| 10 Aug 2026
Deep-learning-based stage-aware classification of Arctic melt ponds and their microtopographic associations from high-resolution UAV imagery
Abstract. Melt ponds are a key component of the summer Arctic sea-ice surface because their formation and evolution strongly affect surface albedo, energy absorption, meltwater redistribution, and sea-ice mass balance. Most previous remote-sensing studies have treated melt ponds as a single surface class, limiting the characterization of heterogeneous pond states during late-summer melt and refreezing. In this study, we developed MP-Unet, a stage-aware semantic segmentation framework for identifying Open, Transitional, and Frozen Melt Ponds from high-resolution unmanned aerial vehicle imagery acquired during the 14th Chinese National Arctic Research Expedition. MP-Unet integrates residual blocks with channel attention, atrous spatial pyramid pooling, attention-gated skip connections, and an auxiliary binary segmentation head. The full model achieved an F1-score of 0.9440 and a mean intersection over union of 0.7466, with class-specific IoU values of 0.5897, 0.7544, and 0.6539 for Open, Transitional, and Frozen Melt Ponds, respectively. Stage-resolved mapping revealed marked spatial heterogeneity among the five observation sites, while Transitional Melt Ponds accounted for approximately 80.3 % of the total pond area in the pooled sample. Pond area–frequency distributions showed a general scale-dependent decline and a sparse large-area tail, although the strength of the fitted scaling relationship varied among sites. Object-level analysis further showed that Frozen Melt Ponds generally had more compact and regular shapes, whereas Open Melt Ponds exhibited broader circularity distributions extending toward lower values. DEM-assisted analysis indicated significant stage-dependent differences in local relative elevation: Transitional Melt Ponds occupied lower local topographic positions than Frozen Melt Ponds, despite the absence of significant differences in distance to the nearest ridge-like feature. These findings demonstrate that stage-aware classification provides information beyond conventional binary melt pond mapping by linking surface-state identification with pond morphology and local microtopographic position. The proposed framework offers a practical basis for fine-scale observations of Arctic sea-ice surface evolution and for the validation and improvement of satellite and numerical melt pond products.
The manuscript presents an interesting high-resolution UAV dataset and a potentially useful machine-learning approach for melt-pond classification. However, I have serious concerns about the physical basis and interpretation of the proposed Open, Transitional, and Frozen Melt Pond classes.
My main concern is that the study proceeds mainly from visually defined RGB (but actually optical) classes to machine-learning classification, physical interpretation to fit the results, rather than establishing the physical states first and then testing whether they can be identified from UAV imagery. Consequently, the manuscript demonstrates that the proposed visual classes can be distinguished, yet does not show that they represent distinct developmental stages of melt ponds.
The manuscript also lacks a sufficiently clear scientific thread connecting machine-learning development to real optical classification, pond morphology, DEM analysis etc... I think the manuscript requires substantial restructuring and clarification of whether its main contribution is a new image-classification method or a physical investigation of melt-pond evolution.
Please see further comments in the supplements.