Spectral–Texture Feature Fusion With Spatial Group Cross-Validation for Sea Ice Mapping Using HY-1C Satellite Imagery

Abstract

Accurate extraction of sea ice extent is of great importance for climate research, maritime navigation safety, and marine environmental monitoring. Traditional optical remote sensing methods are often limited by insufficient feature representation, while conventional machine-learning evaluations may overestimate classification accuracy due to spatial autocorrelation among samples. To address these issues, this article proposes a sea ice classification framework that integrates spectral and textural information through a novel spectral–texture fusion index (STFI) and adopts a spatial group cross-validation strategy. The STFI nonlinearly combines spectral and texture features to enhance the separability between sea ice and seawater, while spatial group cross-validation reduces the overestimation caused by spatial autocorrelation. A multifeature dataset is constructed, and three machine-learning models—random forest, XGBoost, and LightGBM—are trained and evaluated under this validation scheme. Results show that LightGBM achieves the best performance, with an F1-score, overall accuracy, and Matthews correlation coefficient of 92.99%, 89.73%, and 70.40%, respectively. Feature separability analysis, SHapley Additive exPlanations interpretation, and ablation experiments consistently confirm STFI as the most influential feature, and using only the spectral index, texture feature, and STFI already yields excellent classification. Comparisons with threshold segmentation, maximum likelihood estimation, support vector machine, and U-Net further verify the superiority of the proposed method. Cross-sensor experiments demonstrate high consistency when the model is directly transferred to MODIS imagery, and cross-regional validation in the Tatar Strait and the Yellow Sea shows good generalization capability. Moreover, a daily sea ice extent time series for Liaodong Bay during the 2024–2025 winter is generated from HY-1C and MODIS data, exhibiting good agreement with operational ice charts from the Liaoning Maritime Safety Administration (R2 = 0.96, r = 0.98, p < 0.001). The proposed framework offers a reliable solution for operational sea ice monitoring.