Entropy guided spectral phenological feature engineering using Sentinel-2A data for Pan-Arctic vegetation classification

Arctic vegetation plays a critical role in regulating terrestrial ecosystem structure and function. Existing remote-sensing classification products are often limited in thematic detail and accuracy because of the region’s inaccessibility and the underuse of high-dimensional spectral and phenological information. This study established a Pan-Arctic vegetation classification system comprising Barren Tundra (BT), Sparse Tundra (ST), Cyperaceae Tundra (CT), Graminoid Tundra (GT), Wet Tundra (WT), Grass, and Shrub. A total of 161 homogeneous samples were selected across the Pan-Arctic domain through multi-source consistency filtering and visual verification. These samples were used for spectral–phenological feature development, entropy-based feature evaluation, threshold definition, and model construction. We introduced spectral characteristic information entropy (SCIE) and spectral characteristic information weight coefficients (SCIWC) to characterize spectral information content and relative feature importance. The entropy-guided feature set improved inter-class separability; for example, green-band information helped distinguish BT from ST, whereas red-band information contributed to the separation of CT and GT. Final map accuracy was assessed independently using 761 reference points in Yamal’skiy Rayon, which were not used for feature development or model construction. The classification achieved an overall accuracy of 90.67% and a Kappa coefficient of 0.88. A comparison with CALC-2020 indicated broad spatial consistency, while differences reflected the finer thematic detail and differing class definitions of the two products. These results demonstrate the potential of entropy-guided spectral–phenological feature engineering for fine-scale Arctic vegetation mapping, while further multi-region and multi-year validation is needed to assess Pan-Arctic transferability.