A Stacking Approach for Arctic Sea Ice Lead Classification (SALC) via Sentinel-1 SAR Imagery

Leads are linear fractures formed by the deformation of the ice cover and are crucial areas for Arctic heat exchange between the ocean and the atmosphere. Accurately identifying leads is essential for studying climate change, polar ecosystems, and shipping navigation. Recent research on leads detection methods developed using synthetic aperture radar (SAR) tend to ignore the developmental stage of the leads and the distribution context of leads [first-year ice (FYI) or multiyear ice (MYI)], which hinders the generalizability of these methods. In this article, we propose a stacking approach for lead classification (SALC) integrating five specialized learners for five typical lead conditions. Ablation experiments reveal the effectiveness of each component of SALC: 1) the optimal preprocessing method is to input the incidence angle as a feature and apply the advanced thermal noise removal. In addition, the effect of inputting the incidence angle is significantly stronger than the incidence angle correction for original images; 2) the SALC performs best when the data features include polarization, texture, and upscaling features; 3) XGBoost is the most suitable meta classifier as compared to support vector machine (SVM) and random forest (RF). The results show that SALC outperforms both single learners and traditional classifiers across various conditions and quality criteria. In addition, SALC can tolerate at least 15% of erroneous samples.