Stripe segmentation of oceanic internal waves in SAR images based on SegFormer

Abstract

The study of oceanic internal waves remains a critical area of research within oceanography. With the rapid advancements in oceanic remote sensing and deep learning, it is now possible to extract valuable insights from vast datasets. In this context, by building datasets using deep learning models, we propose a novel stripe segmentation algorithm for oceanic internal waves, leveraging synthetic aperture radar (SAR) images based on the SegFormer architecture. Initially, a hierarchical transformer encoder transforms the image into multilevel feature maps. Subsequently, information from various layers is aggregated through a multilayer perceptron (MLP) decoder, effectively merging local and global contexts. Finally, a layer of MLP is utilized to facilitate the segmentation of oceanic internal waves. Comparative experimental results demonstrated that SegFormer outperformed other models, including U-Net, Fast-SCNN (Fast Segmentation Convolutional Neural Network), ORCNet (Ocular Region Context Network), and PSPNet (Pyramid Scene Parsing Network), efficiently and accurately segmenting marine internal wave stripes in SAR images. In addition, we discuss the results of oceanic internal wave detection under varying settings, further underscoring the effectiveness of the algorithm.