Monitoring upwelling regions in major coastal zones using deep learning and sea surface temperature images

ABSTRACT

The coastal region of northwest Africa experiences a persistent and variable upwelling phenomenon throughout most of the year, contributing to the presence of one of the world’s largest and most productive fishing ports. In this study, we introduce Deep$\mathrm{Coas}{\mathrm{t}}_{\mathit{up}}$-Net a convolutional neural network architecture designed for identifying and extracting upwelling regions from satellite sea surface temperature (SST) images. Our model is trained on a dataset of SST images captured along the Moroccan Atlantic coast, utilizing essential input parameters, including SST, latitudinal position ($\mathit{LA}{T}_{\mathit{pos}}$), and distance from the coastline ($D_{\mathit{coast}}$). By incorporating these physical parameters as input features and training the model with corresponding masks for the Atlantic coast of Morocco, the Deep$\mathrm{Coas}{\mathrm{t}}_{\mathit{up}}$-Net network learns to accurately segment upwelling regions in major coastal zones and demonstrates its generalization capability. We validate our methodology using a quantitative index. The results of this validation showcase the effectiveness of our approach. Subsequently, we explore seasonal and interannual upwelling trends within a 21-year time series, spanning from 2000 to 2020.

KEYWORDS:

• Sea Surface Temperature
• Moroccan coastal upwelling
• coastal zones
• deep learning
• U-Net

Disclosure statement

No potential conflict of interest was reported by the author(s).