Intelligent identification of oceanic eddies in remote sensing data via Dual-Pyramid UNet

Abstract

Oceanic eddies are an omnipresent phenomenon of seawater flow and critical in transporting oceanic energy and material. Consequently, mastering and comprehending the characteristics of ocean eddies through detecting and recognizing eddies contributes to the understanding of oceanography. In traditional oceanography, a series of methods to identify eddies with physical or geometric characteristics have been developed. Deep learning frameworks have recently been applied in the eddy detection field. In this paper, a Dual-Pyramid UNet architecture that combines a pyramid split attention (PSA) module and atrous spatial pyramid pooling (ASPP) is proposed to identify oceanic eddies from remote sensing data. The encoder and decoder parts can effectively integrate low-level and high-level features, thus ensuring that feature information is not lost in large quantities after the nonlinear connection mode. In addition, the PSA and ASPP modules are introduced into the encoding, decoding, and skip connections to enhance feature extraction. Experiments were implemented in two typical study areas—the North Atlantic and South Atlantic. The recognition results demonstrate that Dual-Pyramid UNet can outperform four other competitive AI-based methods, especially for eddy edges and small-scale eddies.

摘要

海洋涡旋是大洋中重要的组成部分, 对海洋能量和物质的输送至关重要. 海洋涡旋的检测和表征无论是对于海洋气象学, 海洋声学还是海洋生物学等领域都具有重要的研究价值. 本文基于UNet架构, 并结合金字塔分割注意力(PSA)模块和空洞空间卷积池化金字塔(ASPP)构造了Dual-Pyramid UNet模型, 以平面异常和海表面温度数据中进行海洋涡旋的识别. 实验在北大西洋和南大西洋两个涡旋活跃区域进行并选用多个评价指标对识别结果进行评价以证明模型的优异性能.