≤ Red/SWIR ≤ 2, Red/SWIR ≤ 1.2; 0 ≤ NIR/SWIR ≤ 1	▪ Sentinel-2 ▪ Landsat-8	Cross comparison between OLI pan and OLI red, MSI and OLI red, and MSI and OLI pan	MSI4/MSI11 accurately extracted glaciers compared to others
[49]	Russia, Canada, Mongolia, China, Greenland	Methodology for extracting icy lakes and classifying water-ice using Landsat-8 OLI.	$MNDWI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{SWIR1}}{{\rho }_{Green}+{\rho }_{SWIR1}}}$ $WRI={\displaystyle \frac{{\rho }_{Green}+{\rho }_{Red}}{{\rho }_{NIR}+{\rho }_{SWIR1}}}$ ${NDWI}_H={\displaystyle \frac{{\rho }_{NIR}-{\rho }_{Blue}}{{\rho }_{NIR}+{\rho }_{Blue}}}$ $WRI={\displaystyle \frac{{\rho }_{Red}+{\rho }_{SWIR2}}{{\rho }_{NIR}+{\rho }_{SWIR1}}}$ $WICI={\displaystyle \frac{{{\sigma }_{\rho }}_{SWIR1}+{{\sigma }_{\rho }}_{SWIR1}}{{{2\sigma }_{\rho }}_{NIR}}}$	MNDWI > 0.75 WRI > 1.5 NDWIH < −0.1 ILI > 0.95	▪ Landsat-8	Manual Digitization and Sentinel-1	ILI with Kappa of 0.90 outperformed other three indices, moreover, they suggested WICI for separating shallow water from ice
[62]	The Tibetan Plateau, China	To compare the result of NDSI and RDRI for river ice monitoring	$RDRI={\displaystyle \frac{{\rho }_{Red}-{\rho }_{NIR}}{{\rho }_{NIR}+{\rho }_{SWIR}}}$ $NDSI={\displaystyle \frac{{\rho }_{VIS}-{\rho }_{SWIR}}{{\rho }_{VIS}+{\rho }_{SWIR}}}$	RDRI ≥ 0.13	▪ Landsat-8, ▪ Sentinel-2	Ground measurement	RDRI achieved 99.4% OA and 0.88 Kappa on the Tibetan Plateau; both RDRI and NDSI showed similar accuracy for river ice without snow.
[50]	The Tibetan Plateau, China	To develop indices for the accurate extraction of water and snow cover from glaciers	${NDWI}_{new}={\displaystyle \frac{{\rho }_{Green}-a\ast \left({\rho }_{NIR}\right)}{{(\rho }_{Green}+{\rho }_{NIR})}}$ a = 2 ${NDSI}_{new}={\displaystyle \frac{{\rho }_{NIR}-{\rho }_{SWIR1}-b}{{(\rho }_{NIR}+{\rho }_{SWIR1})}}$ b = 0.05	NDSI > 0.5 for clean glacier	▪ Landsat-5,7 and 8	Google Earth Image high-resolution image, Global Inland Water Dataset and the global snow cover product.	${NDSI}_{new}$ achieved 97% OA and 0.93 Kappa; ${NDWI}_{new}$ reached 95% OA and 0.90 Kappa, outperforming traditional indices.
[51]	The Dhauliganga basin, India	Compare techniques for extracting open glaciers to evaluate sensor effectiveness.	$BR=\left(Red\right)/\left(SWIR\right)$ $BR=\left(NIR\right)/\left(SWIR\right)$ $MNDWI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{SWIR1}}{{\rho }_{Green}+{\rho }_{SWIR1}}}$	Red/SWIR > 2.75 NIR/SWIR > 2.75 NDSI > 0.40 Slope for clean ice 20–40°	▪ Landsat-8, ▪ Sentinel-2	Sentinel-1 and Google Earth Images	NIR/SWIR outperformed other indices and Landsat-8 for open glacier extraction; per-pixel RMSE showed 1.12% uncertainty for Landsat and 0.3% for MSI.
[63]	Nevados Caullaraju-Pastoruri, Peru	To explore the temporal variation of clean ice and total glacier area in Nevados	$NDSI=(\mathrm{T}\mathrm{M}2-\mathrm{T}\mathrm{M}5)/(\mathrm{T}\mathrm{M}2+\mathrm{T}\mathrm{M}5)$	NDSI > 0.5 for clean ice, NDSI < 0.5, slope < 24°, LST < ${LST}_{average}-{LST}_{STD}$, NDSI < 0.1 (not vegetation) for debris covered glacier	▪ Landsat-5, ▪ 15m DEM, ▪ Metrological data	Results compared to ground-truth glacier area data and RGIv7	Average Area Error around 5%
[52]	Canada, China, Sweden, Switzerland-Italy	To propose and evaluate a new index (ANDSI) for distinguishing glacier from water	$NDSI={\displaystyle \frac{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1}{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}$ $CSI={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2}}$ $ANDSI={\displaystyle \frac{\mathrm{C}\mathrm{S}\mathrm{I}-\mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I}}{\mathrm{C}\mathrm{S}\mathrm{I}+\mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I}}}$ $ANDSI={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1)-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1)}{\mathrm{N}\mathrm{I}\mathrm{R}\left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1)}}$	NDSI ≥ 0.42 NDSI ≥ 0.42 and (−0.25 ≤ Ln (ANDSI) < 0) effectively distinguish glaciers from non-glacier areas and water.	▪ Sentinel-2 (Level 1C)	Ground reference samples from Sentinel-2 (visual interpretation + expert knowledge).	ANDSI outperformed NDSI, achieving ~95% accuracy and 0.92 Kappa versus NDSI’s ~91% accuracy and 0.85 Kappa.
[64]	Cordillera Blanca, Peru	To quantify glacier area and change and evaluate threshold sensitivity for glacier mapping.	$NDSI=(\mathrm{T}\mathrm{M}2-\mathrm{T}\mathrm{M}5)/(\mathrm{T}\mathrm{M}2+\mathrm{T}\mathrm{M}5)$	NDSI threshold calibrated at 0.42 (SD 0.13, median 0.45) using high-resolution IKONOS and QuickBird imagery.	▪ Landsat-2,5,7 ▪ ASTER, ▪ IKONOS-2, ▪ QuickBird, ▪ Google Earth	Validated using high-resolution IKONOS-2 and QuickBird imagery, with manual cross-checks on Google Earth.	Error analyzed using ±30 m buffer and ±20% uncertainty for debris-covered glaciers.

summarizes the reviewed studies, highlighting index techniques for glacier mapping.

3.2. Machine Learning Models in Glacier Mapping

Moving beyond the fixed-threshold limits of spectral indices, machine learning (ML) provides a fundamentally different approach to glacier classification by mapping remote sensing images into a multidimensional feature space. Instead of relying on predefined band ratios, ML models treat each pixel as a feature vector that includes spectral bands (e.g., VNIR and SWIR), derived indices, and supplementary variables such as elevation, slope, and aspect. This approach effectively “flattens” the image into a tabular format, with each pixel as a row and each variable or band as a column. This setup allows ML algorithms to detect complex, nonlinear relationships in the data that traditional index-based methods cannot capture.

This ability to process multiple inputs simultaneously offers a significant advantage over spectral indices. While indices like NDSI depend on a few spectral contrasts, machine learning models can integrate multiple spectral bands, terrain features from DEMs, thermal data, and even texture metrics. Consequently, ML methods excel in complex environments, including debris-covered glaciers, shadowed areas, and mixed surfaces. For example, the inclusion of SAR backscatter can improve performance in cloudy or low-illumination conditions, while DEM-derived features help distinguish debris-covered ice from surrounding rock. In addition, incorporating topographic variables (such as slope, curvature, and elevation) and thermal signals helps ML models distinguish glacier surfaces from surrounding terrain by physical context rather than by spectral differences alone. Figure 5 illustrates the typical ML workflow for glacier extraction.

Amsssong ML algorithms, SVM [65,66,67,68,69,70,71], and Random Forest (RF) [65,66,67,68,72,73,74,75,76,77] have consistently demonstrated strong performance in glacier mapping. SVM is particularly effective in high-dimensional feature spaces due to its ability to construct optimal decision boundaries (hyperplanes) that maximize class separation, even with limited training data. This makes it well-suited for remote sensing applications where training samples are often scarce or imbalanced. In contrast, RF, as an ensemble learning method, builds multiple decision trees and aggregates their outputs, improving robustness and reducing overfitting. Its strength lies in handling heterogeneous datasets and capturing complex interactions between variables, making it especially powerful when combining spectral, topographic, and textural features. Additionally, several studies have successfully employed foundational algorithms such as Decision Trees (DT) [70,78], and k-Nearest Neighbors (k-NN) [68,70,79] for land cover classification that includes glacial features, yielding good results due to the varied data inputs.

While the effectiveness of these models relies on the quality and representativeness of training data, ML remains more scalable and adaptable than index-based methods. It acts as an important intermediary between traditional static indices and fully automated deep learning techniques, facilitating feature-based learning necessary for mapping glaciers in diverse environments.

Table 3. Summary of studies highlighting the ML utilized for glacier mapping 1995–2025.

Study	Study Area	Purpose of Study	Input Dataset	Validation	ML Classifiers	Performance Metrics
[79]	Four glaciers in three climatic zones of the Tibetan Plateau	Debris-covered glacier identification	▪ Sentinel-1A; ▪ Landsat-8 OLI; ▪ ALOS AW3D30.	Manual delineation of glacier boundaries.	Boosted Trees, Subspace k-NN	OA: 41.70% to 99.94%
[78]	Eastern Pamir and Hunza Basin, Karakoram	Debris-covered glacier mapping	▪ Landsat-8 (OLI/TIRS); Sentinel-2; ▪ GDEM; ▪ NDSI; ▪ NDVI; ▪ NDWI; ▪ SGDBR; ▪ LST; ▪ [TIRS/(NIR/SWIR)].	GAMDAM, GI_P_K, and GLIMS	DT	OA: 91.11–98.28%
[80]	Gulkana, Wolverine, Lemon Creek, Sperry, South Cascade, and Emmons Glaciers, America	Automated snow cover detection on mountain glaciers	▪ Landsat-8/9; ▪ Sentinel-2; ▪ PlanetScope.	Cloud-free PlanetScope imagery used to manually delineate snow lines at each USGS Benchmark Glacier	Nine supervised ML models were tested; the optimal models were Nearest Neighbors and Support Vector Machine.	SVM (OA > 95%)
[72]	Central Karakoram Range	Object-based classification of debris-covered glaciers	▪ Landsat-8 OLI/TIRS; ▪ SRTM-DEM; ▪ GLIMS glacier database.	Two datasets were used: the Glacier inventory of the Pamir and Karakoram and the Global Land Ice Measurements Space Initiative (GLIMS) glacier database.	RF	RF average for 3 regions (OA: 99.81%, Kapp: 0.98)
[65]	North-western Himalayan, India	Snow and glacier feature identification	▪ Hyperion; ▪ Sentinel-2;	field data (GPS, Snow Fork, ASD Field Spectroradiometer)	SVM RF	OA (RF: 90.98%, SVM: 87.27), Kappa (0.88, 0.84, 0.76)
[73]	Eastern Pamir Plateau	Debris-covered glacier mapping	▪ SDGSAT-1; ▪ Sentinel-2; ▪ GDEM; ▪ ITS_LIVE; ▪ Glacier Velocity; ▪ LST; ▪ Textural Features.	1. CCI (Pamir Plateau and Karakoram Glacier Inventory) 2. CGI2 (Second Glacier Inventory of China) 3. RGI 6.0 4. GAMDAM (Glacier Area Mapping for Discharge from the Asian Mountains)	RF and DT	Hybrid performed better than individuals (OA: 87.13%, k: 0.80)
[66]	Columbia Icefield, Canada	Glacier area and volume change estimation	▪ Landsat-5 TM; ▪ Landsat-8 OLI; ▪ SRTM_DEM; ▪ NDVI; ▪ NDSI; ▪ NDGI; ▪ NDSII.	GLIMS Glacier Inventory, Manual Delineation, Supporting Data	SVM RF Maximum Likelihood Classifier (MLC)	RF (OA: 99.8%, Kappa: 0.99), SVM (OA: 99.7%, Kappa: 0.99), MLC (OA: 99.7%, Kappa: 0.99)
[74]	Eastern Pamir	Glacier mapping	▪ Landsat-8 OLI/TIRS; ▪ SRTM DEM	Manual digitization and the use of existing glacier inventories, using RGI 6.0	RF	RF (OA: 97.42–97.60%, Kappa: 0.95–0.96)
[67]	Tianshan Mountains	distribution and change mapping	▪ Landsat TM/ETM+/OLI; ▪ Sentinel-1; ▪ SRTM DEM; ▪ ERA5-Land data,	Manual digitization and the use of existing glacier inventories.	RF SVM Gradient Tree Boost (GTB), Classification & Regression Tree (CART)	RF (OA: 99.4–99.7%)
[75]	Antarctica (14 training regions, 8 test regions)	Supraglacial lake mapping	▪ Sentinel-2; ▪ TanDEM-X DEM.	manual interpretation and delineation of supraglacial lakes from Sentinel-2 satellite imagery.	RF	RF (Average F1: 86%, Average Kappa: 0.86
[68]	Gilgit-Baltistan, Pakistan and Shaksgam valley, China	Hierarchical mapping of glacier surfaces	▪ Sentinel-1; ▪ Sentinel-2; ▪ Landsat-8; ▪ ALOS World 3D 30 m mesh (AW3D30) DEM.	RGI 6.0 was used as a primary reference.	K-NN, SVM, Gradient Boosting (GB), DT, RF Multi-Layer Perceptron (MLP)	RF (OA: 97%, Kappa: 0.95)
[76]	Alps, French	Automatic detection of glacier snow lines	▪ Sentinel-2; ▪ EU-DEM v1.1.	Manual annotation of glacier snow lines.	K-means clustering, Gaussian Mixture Models (GMM), RF, MLP	RF (OA: 99.8)
[69]	Peru	Glacier mapping	▪ Landsat-8 OLI; ▪ SRTM DEM.	Manual digitization and the use of existing glacier inventories using RGI 6.0.	RF, SVM, K-NN.	K-NN (OA: 78–96%)
[81]	Eastern Pamir and Nyainqentanglha, China	Debris-covered glacier mapping	▪ Landsat-8 OLI/TIRS; ▪ Sentinel-1; ▪ SRTM DEM.	Manual digitization and the use of existing glacier inventories using RGI 6.0.	RF	RF (OA: 98.14%, Kappa: 0.97)
[77]	Parlung Zangbo basin, southeastern Tibetan Plateau	Glacier facies mapping	▪ Landsat-8; ▪ TanDEM-X 90 m DEM; ▪ SRTM DEM.	Manual interpretation and selection of training samples from high-resolution satellite imagery (Gao Fen-1 (GF-1)).	RF	RF (OA: 98.6%, Kappa: 0.98)
[70]	Marguerite Bay, Antarctic Peninsula	Land cover classification	▪ Sentinel-2A; ▪ Landsat-8 OLI.	Manual digitization and the use of existing glacier inventories.	RF, DT, SVM, K-NN.	SVM (pixel-based): OA: 97.31%, F1: 88.35; SVM (object-based): OA: 94.39%, F1: 81.43

outlines the various ML algorithms and datasets used for glacier extraction.

3.3. Deep Learning Models in Glacier Mapping

The emergence of deep learning in glacier mapping marks a significant shift from traditional feature-based classification to end-to-end representation learning. Unlike earlier methods that relied on manually engineered inputs, DL models automatically extract hierarchical spatial–spectral features directly from raw imagery, enabling the precise delineation of complex glacial features such as debris-covered and mixed-surface glaciers.

DL in glacier mapping was pioneered by CNNs, which effectively extract spatial hierarchies from imagery (e.g., studies [82,83,84,85]). These models process image tensors, maintaining spatial structure while learning local and global context.

The mechanism uses convolutional operations with learnable kernels: low-level filters capture edges, mid-level filters detect textures such as crevasses, and high-level filters encode structures such as glacier extent. Localized, weighted operations help networks learn translation-invariant patterns, essential for mapping features across different locations. After convolution, activation functions such as ReLU introduce non-linearity for modelling spectral relationships, while pooling (max or average) reduces resolution, expands the receptive field, cuts computational costs, and boosts robustness to noise, shadows, and variability.

As boundary delineation in challenging environments grew, the field adopted encoder–decoder architectures like U-Net and variants. The encoder extracts features through convolution and pooling, while the decoder builds segmentation maps via up-sampling. Skip connections combine deep semantic features with shallow details, effective for glacier segmentation. Researchers [86,87,88,89,90,91,92,93,94,95,96,97,98] used these architectures to map glaciers globally and monitor features such as grounding lines on the Greenland and Antarctic ice sheets.

Recognizing the unique spectral and morphological characteristics of glacial environments, the scientific community has moved beyond general-purpose models toward domain-specific architectures. A prime example is GlacierNet [99], explicitly designed for alpine glacier mapping, later refined in GlacierNet2 [100].

Performance is further optimized through structural design choices while modules like DeepLabv3+ utilize dilated convolutions to capture multi-scale spatial features, ranging from melt ponds to entire glacier basins, without losing resolution, which is crucial for managing glacial landscape heterogeneity.

Recent advances favor transformer-based models over CNNs, which focus on local features. Transformers use self-attention to capture long-range dependencies, helping them recognize large-scale patterns such as glacier continuity. Hybrid models like LGT-U-Net [92] and GlaViTU [101,102] combine convolutional layers for local features with transformer modules for global context. Attention-gated architectures like AMD-HookNet [103] improve feature representation for complex calving fronts.

Like ML, DL fuses multimodal data, using deep features for better classification. The typical DL workflow for glacier classification (Figure 6) shifts from general CNNs to specialized, hybrid models (

Table 4. Summary of studies highlighting the DL utilized for glacier mapping 1995–2025.

Study	Study Area	Purpose of Study	Input Dataset	Validation	DL Models	Performance Metrics
[86]	Central Karakoram	Debris-covered glacier Mapping	▪ All Landsat-8 Bands; ▪ ALOS DEM; ▪ slope-azimuth index; ▪ slope angle; ▪ tangential curvature; ▪ profile curvature; ▪ Nonspherical curvature.	Manually modified GLIMS and GAMDAM glacier inventories.	Glacier-CNN; DeepLabV3+; Res-Unet; Mobile-Unet; R2UNet; FCDenseNet.	DeepLabV3+ (IOU = 86.23%)
[100]	Central Karakoram	Alpine glacier mapping (clean, snow-covered, debris-covered)	▪ All Landsat-8 Bands; ▪ ALOS DEM; ▪ slope-azimuth index; ▪ slope angle; ▪ tangential curvature; ▪ profile curvature; ▪ Nonspherical curvature.	RGI 6.0 and other manually delineated datasets.	GlacierNet2	GlacierNet2 (IOU = 88.39%)
[99]	Karakoram and Nepal Himalaya	Debris-covered glacier (DCG) mapping	▪ 11 B Landsat-8 Bands; ▪ ALOS DEM; ▪ slope-azimuth index; ▪ slope angle; ▪ tangential curvature; ▪ profile curvature; ▪ Nonspherical curvature.	Modified glacier boundaries from the GLIMS database.	GlacierNet	Karakoram (IOU = 88.05%) Nepal (IOU = 77.96%)
[104]	Tomur Peak Region, Tianshan, Xinjiang	Alpine glacier mapping (clean, debris-covered, glacier lake)	▪ Landsat TM/ETM+	The Second China Glacier Inventory was used as a reference for training and validation.	Deeplabv3+	DeepLabV3+: Pure (IOU = 76%) Debris Covered (IOU = 48%) Glacier lake (F1 = 76.91%)
[105]	Gangshika region, Qilian Mountain	Regional glacier mapping and change analysis (2012–2023)	High-resolution satellite data: ▪ ZY; ▪ Gaofen; ▪ Landsat-8.	Manual visual interpretation was used to create and correct reference glacier boundaries.	Deeplabv3+	DeepLabV3+ (F1 = 95%, 2012)
[106]	Himalayas and Karakoram	Automatic glacier Segmentation/mapping	▪ Landsat-8 (Thermal) ▪ Sentinel-1 (SLC) ▪ Sentinel-2 (2,3,4, 8,12) ▪ ALOS DEM	RGI V6.0 dataset.	Supervised Attention U-Net (SAU-Net)	SAU-Net (Acc = 94.60%)
[107]	Horseshoe Island, Antarctic Peninsula	Glacier segmentation/mapping	▪ 5393 High-resolution UAV orthophotos.	A new glacier segmentation dataset was created from the UAV imagery.	Segformer, DeepLabv3+, K-Space Neural Network (K-Net)	Segformer (IOU = 98.73%) DeepLabv3+ (IOU = 99.09%) K-Net (IOU = 99.58%)
[87]	Hunza Valley, Karakoram	Debris-covered glacier (DCG) mapping	▪ Gaofen-2 (GF-2) ▪ Landsat-8 (4, 5, 10B)\LST ▪ ASTER GDEM V3\Slope ▪ KGIs	Manually digitized DCG boundaries from the GF-2 images.	Fully Convolution Neural Network (FCNN); DeepLabV3+; U-Net.	FCNN (IOU = 71.70%) DeepLabv3+ (IOU = 71.40%) U-Net (IOU = 76.70%)
[108]	Chandra-Bhaga basin, Western Himalayas	Supraglacial debris cover delineation	▪ Sentinel-1; ▪ Sentinel-2, (12B); ▪ ALOS PALSAR DEM; ▪ Landsat-8 (Thermal).	Manual delineation of debris cover was performed to create the training and validation labels.	Saliency-Guided Deep Neural Network (SGDNet)	SGDNet (Acc = 99.00%)
[88]	Antarctica (Getz Ice Shelf and others)	Automated glacier & ice shelf front extraction	▪ Sentinel-1; ▪ TanDEM-X DEM.	Not specified, but likely manual delineation of fronts for creating training labels.	U-Net	U-Net (F1 = 91.00%)
[89]	The THR, Tibetan Plateau	Long time-series glacier outline extraction (1986–2021)	▪ Landsat-5	Existing glacier inventory with manual adjustments	U-Net; U-Net++; SAU-Net; GlacierNet; U-Net+cSE; Band Radio; LandsNet; M-LandsNet.	Scene1: M-LandsNet (OA = 98.16%) Scene2 Test1: U-Net + cSE (OA = 97.72%) Scene2 Test2: M-LandsNet (OA = 96.95%) Scene2 Test3: M-LandsNet (OA = 96.94%)
[82]	Central Himalayan and Karakoram ranges	Debris-covered glacier mapping	▪ Sentinel-2; ▪ Landsat-8; ▪ Sentinel-1; historical Corona KH-4B.	Comparison of a manually delineated glacier inventory.	9. CNN	F1 = 89.2–93.7%
[101]	Global	Global-scale glacier mapping	▪ Sentinel-1; ▪ Sentinel-2; ▪ Landsat-8/9.	Utilized various global glacier inventories, including the RGI	GlaViTU	Most cases: IoU > 85% debris-rich areas: IoU > 75% clean ice regions: IoU > 90%
[102]	Multi-regional	Glacier mapping	▪ Sentinel-1; ▪ Sentinel-2; ▪ Landsat-8/9; ▪ DEM.	-	GlaViTU	Mean of IoU = 87.5%
[83]	Lower Himalayan region (Sutlej basin)	Glacial retreat delineation	▪ MODIS (LST, NDSI, NDWI) ▪ SRTM DEM ▪ ECMWF-R ▪ BESS	G + M25:M34lacier outlines from RGI were used as the basis for training data.	DNN	Acc = 95% AUC = 97%
[90]	Himachal Pradesh, India	Glacier identification and retreat mapping (1994–2021)	▪ Landsat 4\5 (2,3, and 4B) ▪ Landsat-8 (3,5, and 6B)	Not specified, likely manual delineation or use of existing inventories for training.	U-Net	F1 = 94.65%
[109]	Tanggula, Kunlun, and Qilian Mountains	Automatic glacier boundary extraction	▪ High-resolution Gaofen-6 PMS images.	Not specified, but likely manual visual interpretation to create ground truth labels.	Attention DeepLab V3+ with TTA	MioU = 98.21%
[91]	23 Greenland and two Antarctic outlet glaciers	Glacier calving front extraction	▪ Landsat-8	Manual delineation of 728 calving front positions.	U-Net	F1 = 90%
[110]	Bhutan Himalaya (Poiqu, northern Bhutan)	Automated glacial lake mapping	▪ Sentinel-1; ▪ Sentinel-2 (RGB, NDWI); ▪ Landsat-8; ▪ PlanetScope (RGB, NDWI).	Manual delineation of glacial lakes was used to create training and test datasets.	Deeplabv3+	Sentinel-1 F1 = 80% and 74% Sentinel-2 (RGB) F1 = 92% and 77% Sentinel-2 (NDWI) F1 = 89% and 84% Landsat-8 F1 = 88% and 86% PlanetScope (RGB) F1 = 97% and 91% PlanetScope (NDWI) F1 = 87% and 87%
[84]	Hunza Valley, Pakistan	Debris-covered glacier monitoring	▪ Sentinel-2 (2–11B) ▪ SRTM DEM	RGI 6.0 was used as the ground truth.	A new CNN-based architecture.	Sentinel-2 OA = 93.79% DEM OA = 88.50%
[92]	Qilian Mountain	Automated glacier extraction	▪ Sentinel-2 (2,3,4,8, and 11B); ▪ Sentinel-1 GRD (VV, VH); ▪ SRTM DEM.	The Second Chinese Glacier Inventory was used as the basis for the training dataset.	U-Net LGT + LGCB	OA = 97.20%
[111]	Greenland	Calving front delineation	▪ Landsat-8; ▪ Sentinel-1; ▪ Sentinel-2; ▪ TerraSAR-X; ▪ Envisat ▪ ALOS-1; ▪ ALOS-2;	Manual delineation of calving fronts.	Deeplabv3+ (ResNet) Deeplabv3+ (DRN) Deeplabv3+ (MobileNet)	Avg. Error (meter) 1. 61 m 2. 59 m 3. 154 m 4. 271 m
[112]	Alaska	Glacier mapping at fine temporal granularity	▪ Landsat (NDSI, NDVI, NBR, TCB, TCW)	RGI 6.0	GlacierCoverNet	OA = 97%
[113]	Western Kunlun Mountains	Rock glacier mapping and characterization	▪ Sentinel-2; ▪ InSAR; ▪ Google Earth.	Manual annotation of rock glaciers.	Deeplabv3+	-
[93]	Himalayas	Debris-covered glacier mapping	▪ Landsat-8 ▪ Sentinel-2	Manual delineation of debris-covered glaciers.	U-Net Deeplabv3+	Acc = 96.70% Acc = 98.07%
[12]	Svalbard	Glacier mapping	▪ Sentinel-1	RGI 6.0	3D conv LSTM	IoU = 96.40% IoU = 96.40%
[94]	Tianshan Mountains	Mountain glacier mapping	▪ Landsat-8 (RGB) ▪ SRTM DEM	Based on the Second Glacier Inventory of China.	2. U-Net with a cSE	Acc = 97.74%
[85]	European Alps (Valtellina and Val Masino)	Automatic rock glacier mapping	Grayscale SPOT 6	Manual interpretation of aerial orthophotos.	CNN	Acc = 88%
[95]	Greenland	Calving front monitoring	▪ Landsat-8	Manual delineation of calving fronts.	U-Net	mIoU = 94%
[114]	Nepal and China Himalaya (NCH), Karakoram and parts of western Himalaya (KWH)	Large-scale glacier mapping	▪ All Landsat-8 Bands; ▪ ALOS DEM; ▪ slope-azimuth index; ▪ slope angle; ▪ tangential curvature; ▪ profile curvature; Nonspherical curvature.	RGI 6.0	Modified version of GlacierNet2	NCH IoU = 75.25% NCH Iou = 81.15%
[96]	Hindu Kush-Himalayan region	Glacier segmentation (model interpretation)	▪ Landsat-7 (NDSI, NDVI, NDWI) ▪ SRTM DEM (Slope)	RGI.	U-Net	-
[97]	Siachen and South Lhonak Glaciers	Glacier segmentation and recession analysis	▪ SPOT ▪ Landsat-7 ETM+	Manual delineation or use of existing inventories.	U-Net	Acc = 92%
[103]	Benchmark Data	Glacier front segmentation	▪ CaFF	Manual annotation of glacier imagery.	AMD-HookNet	IoU = 74.40%
[98]	Antarctic Peninsula	Glacier calving front segmentation	▪ ERS-1/2; ▪ Envisat; ▪ RadarSAT-1; ▪ ALOS; ▪ TerraSAR-X; ▪ TanDEM-X.	Not specified.	Optimized U-Net	IoU = 87.24%

), illustrating AI’s growing role in glacier science. Note that the studies of the Glacial Lake have been moved to Supplementary File S1 due to the large number of studies in DL.

4. Discussion and Future Research Directions

The extraction of glaciers from remotely sensed data presents considerable challenges due to the complexity of mountainous landscapes and the heterogeneous nature of glacier surfaces, including shadows, debris cover, snow, meltwater, variable albedo, and frequent cloud obstruction. The extraction of glaciers from remotely sensed data has evolved significantly over recent decades, shifting from traditional spectral ratio techniques to more advanced ML and DL frameworks. Each method offers distinct advantages but also presents inherent limitations, particularly when applied across diverse glacier environments. However, comparing the results of all the studies reviewed is challenging due to differences in study areas and datasets. Nonetheless, it is feasible to evaluate select spectral indices, ML, and DL techniques to provide a comprehensive overview for researchers in this field. Studies using deep learning make up about 56% of the studies reviewed, followed by spectral indices for glacier mapping at 24%, and finally, studies involving machine learning at around 20%, as shown in Figure 7. The distribution of research focus across different glacial features reveals a clear methodological hierarchy. The research focus across glacial features reveals a clear hierarchy in Figure 8. Deep learning is the most versatile, leading in all areas, especially in clean glacier mapping (19 studies) and glacier lake detection (11 studies). Spectral indices are crucial for clean glaciers (10 studies) and snow-covered glaciers (4 studies) due to high reflectance. For complex tasks like debris-covered glaciers, machine learning (5 studies) and deep learning (8) are favored over indices (3 studies) to differentiate spectral similarities with terrain. Monitoring calving fronts mainly relies on deep learning, emphasizing its ability to handle the complexity of ice shelf margins.

4.1. Spectral Indices: Performance, Challenges, and Future Directions

Spectral index-based techniques remain among the most widely used methods for glacier extraction, particularly due to their computational simplicity and effectiveness in identifying clean or snow-covered glaciers. BRs, such as the NDSI, leverage the strong reflectance of snow and ice in the visible spectrum and their absorption in the shortwave infrared (SWIR), making them especially suitable for large-scale monitoring with medium-resolution sensors like Landsat and Sentinel.

Despite their popularity, traditional spectral indices face significant limitations when applied to debris-covered or mixed-surface glaciers, where spectral confusion with surrounding terrain reduces classification accuracy. To overcome these challenges, researchers have proposed modified indices, alternative BRs, and spectral unmixing methods, often in combination with auxiliary datasets such as DEMs and LST.

Within the reviewed literature, 65% of studies employed the Normalized Difference Snow Index (NDSI) as the primary method for glacier delineation, often benchmarking its performance against other or newly proposed indices. This underscores NDSI’s longstanding role as the baseline tool for distinguishing snow and ice. However, approximately 42% of studies introduced novel indices that demonstrated improved accuracy, while a smaller subset combined optical data with Digital Elevation Models (DEM), LST, or Synthetic Aperture Radar (SAR) data to enhance classification reliability.

A consistent theme across the reviewed studies is the critical role of threshold calibration—small variations in threshold values can substantially alter glacier extent estimates, particularly in debris-covered or shadowed regions. About 25–30% of the studies integrated spectral indices with auxiliary variables such as DEM and LST to improve mapping accuracy in complex terrain, emphasizing the growing recognition that spectral information alone is insufficient for robust glacier mapping. Roughly one-third of the studies proposed new or modified indices (e.g., ANDSI, AGEI, SWI, RDRI) to overcome NDSI’s limitations, particularly for differentiating glaciers from water bodies, snow, or supraglacial debris.

While traditional indices like NDSI perform reliably in clean-ice conditions, their effectiveness decreases in mixed or debris-laden environments, where hybrid or newly developed indices consistently outperform. Across studies, there is broad agreement that spectral indices remain fast and computationally efficient, but their transferability across regions and sensors remains a challenge, with many indices proving highly context dependent. Collectively, these findings indicate a paradigm shift toward context-specific, multi-index approaches and the integration of topographic and thermal variables for improved accuracy.

To illustrate these trends, several key studies provide representative examples.

For instance, in China, Ref. [46] developed the Normalized Difference Ice-Snow Index (NDISI) using multi-angle BRDF observations to enhance discrimination between snow, ice, and metamorphic ice-snow. The index, derived from reflectance in the 560–580 nm and 1600–1680 nm bands, was validated using spectral measurements from regions such as Altay, Qinghai Lake, and the Babao River Basin. Field investigations confirmed that NDISI provided superior separation of snow, ice, and ice-snow types compared to traditional indices.

Similarly, ref. [51] evaluated multiple BR approaches (NIR/SWIR and Red/SWIR), NDSI, and multisensory data (Landsat-9 and Sentinel-2). They reported that a threshold of 2.75 was optimal for both BRs in clean glacier ice mapping, while NDSI performed best at a threshold of 0.40. Among all tested methods, the Sentinel-2 MSI NIR/SWIR BR consistently outperformed others, particularly in shadowed and debris-covered regions, due to its higher spatial resolution and superior radiometric calibration.

Other studies compared multiple automated methods in parallel [51,55]. For example, Ref. [63] examined glacier area changes in the central Southern Alps, New Zealand (1978–2002), using ASTER imagery, historical aerial photographs, and field surveys. The research tested three automated classification techniques—NDSI, the ASTER BR (3/4), and supervised maximum likelihood classification. Among these, the ASTER 3/4 BR achieved the highest accuracy (97.01%), while NDSI reached 96.74%. Despite high overall performance, automated methods struggled with debris-covered glaciers, necessitating manual digitization for refinement. Validation against photo-interpreted outlines and GPS-based field mapping confirmed a ~17% reduction in glacier area, highlighting both the effectiveness and limitations of multispectral approaches.

Several studies further integrated spectral indices with DEM, LST, and SAR data to refine classification results [53,58,63]. For instance, Ref. [63] investigated glacier retreat in the Nevados glacier group, Peru, between 1975 and 2010 using Landsat-5 imagery, DEM data, and meteorological records. Clean ice was identified using NDSI > 0.5, whereas debris-covered areas were delineated through decision-tree rules involving NDSI < 0.5, slope < 24°, and specific LST thresholds. Results validated against datasets from the Autoridad Nacional del Agua (ANA) revealed a consistent retreat in both clean and total glacier areas, confirming significant glacier shrinkage in the Cordillera Blanca over recent decades.

Overall, the reviewed body of research emphasizes that while spectral indices remain central to glacier mapping, integrative and adaptive methods are increasingly essential for capturing the complexity of glacier environments. Future research should focus on addressing several key gaps identified across the literature, including the development of universally robust indices capable of performing reliably across diverse glacier environments, particularly debris-covered and shadowed regions, and the establishment of standardized thresholding protocols to enhance reproducibility and comparability across regions and sensors. Moreover, integrating multi-sensor datasets that combine hyperspectral, optical, and Geomorphometric data can significantly improve mapping precision and cross-sensor consistency. Expanding systematic validation through high-resolution or field-based ground-truth data is also crucial to minimize reliance on visual interpretation and enhance the reliability of automated classifications. Finally, implementing long-term glacier monitoring frameworks using cloud-computing platforms such as Google Earth Engine (GEE) will support efficient time-series analysis and automation. By focusing on these priorities, future studies can improve the accuracy, transferability, and operational applicability of glacier mapping methods, ultimately contributing to a deeper understanding and more effective monitoring of glacier changes under ongoing climatic variations.

4.2. Machine Learning: Performance, Challenges, and Future Directions

ML techniques represent a significant methodological advancement over traditional spectral indices, offering a more robust and data-driven framework for glacier classification. By leveraging supervised algorithms, ML models can learn complex, non-linear relationships between spectral data, topographic variables, and land cover classes, providing a more nuanced and accurate approach than fixed-threshold methods. These models have proven particularly effective in handling the inherent complexity of glacial environments, where features like debris-covered glacier and shadowed areas often lead to spectral confusion.

The application of ML in the reviewed literature shows a clear progression from foundational pixel-based classifiers to more sophisticated ensemble and object-based methods. Early or comparative studies often employed algorithms such as SVM, DT, and K-NN. While effective, these single classifiers can be sensitive to noisy data. A major leap in performance was achieved with the widespread adoption of ensemble methods, with RF emerging as the most dominant and successful algorithm. Approximately 60% of the reviewed ML studies utilized the RF classifier. Its strength lies in its ability to handle high-dimensional data and effectively integrate multisource inputs—such as optical bands, SAR data, and DEM-derived parameters to produce highly accurate classifications.

Within the literature, there is strong agreement that ML classifiers, particularly RF, consistently outperform traditional spectral index methods, especially in complex terrain. For instance, studies mapping glacier facies in the Tibetan Plateau and land cover in the Antarctic Peninsula reported overall accuracies exceeding 98% and 97%, respectively, using RF and SVM models. A clear pattern is the increasing integration of ancillary data; over 80% of the ML studies incorporated topographic variables from DEMs (e.g., slope, aspect, curvature) as crucial input features, confirming that spectral information alone is insufficient for robust glacier mapping.

Despite the high accuracies reported, disagreements and challenges persist. There is no universal consensus on the optimal set of input features, with different studies finding success with varied combinations of spectral, textural, and topographic data. The performance of ML models remains highly dependent on the quality and representativeness of the training data, and the manual effort required to create these datasets is a significant bottleneck. Furthermore, the persistent challenge of accurately mapping debris-covered glaciers remains a key limitation, even for advanced ML models. A significant research gap is the lack of standardized validation protocols and benchmark datasets, which makes direct comparison of model performance across different studies difficult.

4.3. Deep Learning: Performance, Challenges, and Future Directions

DL represents the current state-of-the-art in glacier mapping. By employing deep neural networks, these methods can automatically learn intricate spatial and spectral hierarchies directly from raw imagery, making them exceptionally powerful for complex semantic segmentation tasks. DL models have demonstrated superior performance in accurately delineating glacier boundaries, especially in challenging environments that confound simpler methods, such as debris-covered glacier, deep mountain shadows, and dynamic calving fronts.

The evolution of DL models in the reviewed studies shows a clear and rapid progression. Initial studies leveraged foundational CNNs, but the field quickly standardized on encoder–decoder architectures, with U-Net becoming the most ubiquitous and influential model. Its unique architecture, featuring skip connections that preserve fine-grained spatial information, is perfectly suited for precise boundary delineation. Over 50% of the reviewed DL studies employed a U-Net or a U-Net-based variant, cementing its role as the de facto standard for glacier segmentation. Building upon this, more advanced architectures such as DeepLabv3+, which uses convolutions to capture multi-scale context, have also been successfully applied, often achieving top-tier performance in comparative studies. For instance, a study in the Central Karakoram benchmarked seven different DL models and found that DeepLabv3+ achieved the highest Intersection over Union (IoU) of 86.23% for mapping debris-covered glaciers.

A significant trend evident in the literature is the development of highly specialized, domain-specific models tailored explicitly for glaciological challenges. Researchers have moved beyond adapting off-the-shelf models to creating bespoke architectures like GlacierNet and its successor, GlacierNet2. These models are specifically designed to handle the unique characteristics of clean, debris-covered, and snow-covered ice by integrating multisource data, achieving reported IoU scores as high as 88.39%. More recently, cutting-edge research has begun to incorporate transformer architectures (e.g., in GlaViTU, LGT-U-Net) and specialized attention mechanisms (e.g., in AMD-HookNet) to better capture global context and refine feature representation, pushing the boundaries of segmentation accuracy even further.

Across the reviewed studies, there is overwhelming agreement that DL models consistently outperform both spectral indices and traditional ML classifiers. A clear pattern is the mandatory integration of ancillary data; nearly all DL studies (over 90%) incorporate Geomorphometric Data alongside optical imagery to achieve state-of-the-art results. However, significant challenges remain. The primary bottleneck for all supervised DL models is their reliance on large, accurately labelled training datasets, the creation of which is time-consuming and requires expert knowledge. The computational cost of training these complex models is also a considerable factor. Furthermore, while models often perform exceptionally well in their specific study regions, their transferability to new, geographically distinct areas with different glacier characteristics remains a major research gap. Disagreements persist on which specific architecture is universally superior, suggesting that model performance is often task- and data-dependent.

DL models perform well in complex glaciological conditions where traditional methods often fail. In clean-ice areas with high spectral contrast, they provide only minor improvements over spectral indices. However, on debris-covered glaciers, where spectral similarities between ice and moraines limit the effectiveness of index-based techniques, DL models rely on spatial texture and morphological cues learned from convolutional filters to improve classification accuracy. Similarly, in areas with topographic shadows, DL models leverage information from surrounding terrain to identify glaciers, overcoming spectral limitations. Their ability to include multiple data sources, including factors derived from DEMs such as slope and aspect, further enhances performance by accounting for physical factors that influence glacier distribution. These features make DL particularly effective for diverse and changing glacier systems, even though they require more data and computational power. DL is most useful in complex environments, while simpler methods may still work for clear, well-defined glacier surfaces.

In this context, the choice of method mainly depends on glacier characteristics and the complexity of the data. Spectral indices are sufficient for clear, well-defined glacier surfaces with high spectral contrast. Machine learning techniques improve accuracy in moderately complex situations by combining various features, while deep learning approaches are essential in highly diverse conditions, such as debris-covered, shadowed, or actively changing glacier systems.

Looking ahead, several opportunities exist to advance glacier extraction. First, hybrid methods that integrate spectral indices, unmixing, and DL across optical, thermal, SAR, and LiDAR datasets offer strong potential to address the limitations of individual techniques. Second, multi-temporal and time-series analyses can improve boundary detection by distinguishing permanent glaciers from seasonal snow and by capturing inter-annual changes linked to climate variability. Third, advances in DL methodologies, including transfer learning and semi-supervised learning, can reduce dependence on large, annotated datasets, while the development of explainable AI frameworks will enhance interpretability and trust in these models. Finally, the increasing availability of open-access, high-resolution satellite data (e.g., Sentinel-1/2, Landsat-9, ICESat-2, GEDI) combined with cloud computing platforms like GEE provides an unprecedented opportunity to operationalize automated, near-real-time glacier monitoring systems. Such systems will be vital for supporting scientific understanding, water resource management, and climate change adaptation strategies at regional to global scales.

4.4. The Geomorphology-Classification Conundrum

In addition to clean, snow-covered, debris-covered glaciers, and supraglacial lakes, rock glaciers are an important cryosphere landform closely related to glaciers and should be explicitly considered within glacier mapping frameworks. In debris-dominated environments, it is important to distinguish among several related but distinct surface types, including debris-covered glaciers with an active, continuous ice core, buried or residual ice within debris-dominated terrains, dead ice bodies, and rock glaciers. Rock glaciers are debris-mantled bodies that often contain substantial volumes of glacier ice and, in many mountain regions, originate from debris-covered or stagnating valley glaciers rather than exclusively from permafrost processes.

Although much of the recent machine learning literature treats rock glaciers primarily as indicators of permafrost [115,116,117,118], growing glaciological evidence shows that glaciers, debris-covered glaciers, and ice-rich rock glaciers are genetically linked by shared ice content and evolutionary pathways [119]. However, these features differ in ice continuity and dynamics, with debris-covered glaciers typically remaining active, buried ice and dead ice representing stagnant or degrading remnants, and rock glaciers characterized by creep-driven movement.

From a remote sensing perspective, this distinction is critical, as most methods, including optical, SAR, and ML/DL approaches, primarily detect surface characteristics such as texture, reflectance, and morphology rather than directly identifying subsurface ice continuity. Consequently, these different ice-related landforms may appear similar in imagery, and high classification accuracy does not necessarily imply glaciological validity.

Consistent with guidance provided in the GLIMS framework [24,25,26], rock glaciers should therefore be included in comprehensive glacier mapping efforts. Their exclusion can lead to systematic underestimation of glacier-related ice volume and misinterpretation of long-term cryosphere change, while their misclassification highlights the need for careful interpretation of automated mapping results.

Optical imagery is particularly limited by spectral similarity between debris-covered ice and surrounding moraines, while SAR data capture surface structure but cannot reliably confirm internal ice continuity, and ML/DL models primarily learn surface patterns rather than subsurface properties.

4.5. Automating Historical Glacier Inventories for Mapping

While recent ML and DL approaches demonstrate impressive performance in automated glacier delineation, a recurring limitation identified in the literature is the insufficient integration of historical glacier records. Several studies develop region-specific models without explicit reference to past inventory data, despite the fact that glaciological history, including long-term retreat trends, hypsometry, and aspect-controlled melt patterns, is critical for understanding glacier response to climate change. Linking automated methods to established inventories such as GLIMS and RGI ensures temporal consistency, facilitates cross-regional comparison, and allows new outlines to be interpreted within a broader climatic and geomorphological context.

Recent glacier-focused studies have begun to address this gap by coupling ML-based classification with existing inventories and long-term datasets. For example, Alifu et al. [68] and Baraka et al. [120] integrated multi-sensor satellite data with inventory-based validation for debris-covered glacier mapping, while Bouchayer et al. [121] and Qiu et al. [122] emphasized long-term glacier change analysis using automated methods anchored to historical outlines. More recent work [123] further demonstrates the value of inventory-linked deep learning for ensuring model transferability across regions and decades. These studies clearly indicate that future advances in automated glacier mapping should prioritize tighter integration with established glacier inventories and historical records.

4.6. Limitations of Performance Metrics in Glacier Mapping

Although metrics like Overall Accuracy (OA), F1-score, and Intersection over Union (IoU) are standard for evaluating ML and DL models, they present notable limitations in glacier mapping. Since these metrics are usually calculated on a per-pixel basis, they often fail to accurately reflect spatial errors along glacier edges, which are vital for precise mapping.

Small boundary shifts can lead to minor changes in IoU but significantly impact glacier area estimates. This is particularly true for fragmented glacier tongues and narrow ice margins, and it is even more problematic for debris-covered glaciers. The surface variability and gradual transitions between ice and land lead to ambiguous class boundaries. Consequently, models may demonstrate high per-pixel accuracy without accurately capturing the true extent of glacier ice.

Additionally, thin or stagnant ice bodies, such as buried or dead ice, can be hard to tell apart from nearby debris. This leads to consistent classification uncertainties that standard metrics do not fully capture. Therefore, high reported accuracies do not always mean reliable glacier inventories or meaningful mapping.

To improve the reliability of glacier mapping evaluations, assessment strategies beyond pixel-based metrics are important. Boundary-oriented measures, such as error metrics and boundary displacement analysis, enable the detection of errors along glacier edges. The validation methods, such as region-specific or facies-based checks (e.g., clean ice vs. debris-covered ice), provide important information on model performance under different surface conditions. Quantifying uncertainty, e.g., area uncertainty estimation and sensitivity analysis, is crucial in glaciology for the precise interpretation of results. The selection of evaluation metrics should be application-driven: pixel-based metrics may suffice for large-scale mapping, but boundary-sensitive and uncertainty-aware approaches are needed for detailed glacier change analysis and hazard assessment.

4.7. Impact of Data Fusion on Model Performance

The choice of input data is a determinant factor in the performance of any ML or DL model for glacier mapping. The reviewed studies demonstrate a clear and consistent trend: multi-source data fusion significantly enhances classification accuracy by overcoming the inherent limitations of individual sensor types. This section analyzes the performance of models based on five categories of data inputs: optical only, Synthetic Aperture Radar (SAR) only, and their various combinations with geomorphometric data.

• Optical Datasets:

Models trained exclusively on optical data, such as Landsat or Sentinel-2 imagery, can achieve high accuracy, particularly for clean ice delineation. For instance, studies using U-Net architectures with only Landsat or Sentinel-2 inputs have reported F1-scores as high as 94.65% and overall accuracies exceeding 95% [87,90,109]. While effective, the performance of optical-only models is fundamentally constrained by meteorological conditions, such as cloud cover, and they often struggle to accurately distinguish debris-covered glacier tongues from the surrounding terrain due to spectral similarities.

• SAR Datasets:

The primary advantage of SAR data, like data from Sentinel-1, is the ability to penetrate clouds and operate regardless of solar illumination. This all-weather capability makes the data invaluable for consistent monitoring. Studies relying solely on SAR have demonstrated strong results, with a U-Net model achieving an Intersection over Union (IoU) of 87.7% [124] and another study using Sentinel-1 SAR data reaching an F1-score of 93.3% [125]. However, SAR-only approaches are susceptible to challenges such as speckle noise, layover and shadow effects in steep terrain, and ambiguity in backscatter signals caused by variations in surface wetness.

• Geomorphometric and Optical Fusion:

The combination of optical imagery with geomorphometric parameters derived from Digital Elevation Models (DEMs) represents a significant leap in performance, especially for mapping debris-covered glaciers. Topographic features like slope, curvature, and elevation provide critical context that helps differentiate spectrally ambiguous debris-covered ice from stable rock debris. This approach yields excellent results, as seen with the GlacierNet2 model, which achieved an OA of 88.39% using Landsat and ALOS DEM data [100].

• Optical and SAR Fusion:

Fusing optical and SAR data combines the rich spectral information from the former with the structural and textural sensitivity of the latter, producing a more resilient input stack that partially compensates for the weaknesses of each sensor type in isolation. Optical imagery supplies high spectral resolution for clean-ice discrimination, while Sentinel-1 SAR backscatter enables cloud-independent observations and captures surface-roughness contrasts that help separate debris-covered ice from surrounding terrain. This complementarity is demonstrated by Lin et al. [82], who trained a CNN on a combination of Sentinel-2, Landsat-8, and Sentinel-1 imagery to map debris-covered glaciers across the Central Himalayan and Karakoram ranges, achieving an F1-score of 89.2–93.7%. At a larger scale, Maslov et al. [101] showed that adding SAR coherence and backscatter from Sentinel-1 to optical Sentinel-2/Landsat inputs consistently increased accuracy across all tested regions, with their GlaViTU model achieving an IoU exceeding 0.85 globally and above 0.90 in clean-ice-dominated areas.

• Integration of Geomorphometric, SAR, and Optical Fusion:

Across the reviewed literature, the tripartite fusion of optical, SAR, and geomorphometric data consistently yields the most accurate and reliable results. This combination provides a comprehensive characterization of the glacierized environment, leveraging spectral, textural, and topographic information simultaneously. This approach allows models to effectively map both clean and debris-covered ice under a wide range of conditions. The superiority of this method is evident in the highest reported performance metrics; for example, a study utilizing an RF model with Landsat, Sentinel-1, and a DEM achieved an overall accuracy of 98.14% with a Kappa of 97% [81], while another DL study reported an F1-score of 97.2% using a similar data fusion strategy. This multi-sensor approach has clearly become the best standard for achieving state-of-the-art performance in automated glacier mapping.

The Sankey diagram in Figure 9, which analyzes research trends in glacier mapping, reveals a clear focus on certain features and a strong preference for specific modelling techniques. Clean glaciers are the most studied feature, likely serving as a baseline for model development, followed by glacial lakes and debris-covered glaciers, which are significant due to the hazards they pose. Snow-covered glaciers receive the least attention. Methodologically, there is an overwhelming dominance of DL, marking a shift towards sophisticated, automated feature extraction. Traditional ML remains relevant, while older spectral indices are still widely used, particularly for snow cover, where they are highly effective.

Drilling down into specific algorithms, the DL architectures U-Net and DeepLabV3+ have become the standard for mapping clean glaciers and glacial lakes. For ML applications, RF and SVM are the most common classifiers. The complex task of mapping debris-covered glaciers sees a more diverse methodological approach, with both DL and ML models being actively explored. The use of spectral indices is highly tailored to the target feature, with the (NDSI) being favoured for snow and clean ice, and various water-centric indices like NDWI being used for lakes. The limited effectiveness of indices on debris-covered terrain reinforces the necessity of advanced DL and ML models for this challenging task.

Research on remote sensing for glacier mapping shows a significant global imbalance (Figure 10). High Mountain Asia, which encompasses the Himalayas, Tibetan Plateau, and Karakoram, accounts for 57 studies. This prominence is due to its extensive glacier coverage, its importance as a water source, and the rapid climate-induced changes occurring there. Meanwhile, regions like Antarctica, Greenland, and the European Alps are underrepresented, with very few studies. Western North America, Arctic areas, the Caucasus, New Zealand, and low-latitude glaciers have received minimal research, often limited to only a few studies that take a multi-regional or global perspective.

This imbalance significantly influences model development and transferability. Most machine learning and deep learning models are trained on datasets from High Mountain Asia, where glaciers often have thick debris cover and marked seasonal variations. Consequently, these models might learn features unique to this region that do not generalize well to other environments, especially in areas with thin or sparse debris cover or clean ice conditions typical of polar and alpine zones. For example, studies conducted in polar regions such as Greenland [49,91] and Antarctica [75,88] often rely on different spectral and structural characteristics (e.g., calving fronts and clean ice) that are not well represented in High Mountain Asia-trained models. Therefore, the performance metrics reported in the literature may not reliably predict how these models will perform globally.

These findings emphasize the need for more diverse datasets and multi-regional validation. This would improve model reliability and ensure accurate glacier mapping in various climatic and geomorphological contexts.

4.8. Key Challenges and Future Research Needs

The literature review underscores several interconnected challenges that set key priorities for future glacier mapping research. These include difficulties in obtaining reliable, standardized ground truth data; challenges in integrating multi-source datasets such as DEM, SAR, and InSAR; and issues with maintaining FAIR (Findable, Accessible, Interoperable, and Reusable) data principles. Additional hurdles involve limited geolocation accuracy, restricted access to derived datasets, and the complexities of mapping debris-covered and transitional glaciers. Collectively, these problems impact the reliability, reproducibility, and transferability of models across different regions. Although individual studies focus on specific issues, synthesizing these challenges offers a clearer framework for advancing methods and data practices in automated glacier monitoring.

4.9. Operational Readiness of Glacier Mapping Approaches

While important progress has been made in glacier mapping methods, their practical use varies widely. Spectral index-based approaches remain the most useful for large-scale and near-real-time monitoring because they are low-cost, require little training, and work well with platforms like Google Earth Engine (GEE).

Machine learning methods can improve accuracy in moderately complex environments, but they require well-organized training datasets and careful feature selection, which can limit their adoption.

Deep learning methods show the best results in difficult conditions, such as debris-covered and mixed-surface glaciers. However, using them in practice is challenging due to their high computational needs, reliance on large, labeled datasets, and limited ability to transfer across different areas. As a result, many deep learning applications are still in the research or experimental phase rather than fully operational.

In practice, combining spectral indices, machine learning or deep learning models, and cloud-based platforms offers the best path for scalable and near-real-time glacier monitoring. Future efforts should aim to reduce data needs, enhance model generalization, and integrate automated workflows into operational systems.

5. Opinions and Perspectives

▪

While ML and DL now dominate the field, it is crucial not to dismiss the foundational role of spectral indices. For rapid, large-scale assessments, especially for clean ice and snow, indices like the NDSI remain computationally efficient and easy to implement. However, the literature clearly shows a move towards developing more robust indices, like the ANDSI, to tackle specific challenges like differentiating ice from water. The future of indices may not be as standalone tools, but as powerful input features for more complex ML and DL models, providing a baseline of spectral information that these models can then refine.

▪

The effectiveness of supervised ML and DL models for glacier mapping largely depends on the quality, consistency, and representativeness of the training data. Most research uses manually delineated glacier boundaries or existing inventories like the RGI and GLIMS as “ground truth.” However, these are not true ground truths, as they are based on semi-automated or expert methods with varying spatial resolutions, temporal consistency, and assumptions. A key issue is the systematic underrepresentation of debris-covered glaciers, which are hard to detect in optical imagery, leading to the omission or misclassification of glacier areas in training data. This creates bias, causing models to learn from inaccurate labels that do not reflect the true extent and state of glaciers. While in situ surveys provide more precise reference data, they are limited in scope and difficult to carry out. Consequently, current methods depend on imperfect reference datasets, which restrict reliable model comparison, regional generalization, and accurate glacier inventory development. Developing standardized, high-resolution, and physically consistent benchmark datasets that adhere to the FAIR principles is therefore a priority for the scientific community.

▪

The DL table shows a clear dominance of U-Net and its variants (e.g., Res-Unet, R2UNet, SAU-Net). While U-Net is a powerful architecture for image segmentation, the field would benefit from greater exploration of other model architectures. For instance, models that are more computationally efficient (like Mobile-Unet) could be crucial for large-scale, operational glacier monitoring systems. Furthermore, there is a need for more research into hyperparameter optimization and novel loss functions tailored to the specific challenges of glacier mapping, such as the imbalance between glacier and non-glacier pixels.

▪

While many studies now incorporate multiple data sources, such as optical imagery, SAR, and DEMs, which is a positive trend for overcoming the individual limitations of each sensor type, there is still much to explore in this area. Future research should focus on how to best fuse data with different spatial and temporal resolutions, the potential of incorporating less commonly used data types like hyperspectral or thermal imagery within DL frameworks, and the development of more sophisticated data fusion techniques that go beyond simple data stacking. Most studies focus on developing models for a specific region. There is a need for more research on the transferability of these models to other geographic areas with different glacier types and environmental conditions. A model trained on Himalayan glaciers, for example, may not perform well on glaciers in the Andes or the Alps due to differences in glacier morphology, climate conditions, and training data distribution [102].

▪

Debris-covered glaciers are among the most challenging to map due to their spectral similarity to surrounding terrain and their reliance on debris thickness, thermal insulation, and surface dynamics. Future studies should focus on physically based methods that combine thermal data, surface velocity, and geomorphological indicators to better detect active ice beneath debris.

▪

While many studies demonstrate the potential of ML and DL for automated glacier mapping, there are few examples of these models being implemented in operational, near real-time monitoring systems. Bridging the gap between research and operational use is a key challenge for the future.

6. Bias and Uncertainty in Glacier Mapping

6.1. Defining Bias in Remote Sensing

In glacier mapping, bias refers to systematic errors that consistently distort results in a specific direction, leading to overestimation or underestimation of glacier size. Unlike random noise, which varies, bias typically arises from inherent limitations in data collection, environmental factors, or algorithmic design. Recognizing and measuring these biases is crucial to making sure that observed changes in glacial mass truly reflect climate trends rather than errors introduced by the methods used.

6.2. Sources of Bias and Systematic Uncertainty

The transition from manual delineation to automated deep learning (DL) models has introduced specific forms of bias that this review categorizes as follows:

Dataset and Geographic Bias: Most state-of-the-art models are trained on high-quality, standardized datasets, often concentrated in specific regions (e.g., the European Alps). This results in a transferability gap, where models perform poorly on glaciers with different characteristics, such as tropical or high-latitude Arctic glaciers.

Ground Truth and Labelling Bias: Progress in glacier mapping is constrained by the difficulty of obtaining accurate ground truth data. Many studies rely on inventories such as RGI and GLIMS, which are essential but carry uncertainties. These inventories often diverge from expert delineations due to differences in mapping methods, resolution, and interpretation. Debris-covered glaciers are often underestimated or misclassified because surface debris obscures the ice in optical images. These issues introduce biases that affect supervised ML and DL models, which may inherit and reinforce inaccuracies. High performance metrics might reflect bias rather than true glaciological accuracy, thereby impacting model generalization across different areas or data qualities.

Atmospheric and Topographic Bias: Optical sensors such as Landsat and Sentinel-2 are often hampered by clouds and shadows, leading to the omission of glacier sections. SAR sensors reduce cloud contamination but introduce geometric distortions in steep terrain, potentially biasing glacier measurements.

Debris-covered Glaciers Mapping Bias: Debris-covered Glaciers pose a unique uncertainty because their surface appearance often does not reveal the conditions of the ice beneath. Optical and SAR-based techniques mainly detect surface features, which may not accurately reflect the presence, extent, or activity of the subglacial ice. This can cause misclassification of debris-covered glaciers, buried ice, or rock glacier systems, especially in areas with varying debris thickness and complex morphodynamics. Consequently, classification results and inventories may contain systematic errors that are not fully captured by standard accuracy metrics. Therefore, incorporating thermal observations, surface velocity data, and field validation is essential to lessen this uncertainty.

Model Architecture Bias: Relying too heavily on a limited set of model architectures, such as standard U-Net variants, can lead to systematic failures in detecting fine-scale features or transient supraglacial lakes. These models tend to favor smooth boundaries, leading to the underrepresentation of highly fragmented glacial snouts.

7. Conclusions

This review has charted the evolution of glacier mapping techniques using remote sensing data, tracing the progression from traditional spectral indices to the current era dominated by sophisticated ML and DL models. Our analysis reveals a clear and significant trend: the increasing adoption of automated, data-driven approaches has substantially improved the accuracy, efficiency, and scale of glacier delineation. Methods have evolved from the computational simplicity of indices like the NDSI, which remain useful for preliminary assessments, to the nuanced classificatory power of ML algorithms such as RFs and Support Vector Machines. The contemporary landscape, however, is unequivocally shaped by DL, with architectures like U-Net and its derivatives demonstrating unparalleled performance in semantic segmentation, particularly for challenging tasks like identifying debris-covered glaciers and calving fronts.

Furthermore, while the field has embraced data fusion, combining optical, SAR, and DEM data, there is significant room to develop more sophisticated fusion techniques and to explore the potential of underutilized data sources. The current over-reliance on a limited set of model architectures, primarily U-Net, highlights a need for greater diversity in model exploration, including a focus on computational efficiency and hyperparameter optimization to move from research to operational monitoring.

A major challenge in the literature is the persistent shortage of high-quality, transparent, and reusable ground truth data. This problem extends beyond quantity, encompassing issues such as data accessibility, thorough documentation, and accurate geolocation. Many studies lack specific geolocation details, such as exact decimal coordinates, which hampers reproducibility and reduces the usefulness of the dataset.

Providing machine-readable geographic coordinates (e.g., decimal latitude and longitude) for features would enhance interoperability, facilitate cross-study comparisons, and enable integration with emerging frameworks, including large language models and automated geospatial reasoning. Vague spatial references or map edge coordinates are less effective and do not meet FAIR standards.

Future glacier mapping relies not only on advances in machine learning but also on a cultural shift toward open, well-documented, and precisely geolocated datasets. Applying FAIR principles, improving ground-truth transparency, and enforcing strict geospatial standards are vital to scalable, trustworthy climate-related glacier monitoring.

In conclusion, the application of ML and DL has revolutionized glacier mapping, offering powerful tools to monitor these critical indicators of climate change. The future of the field lies in addressing the key limitations identified in this review. Progress will be driven not only by the development of more advanced algorithms but also by a community-wide focus on creating robust validation datasets, improving model transferability and uncertainty quantification, and bridging the gap to create operational, near real-time glacier monitoring systems. By integrating cutting-edge AI with improved data practices, the remote sensing community can provide the crucial data needed to understand and respond to the impacts of global climate change.