Data-attention-YOLO (DAY): A comprehensive framework for mesoscale eddy identification. (November 2022)

• Record Type: Journal Article
• Authors: Wang, Xinning; Wang, Xuegong; Li, Chong; Zhao, Yuben; Ren, Peng
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Mesoscale eddy identification -- Attention mechanism -- Data-attention-based YOLO -- One-stage detection; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108870
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a novel framework, namely DAY, including data integration and dynamic attention modules to precisely identify mesoscale eddies. We devise an eddy detection module based on one-stage detecting mechanism to deeply characterize eddy features. The performance of our DAY framework is analyzed in comparison to other deep learning based methods with 17.3% mAP improvement. Abstract: The accurate mesoscale eddy identification methods with deep learning framework depend on either single eddy characteristic from altimeter missions or multi-step eddy examination strategies, disregarding those indistinguishable features from multiple eddy data integration. In this article, we first propose a data-attention-based YOLO (DAY) to precisely recognize mesoscale eddies in the South China Sea (SCS), which can hierarchically unite multiple eddy attributes and efficiently predict eddies with one-step strategy involving detection and classification. It consists of two main components: heterogeneous eddy data integration module and dynamic attention detecting module for eddy identification. The data integration component empirically transforms the field of multi-source eddy data and propagates eddy labels through automatic labeling method, which sustains a good supply for our dynamic attention-base detecting network. To thoroughly identify mesoscale eddies based on spatio-temporal patterns, DAY efficiently learns the characteristics of mesoscale eddies with an improved one-step identification YOLO network. The comparative evaluation results demonstrate that DAY achieves 54% performance improvement over the state-of-the-art methods on single gray SLA data and outperforms two-stage detecting technique Faster R-CNN by 51%.