POSTER SESSION 4
10:30–12:30, Thursday, September 17
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
ABSTRACT 1387 | POSTER 68
A DEEP LEARNING FRAMEWORK FOR OCEAN OIL SLICK SEGMENTATION USING SENTINEL-1 IMAGERY
Ocean oil slicks pose severe threats to marine ecosystems. Although Synthetic Aperture Radar (SAR) remote sensing provides global-scale and high spatiotemporal resolution data, the wide distribution, random occurrence, variable morphology, and relatively short persistence of oil slicks—coupled with interference from lookalikes—present great challenges for accurate oil slick detection. Deep learning techniques have demonstrated substantial potential for oil slick segmentation in SAR imagery; however, their performance depends heavily on the quantity and quality of training data. This study aims to address two critical issues: (1) constructing a globally applicable oil slick segmentation method based on deep learning, and (2) validating its performance in real-world applications. Based on a previously constructed global oil slick inventory, this study established a dataset (GlobalOSD-SAR) comprising 100,329 oil slick images and 100,134 lookalike images using Sentinel-1 imagery from 2014 to 2020. We proposed an iterative optimization algorithm that refines the oil slick segmentation model through continuous improvement and supplementation of training samples. Our results indicated that enhancing dataset diversity can effectively improve model generalization capability. The optimal model achieved mIoU scores of 95.99% and 82.63% on two independent test sets representing real-world scenarios. Future work will focus on expanding model parameters, optimizing and augmenting the dataset, and integrating SAR image textural features with auxiliary data to further enhance model robustness, generalization, and transferability.
Yanzhu Dong, East China Normal University, [email protected], https://orcid.org/0000-0001-5738-3768
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
