Abstract
Sea ice mapping on synthetic aperture radar (SAR) imagery is important for various purposes, including ship navigation and usage in environmental and climatological studies. Although a series of deep learning-based models have been proposed for automatic sea ice classification on SAR scenes, most of them are flat N-way classifiers that do not consider the uneven visual separability of different sea ice types. To further improve classification accuracy with limited training samples, a hierarchical deep learning-based pipeline is proposed for sea ice mapping from SAR. First, a semantic segmentation model with encoder-decoder structure is implemented to accurately separate ice and open water on each SAR scene. To classify different ice types, a two-level category hierarchical convolutional neural network (CNN)-based model is then trained using limited numbers of labeled image patches. Experimental results on dual-polarized SAR scenes collected from C-band satellite RADARSAT-2 show that ice-water mapping results are in very good accordance with pixel-based labels under different combinations of encoders and decoders. Also, compared to a flat N-way CNN, the hierarchical CNNs further boosts the classification accuracy among all the ice types.
Presenter
Xinwei Chen, PhD candidate in Systems Design Engineering