| 16 Jun 2026
High-resolution glacier mapping reveals inventory biases and terrain controls on debris-covered glaciers in the Karakoram
Abstract. Accurate glacier inventories are fundamental for quantifying glacier change, estimating ice volume and assessing meltwater resources. However, medium-resolution inventories often fail to resolve critical glacier characteristics in topographically complex and debris-rich mountain environments. Here, we present the 2 m Karakoram Glacier Inventory (2mKGI), developed from high-resolution ZY-3 optical imagery, a co-registered ZY-3 digital elevation model (DEM) and auxiliary optical datasets through an integrated deep-learning and manual-refinement framework. The inventory identifies ~13,900 glaciers covering 21,261.8 ± 278 km², including 2,239.9 ± 82.8 km² of supraglacial debris, with an overall mapping uncertainty of ±4.7 %. Comparison with existing inventories reveals that previous medium-resolution products commonly underestimate glacier numbers while simultaneously overgeneralizing debris-covered glacier margins. These biases substantially influence glacier-count statistics, estimates of debris-covered area, and the interpretations of glacier-change signals. The newly identified glaciers are mostly <0.1 km². Despite their limited area, their thin ice and shorter response times may make them particularly sensitive to warming. Topographic analysis further demonstrates that supraglacial debris is preferentially distributed across low-elevation and low-slope glacier tongues, highlighting the strong controls of valley geometry, ice transport processes, and ablation-zone morphology on debris persistence. The 4 m ZY-3 DEM highlights that high-resolution topographic information improves the delineation of glacier-units, accumulation–ablation zone structure and debris-covered tongues by preserving steep headwalls, slope discontinuities, tributary junctions, local relief and low-gradient terrain. The 2mKGI will provide a high-resolution geometric and topographic benchmark for glacier-change assessment, ice-thickness inversion, glacier-evolution modelling and next-generation automated glacier mapping in the Karakoram.
The authors present a comprehensive and high-resolution glacier inventory (2mKGI) for the Karakoram range, integrating multi-source satellite imagery and a deep learning approach (U-Net+CBAM). The methodology is innovative, and the resulting dataset holds significant value for cryospheric research. The manuscript is well-structured; however, there are several areas where additional methodological clarification, transparency regarding data processing, and adherence to established glaciological standards would significantly improve the clarity, reproducibility, and utility of the dataset. I have compiled the following specific comments to assist in strengthening the manuscript for publication.
1. Figure 3 provides a clear overview of the U-Net+CBAM workflow. However, the specific activation functions (e.g., ReLU, Sigmoid) used in the convolutional blocks are not explicitly stated. Including these details in the figure caption or the main text would help the reader better understand the model's non-linear fitting capabilities.
2. As the glacier boundary extraction relies heavily on deep learning, I strongly recommend providing the core training and inference code (or pseudocode) via a public repository such as GitHub. This transparency is crucial for the reproducibility of the methodology.
3. The manuscript mentions that Land Surface Temperature (LST) was retrieved from Landsat 8 imagery. Given that the ZY-3 imagery was acquired between July and October 2020–2021, could you clarify whether the LST represents a single scene matching the specific ZY-3 acquisition date, or if it is a composite/average product for the ablation season (July–October)?
4. In the manual refinement phase, the manuscript mentions using features such as lateral moraines and meltwater outlets as geomorphological indicators. However, the distinction between debris-covered ice and lateral moraines is not explicitly detailed. Please clarify the specific visual texture features or criteria used to manually differentiate debris-covered glaciers from lateral moraines.
5. The legend for Figure 6 includes ‘Karakorum’, but there are no corresponding boundaries, nor are any required here. Could this be an error in the legend?
6. The manuscript provides a total glacier area of 21,261.8 km². According to standard glacier inventory protocols (e.g., RGI standards), internal bedrock outcrops (Nunataks) should typically be excluded. Please confirm whether the 2mKGI dataset excludes these Nunataks. If they are retained, what is their approximate proportion of the total area?
7. Regarding Section 4.4, which identifies 192 surge-type glaciers, please clarify the data source for these "known" surge-type glaciers. Were they referenced from existing literature, or were they identified independently based on surface morphology observed in the ZY-3 imagery?
8. The study notes that 21.5% of the glacier boundaries were supplemented by 10m-resolution Sentinel-2 imagery where ZY-3 coverage was unavailable. Could you provide a breakdown of whether the overall extraction uncertainty of ±4.7% differs significantly between the ZY-3 (2m) and Sentinel-2 (10m) data sources?
9. The 2mKGI dataset provides high-resolution benchmarks for approximately 13,900 glaciers. As this paper is intended to facilitate data sharing, I recommend stating in the text or supplementary material whether the attribute table of the final vector Shapefile includes standard GLIMS or RGI fields (e.g., standard glacier ID format, specific acquisition dates, area). This would greatly enhance the usability and interoperability of the dataset for the global glaciological community.
10. Glacier unit subdivision uses ZY-3 4 m DEM where available and ASTER GDEM V3 elsewhere. The two DEMs differ in resolution, vertical accuracy, and ridge-line definition, which will introduce systematic inconsistencies in glacier unit boundaries across the mosaic edge. Please assess ridge-line positional differences between the two DEMs in overlapping areas, and report how many glacier units straddle the DEM mosaic boundary. Also discuss whether the use of ASTER GDEM introduces detectable biases in glacier count, elevation statistics, or slope metrics for the affected subregions.
11. The inventory includes glaciers down to 0.005 km² (~1250 pixels at 2 m resolution). No justification is given for this low threshold, and there is no discussion of whether very small features may be perennial snow patches.