The primary imagery used for glacial landform mapping was an ESA Sentinel-2 image acquired on 9/17/2017 and freely downloaded from USGS EarthExplorer. This scene was selected because it covers the entire study area with no clouds and minimum seasonal snows and the date is not far from the field campaign. Compared to other freely accessible, commonly-used satellite imagery (e.g. Landsat TM, ETM+) in geomorphological mapping (e.g. [25] [46] [47] [48] [49] ), the Sentinel-2 imagery with a spatial resolution of 10 m in the RGB and near-infrared bands provides a better choice for years after 2015. The false-color composite (7-4-3 bands) image was imported to the ArcGIS environment where each geomorphological feature was then digitized/vectorized on screen. High resolution Google Earth imagery was frequently consulted for cross-checking and better identification of the land features with its flexible 3D visualization. Some small-scale landforms like moraine ridges were directly mapped in Google Earth and then exported and converted from the kml format to shapefiles in ArcGIS.

The NASA Shuttle Radar Topography Mission (SRTM) Global 1-arc second (~30 m resolution) Digital Elevation Model (DEM) was applied with hillshade effect to enhance the visualization of the terrain. It was used to help identify and correct the boundary of some glacial landforms with relief changes. It was also the input data for the derivation of basin boundary, contours, summits, and river channels in ArcGIS. Ridgelines were extracted automatically from the inverse DEM with the terrain analysis tool in SAGA. All of these topographic features were checked visually with the overlay of the Sentinel-2 image, and some were smoothened, truncated, and removed to create a clear representation of the topographic context. The mapped glacial geomorphological features include glacial valleys, moraine complex, moraine ridges, and trimlines, and are overlain on the grey-scale hillshaded DEM. All datasets and vectorized features were projected to the Universal Transverse Mercator (UTM) coordinate system Zone 44N. In the Xiata basin, GPS waypoints and panoramic photos from the field serve as validations for some on-screen interpretations of land features, especially when features have a dimension smaller than 10 m. In areas without field data, a combination of all available materials, namely the Sentinel-2 image, DEM, Google Earth, and maps from previous publications, was carefully examined to help interpret and delineate the landforms.

Relying on the differences in spectral and thermal reflectance characteristics, many methods have been developed to detect glaciers facies (e.g. [36] [50] [51] [52] ) and supraglacial debris cover, where the latter has been challenging due to its similarity with the surrounding terrain (e.g. [53] [54] [55] [56] [57] ). In this study, the classification of glacier surfaces, including clean ice, dirty ice, snow, slush zone, and debris cover, was attempted by applying an automated mapping algorithm proposed by [38] . The method utilizes 3 optical bands and 1 thermal band of Landsat 8 data and then calculates a band ratio for thresholding different types of the surface. The remote sensing analysis steps are outlined in a flowchart (Figure 3). Prior to the band math, 16-bit unsigned DN (digital number) values of the Landsat 8 optical and thermal bands need to be converted to the ρ_λ, “TOA (top of atmosphere) reflectance” and T, “at-satellite brightness temperature” with parameters from the metadata file of the Landsat scene. The conversion formulae are provided and updated by the USGS Landsat Missions, and the most recent version of the formulae was applied to obtain ρ_λ for optical bands of 1, 5, 6, 9 and T for the thermal Band 10 in this case (Figure 3). A late summer Landsat 8 OLI/TIRS scene acquired on 9/4/2017 was selected because of its minimal cloud cover and no snow cover to ensure only capturing the glacier features. Glacier boundaries were obtained from the Randolph Glacier Inventory Version 6.0 (www.glims.org/RGI/) and were used for subsetting the ratio image. Threshold values for dividing various surface types were slightly modified from the original divisions in [38] depending on this particular scene and are provided in Table 1.

The geomorphological configuration was investigated in detail at the Xiata site where field work was conducted. GPS tracks show the spread of the footprint at the top of the sparsely vegetated marginal moraines on both sides of the valley and across the outwash plain in front of the debris covered glacier snout (Figure 4(a)). The terminal moraine is located approximately 2 km down the clean ice

edge of the glacier. Based on geomorphological analysis and numerical dating of similar moraines in the TS, previous studies have suggested the timing of such formation during the LIA (e.g. [6] [17] [21] [58] [59] [60] [61] ). These LIA moraines not only mark the extent of the most recent glacier advances, but are also used to define the boundary of a proglacial environment in which intense and variable glacio-fluvial dynamics occur (Figure 4(b), Figure 4(c); [62] ). The proglacial area of the Aerqialeteer glacier covers about 0.8 km², and the outer wall of the LIA terminal moraine reaches down to 2720 m a.s.l., below the local treeline (~2900 m a.s.l.). Dense forests, dominantly composed of Picea schrenkiana, are distributed beyond the edge of the LIA moraines, whereas on top of the morainic terrains, a few pioneer species have sparsely established colonization after the vegetation elimination and creation of the moraine from the last glacier advance (Figure 4(d)). Such a contrast marks a distinct boundary on the satellite imagery. Other prominent features such as small proglacial lakes and subtle lateral moraine ridges can only be distinguished with help of higher resolution image, i.e. Google Earth and/or field panoramic photos (Figure 4(d), Figure 4(e)). The stream channel pattern is short-lived, indicating the frequent redistribution of glaciofluvial sediments and a high instability in the proglacial system. It is well reflected by comparing the 9/7/2007 Google Earth imagery with the 9/17/2017 Sentinel-2 image, which demonstrates a disappearance of a large proglacial lake (about 26000 m²) within the period (Figure 4(b), Figure 4(c)). Such dramatic changes could impose a potential glacial lake outburst flood (GLOF) hazard to the downstream. The Aerqialeteer Glacier has two tributaries contributing meltwater to the runoff in the proglacial area, but the mechanism and processes are remarkably different due to the difference in surface type (discussed in Section 4.3). The southern branch is a debris-covered glacier with stagnant ice bodies buried underneath the heavy amount of debris at the lower end, whereas the northern branch is a clean-ice active glacier. Several exposed ice cliffs are witnessed in the field at various parts underneath the surface debris cover which extends a few hundred meters further than the tongue of the clean ice glacier branch (Figure 4(d), Figure 4(f)). For ice cliffs next to the main drainage channel, the disintegration of the ice bodies is prominent at the base, showing active melting in the late summer (Figure 4(f)). The thickness of the debris-cover is visually estimated to be about 10 m - 20 m across various parts. The land surface patterns and processes within the proglacial environment are rapidly changing and represent a revenue that displays a continuity between the late Holocene glacial heritages and the present-day glacier characteristics.

Glacial valleys are delineated to confine the extent of the U-shaped cross-profile valleys worked by the LIA glaciers (Figure 5). The large body of viscous ice is powerful to abrade the bedrock to broaden and deepen the valley, leaving the eroded materials deposited as till or moraines at the valley floor. The resulting steep, smooth valley walls and gentle valley bottom characterize glacial valleys at the highest mountain areas, whereas transitioning to the lower terrain, fluvial valleys with V-shaped cross-sections often signify the non-glacial condition. In

our cases, the extent of glacial valleys is constricted to the recently deglaciated area not far (<5 km) from the margins of the modern glaciers. The maximum extent of the late Holocene portion of the glacial valleys was delineated based on the location of the fresh morainic deposition and the sharp change in slope transitioning to the relatively older deposition. Small valley channels outward the cirques are included as glacial valleys as well, and they often connect to the mainstream valley as truncated spurs (Figure 6). These young glacial valleys are primarily identified and mapped based on the DEM and Sentinel-2 imagery.

Moraine complexes and moraine ridges are positive relief formed by deposition of unstratified till materials from past glacial advances. For the purposes of describing the late Holocene glacial geomorphology, moraine complexes and ridges mapped in the two glacier basins are the first set of such landforms from the head of the valley, and the sequence of older moraines down in the valley is not discussed here. The key component in the moraine complexes category is the LIA moraines which clearly delimit the margins of the LIA glaciers. These moraines are composed of a large number of erratic boulders, gravels, and other debris, and are sparsely vegetated with limited soil development. Large boulders

with height more than 2 m can be commonly found at the surface of latero-frontal moraines (Figure 4(g)). The LIA terminal moraines can stand up to 30 - 40 m above the adjacent lower terrain and are often in arc-shape encompassing a moraine-dammed proglacial lake or incised by meltwater streams. Lateral moraines are sharp-crested and linearly located along the valley sides. Moraine ridges are outlined for the most prominent crests exhibited on the well-formed marginal moraines. From the field, several parallel, shallow ridges are found on morainic deposit at the steep slope of the northern valley side of the Aerqialeteer Glacier (Figure 4(e)). Moraine ridges are good approximation of the former stillstands of the glacier during the retreat. Hummocky moraines are also included in the moraine complexes category. Across the Muzart Pass which lies between the two basins, thick till deposits form an undulating surface of ground moraine. Such complexes are likely to suggest a stagnating status of former extensive valley glaciers at this location.

Trimline is a near-horizontal linear feature along the sidewall of a glacial valley that marks the higher stand to which the former glacier was able to erode. It is an indicator of the previous thickness of the glacier. During the LIA maximum, a thicker glacier advanced down the valley and impacted the surface that was in contact with the moving ice. In the studied basins, erosional actions removed surface vegetation, and as the LIA glacier retreated and its surface was lowered, a clear boundary is exposed to separate the vegetated surface and the eroded, vegetation-free eroded region. The presence of a sharp contrast in color and texture on valley slopes allows easy identification of trimlines in the late-Holocene glacial valleys. However, LIA trimlines do not always form on valley slopes or are not discernible especially when in conjunction with moraine formation. At the field site, a section of distinct trimline was found extending about 100 m on one side of the valley, clearly illustrating the abrasive marks where the former glacier encountered the resistant bedrock (Figure 4(a), Figure 4(d)). Other trimlines are mapped from inspection of satellite imagery and Google Earth.

Because the dynamics of a glacier are largely determined by the mass and energy exchange between the overlying atmosphere and the glacier surface, it is an important task to discriminate zones of different surface materials. Due to the high terrain relief in the Harijoriha glacier basins, direct field-based observations or measurements from the surface are extremely difficult. Fortunately, previous studies have long explored the feasibility to distinguish surface zones using remote-sensing data and have shown applicability of such methods in good agreement with field validation (e.g. [36] [38] [50] [51] [63] [64] ). The employment of a recently developed mapping algorithm on the studied glaciers using a Landsat 8 satellite image acquired near the end of the ablation season was able to generate a pixel-based classification of different glacier surface types, including clean ice, snow, dirty ice, slush zone, and debris covers [38] .

The output pattern of each distinct type of glacier surfaces reflects the general zonation of accumulation and ablation in glacier mass balance. As shown on the classified map (Figure 5), glaciers in both the Xiata basin and Muzart basin are predominantly composed of clean ice at the surface, indicating the ablation-induced mass loss exceeding the mass gain through the accumulation of snow. This can be explained from the acquisition date (9/7/2017) of the selected Landsat 8 image. This time of the year corresponds to the end of the ablation season, when the accumulation of snow from the previous winter has been mostly melted and the solid bare ice gets exposed to the surface before the new accumulation season starts. The snow zone where the winter snow does not melt throughout the summer is only present at limited extents, particularly at higher altitudes (above 5000 m a.s.l.) and north-facing slopes in the Muzart basin (Figure 5). In contrast, the highest elevations of the Xiata glaciers are below 5000 m a.s.l. and snow facies are almost completely absent at their surfaces, suggesting the lack of permanent accumulation zone at the end of ablation season. On the one hand, this observation implies that local topographic factors, such as elevation and aspect orientation, can strongly influence the evolution of glacier mass gain or loss over time. On the other hand, it may suggest other potential sources for accumulation other than the persisting solid precipitation that account for the mass input to the glaciers. More investigation is needed to examine the role of wind-blown snow or avalanching in contributing the mass gain to these shrinking glaciers. During the ablation season, the slush zone may appear as an area completely saturated with water and graded up glacier into the wet snow facies, as defined in [36] . The discrimination of this zone in the Harijoriha shows a small area along the margin of snow zones (Figure 5). This agrees with its concept as an ephemeral transition zone from the ice facies and the snow facies [36] . Crevasses were extracted when the mapping algorithm applied to glaciers in the Indian Himalaya [38] , however, they cannot be derived in our study area. Another prominent feature distinguished from the automated mapping is the supraglacial debris. Using the pre-defined glacier boundaries from the glacier inventory, extensive debris-covered regions were detected in the ablation areas of some valley glaciers (Figure 5, Figure 6). Dirty ice pixels, which delineate the ice mixed with debris, are often found in between the debris areas and the clean-ice. Large patches of dirty ice are clearly seen at the lowest portion of the two main glaciers, the Aerqialeteer and the Muzart Glaciers (Figure 6). From satellite images and field observations, this part of the glacier demonstrates a thick and rugged debris surface with dead ice buried underneath (Figure 4(d), Figure 4(f)), whereas the debris surface with active ice up glacier is smoother and shows more streamlined flow features (Figure 5).

The results are considered to provide a detailed record of the late-Holocene glacial geomorphology and glacier surface features in the studied basins of the central Tian Shan. Remote sensing data were carefully examined when applied in the manual delineation of glacial landforms and in the automated classification of surface types. The optimal selection of Landsat 8 and Sentinel-2 images with no disturbance of snow and clouds were acquired at the end of ablation season, representing a complete and accurate picture for inspection and interpretation of the glaciers and the surrounding landscape. The acquisition dates (9/17/2017 for Sentinel-2 and 9/4/2017 for Landsat 8) are close to the days of field reconnaissance of last days in July 2017, which enables a plausible comparison between field checks and image extractions. Investigation in the field is strictly limited by the difficulties of the terrain, and using panoramic photos derived from drone images lends a broader view than the sights obtained from the ground. Uncertainties from different sources may influence the accuracy of the mapped glacial imprints. First, the resolution of the DEM and satellite imagery limits the size of land features that can be identified for mapping, and some small glacial landforms, such as trimlines and narrow moraine ridges, may not be distinguishable if present. Higher resolution images, if available by Google Earth, helped to detect these features as much as possible. The mapped glacial valleys and moraine complexes are larger features (>100 m) and less affected by the resolution of data sources. Second, the automated classification based on optical and thermal bands of Landsat 8 imagery could be an effective method, but the choice of threshold values to differentiate glacier surface types is somewhat empirical and requires multiple rounds of modifications where uncertainties might be introduced. The third source for ambiguities lies in the glacier boundaries which were used for subsetting the ratio image. Uncertainties or inaccuracies in the RGI glacier outlines may cause misrepresentation of the actual glacier extents, and when processing for surface classifications, the results may contain some errors. For examples, the rock outcrops at the northern slopes of the Muzart basin were not excluded from the glacier extent and were consequently misclassified as debris cover due to the similarity of the two in the spectral analysis. Increasing the accuracy level of the prerequisite glacier data offers a way to improve the accuracy of the mapping results.