Deep residual U-Net for automatic detection of Moroccan coastal upwelling using SST images

Upwelling phenomenon is one of the most important dynamic process in the ocean, which brings nutrients from the depths of the ocean into the surface layer, leading to an enhancement of the primary production and playing a considerable role in the coastal ecosystem. Deep learning (DL) based segmentation methods have been providing state-of-the-art performance in the last few years. These methods have been successfully applied to oceanic remote sensing image segmentation, classification, and detection tasks. In particular, U-Net, has become one of the most popular for these applications. This paper proposes UpwellRes-Net, a deep fully convolutional neural network architecture, for automatic upwelling detection and pixel-segmentation on sea surface temperature (SST) images. The proposed model is based on U-Net structure and residual learning, thus, combining the strengths of both approaches. The main objective of this study is to investigate the performance of deep learning in the extraction of upwelling area. Hence, UpwellRes-Net is trained and optimized on satellite-derived SST database provided by the Moderate Resolution Imaging Spectroradiometer (MODIS). Experiments on the southern Atlantic Moroccan coast show the superiority of the proposed model to a transfer learning based model developed for the same. Deep learning based upwelling detection system can be a cost effective, accurate and convenient way for objective analysis of upwelling phenomenon.