Use of Vision Transformer to Classify Sea Surface Phenomena in SAR Imagery

This is the authors' abstract. We don't add key points for this paper.

The rapid advancement of satellite technology has led to a substantial increase in the volume of remote sensing data, particularly Synthetic Aperture Radar (SAR) imagery, demanding efficient processing and analysis solutions. This study pioneers the application of Vision Transformers (ViTs) in classifying geophysical phenomena within SAR images, achieving higher recall (0.985), precision (0.984), and F1 scores (0.986) than existing methods and outperforming Convolutional Neural Networks across multiple geophysical categories. We leveraged the single-polarization TenGeoP-SARwv dataset, comprising over 37,000 SAR vignettes, to train and evaluate the ViT models on a balanced subset, improving model generalizability by jointly using WV1 and WV2 images with advanced data augmentation techniques. Additionally, our study is the first to apply a pre-trained ViT model to a dataset with different polarizations and spatial resolutions-the AI4Arctic Sea Ice Challenge dataset-to rigorously assess model adaptability. Without extensive preprocessing, the ViT model achieved approximately 80% accuracy on HH-polarized images and 47% on HV-polarized images, underscoring the effect of polarization differences on classification performance. Notably, this study introduces the use of the attention mechanism in ViTs to elucidate model decision-making, providing interpretability by highlighting regions that influence predictions and revealing why the model succeeds or fails in specific cases. This comparative analysis also underscores the limitations of CNNs, which perform well on texture-based features but struggle with structural classifications, such as rain cells and oceanic fronts, where ViTs excel. These findings advance the application of deep learning in remote sensing, highlighting the robust adaptability of ViTs for diverse SAR imagery.