Enhanced Ship/Iceberg Classification in SAR Images Using Feature Extraction and the Fusion of Machine Learning Algorithms

Abstract

Drifting icebergs present significant navigational and operational risks in remote offshore regions, particularly along the East Coast of Canada. In such areas with harsh weather conditions, traditional methods of monitoring and assessing iceberg-related hazards, such as aerial reconnaissance and shore-based support, are often unfeasible. As a result, satellite-based monitoring using Synthetic Aperture Radar (SAR) imagery emerges as a practical solution for timely and remote iceberg classifications. We utilize the C-CORE/Statoil dataset, a labeled dataset containing both ship and iceberg instances. This dataset is derived from dual-polarized Sentinel-1. Our methodology combines state-of-the-art deep learning techniques with comprehensive feature selection. These features are coupled with machine learning algorithms (neural network, LightGBM, and CatBoost) to achieve accurate and efficient classification results. By utilizing quantitative features, we capture subtle patterns that enhance the model’s discriminative capabilities. Through extensive experiments on the provided dataset, our approach achieves a remarkable accuracy of 95.4% and a log loss of 0.11 in distinguishing icebergs from ships in SAR images. The introduction of additional ship images from another dataset can further enhance both accuracy and log loss results to 96.1% and 0.09, respectively.

Research on Input Schemes for Polarimetric SAR Classification Using Deep Learning

“…Nie et al [12] utilized 12 polarimetric features from Freeman-Durden decomposition, Van Zyl decomposition [44], and Cloude-Potier decomposition, applying an enhanced learning framework for PolSAR image classification. Jafari et al [45] used VGG16, ResNet-50, and ConvNeXt networks to fuse the features extracted from SAR images, as well as the statistical and spatial features and incident angles, to classify ships and sea ice in the images. However, the features used in CNNs do not have clear physical meanings, meaning that they do not have physical interpretability.…”