Mapping Arctic Polynyas from Sea Ice Concentration Data Using Hierarchical Partitioning and Differentiated Region Growing: Sensitivity to Threshold and Product Choices

The hierarchical classifier is well-motivated and described with enough detail to be reproducible in principle: concentration defines candidate pixels, then connectivity and neighborhood context assign them to polynya or ocean-connected classes. The 63-scene stratified visual assessment provides credible support, yielding "97.32%/95.94%" user/producer accuracy for polynyas and a robust mean $\kappa = 0.93$. The main sensitivity conclusion—that "the absolute magnitude of SIC-defined polynya extent is strongly configuration-dependent, whereas major temporal features are comparatively more stable"—is richly illustrated with pan-Arctic and regional curves, and its geographic and seasonal structure (strongest in the Central Arctic Ocean, weakest in the Gulf of St. Lawrence) is physically plausible. The authors also responsibly anchor their work in prior SIC intercomparisons and Antarctic threshold studies.

Several sources of uncertainty in the operational definition are frozen but untested. The authors note that "Alternative neighborhood configurations and geographic exclusion rules were explored during method development... Their individual effects on mapping accuracy and extent have not been systematically quantified" (Section 5.3), leaving the 15% seed threshold, the 150,000 km$^2$ geographic-exception criterion, and the 5$\times$5 open-water exclusion as unquantified degrees of freedom. The product experiment also bundles resolution with algorithm: "the products differ in retrieval construction as well as grid spacing" (Section 4.2), so the 68.19% MAPE between 6.25 km and 25 km cannot be decomposed into a pure resolution effect versus a retrieval bias. The 2011/12 season is entirely excluded rather than bridged with alternative data, introducing a discontinuity in the 22-season record. Finally, the reassuring claim that interannual structure "remains recognizable" is supported by overlapping curves, not by reported cross-configuration correlations or variance decompositions.

The core evidence for magnitude sensitivity is strong and cleanly quantified: relative to 75%, aggregate differences are "13.48%, 32.47%, and 65.04% at 80%, 85%, and 90%" (Section 4.1), and at 80% threshold the "6.25–25 km comparison" yields a MAPE of "68.19%" (Section 4.2). Regional and seasonal stratification add credibility. Comparisons to related work are generally fair, citing Landsat-based SIC intercomparisons (Kern et al., 2022) and recent Antarctic threshold-sensitivity studies (Landrum et al., 2026; Hurst & Boehme, 2026). What is missing is a head-to-head quantitative benchmark against contemporary deep-learning Arctic polynya methods, and the temporal-stability claim would be more convincing with reported cross-configuration correlation coefficients or shared-variance metrics rather than visual overlap alone.

The operational definition is unusually explicit: equations for gap filling ($C_{fill}$), daily extent ($E(t) = \alpha N_t$), seasonal integrals ($S_s$), and MAPE are all stated in Section 3.4, and the geographic exclusion logic is described in Section 3.1. This traceability is a genuine strength. However, no code, parameter files, or exact mask vectors are shared as supplementary material, and the manually delineated coastal masks are described only qualitatively. The short-gap interpolation "does not add independent information about opening or closing events during the missing dates" (Section 3.1), yet the specific interpolated scenes are not listed. Hyperparameters such as the 150,000 km$^2$ exception threshold are stated without derivation. An independent team could reproduce the study from the prose, but it would require significant manual reconstruction of masks and rules.