A novel Otsu-UNet coupling approach to extract cage aquaculture in changing bay environments

Abstract

The accurate extraction and mapping of cage aquaculture are of substantial significance to the sustainable development of bay ecosystems but remain challenging due to diverse aquaculture practices in changing bay environments. Existing methods rely primarily on complex classifiers; however, their generalization capabilities are limited. To address this challenge, we developed a novel classification method called Otsu-UNet by integrating the Otsu algorithm with the U-shaped Network method. Based on the Google Earth Engine platform, cage aquaculture data were autonomously extracted from Sentinel-2 imagery using the Otsu technique, and high-quality samples were curated to train the UNet model for cage aquaculture extraction across bay locales. A comparative analysis was further performed with random forest and object-oriented methods. The model-driven extraction attained an F1 score of > 97 % across three distinct study areas—Pulandian, Shidao, and Dongshan Bays in Liaoning, Shandong, and Fujian Provinces—with accuracy metrics of > 95 %. Its accuracy was also superior to that of the random forest method and the object-oriented method. This methodology surpasses conventional approaches in its simplicity, efficiency, precision, and adaptability in deciphering cage aquaculture in the intricate tapestry of bay environments.