Transformer-Based Deep Learning for Mesoscale Eddy Detection in Sea Surface Temperature Maps
Abstract
Mesoscale eddies are dynamic oceanic phenomena significantly influencing marine ecosystems' energy transfer, nutrients, and biogeochemical cycles. These eddies' precise identification and categorization can improve climate modeling, ocean circulation research, and environmental surveillance. This study presents an innovative methodology for mesoscale eddy detection utilizing Transformer-based deep learning models, namely Swin Transformer U-Net (Swin-Unet) and SegFormer, to categorize ocean eddies from Sea Surface Temperature (SST) maps sourced from the Copernicus Marine Environment Monitoring Service (CMEMS). In contrast to traditional convolutional neural networks (CNNs) that have prevailed in the domain, Transformer-based models provide superior global attention mechanisms, facilitating greater feature extraction and segmentation precision. The models are trained on labeled SST datasets and assessed using Intersection over Union (IoU), Dice coefficient, precision, recall, and F1-score. Experimental findings demonstrate that Transformer-based designs surpass conventional CNN-based techniques, yielding enhanced generalization and superior accuracy in classifying cyclonic and anticyclonic eddies. This study illustrates the efficacy of attention-based segmentation algorithms for resilient oceanographic applications.