A method for generating obstacle sample datasets in coastal remote sensing scenes based on out-of-distribution detection

The navigation and docking of coastal vessels are significantly affected by various maritime obstacles, including natural obstacles like reefs and artificial ones such as breakwaters. Obstacle detection through remote sensing images can form a navigation map, which plays an important role in navigation safety and environmental assessment. Constructing deep learning models is an effective approach for tasks like obstacle detection, but these models heavily rely on labeled sample data. However, sample data for coastal maritime obstacles are scarce, making it critical to develop a method for generating such datasets. Out-of-distribution (OOD) detection methods can enhance a model’s adaptability to unknown scenarios, and previous studies have shown that post hoc analysis-based OOD detection methods are suitable for remote sensing scenarios. This paper proposes a method for generating maritime obstacle datasets based on post hoc analysis-based OOD detection. An experimental analysis was conducted in the coastal waters around the main island of Taiwan, China, to expand the obstacle sample dataset. This paper evaluates the performance of the OOD detection method by the area under the receiver operating characteristic curve (AUROC). Results show that this method achieved an AUROC value of 80.33% on the test dataset, demonstrating a high capacity for OOD detection. Building on this, an OOD detection was carried out in the coastal waters of the study area, resulting in 8,161 OOD scene images. These images were used to construct a coastal maritime obstacle remote sensing sample dataset comprising 2,942 images across six categories: reefs, sandbanks, breakwaters, fish net cages, fishing lines, and coastal bridges. This demonstrates the effectiveness of the proposed method in expanding the OOD scene sample dataset, effectively filling the gap in maritime obstacle sample datasets and laying a data foundation for future maritime obstacle recognition research.