Deformable Large Kernel Approximation For Semantic Segmentation of Shorelines
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3142214Utgivelsesdato
2024Metadata
Vis full innførselSamlinger
Sammendrag
Marine oil spills have emerged as a major threat to marine ecological safety. Large incidents related to oil extraction and maritime operations has lead to an estimated 1,522,000 tonnes of oil being released into the oceans. The detrimental effects of such incidents can significantly impact the marine ecosystem. To mitigate the short and long-term effects, rapid response is cruicial. However, effective cleanup and preparedness require comprehensive knowledge of the geological characteristics of the affected areas. Acquiring this knowledge is challenging due to the complex and dynamic nature of the vast shorelines.This thesis investigates the effects of using AI to detect and classify shorelines in aerial images. For this purpose, a dataset for semantic segmentation of shorelines is presented along with the challenges it presents for an AI-based solution. To tackle these challenges, we present a convolutional attention-based module, armed witha novel deformable convolution operator which is scalable to large kernels and capable of extracting global information. By inheriting the benefits of convolution and attention and introducing single directional deformable convolution, our proposed solution effectively address the unique challenges of shoreline semantic segmentation.Our method is benchmarked on popular remote sensing datasets including the shoreline dataset presented in this thesis. The results highlights our methods superior performance by outperforming well-established methods for semantic segmentation of shorelines. Moreover, our model achieves performance comparable to the state-of-the-art on the Potsdam dataset, including the Clutter category. To the best of our knowledge, it attains the second-best results across all three performance metrics. Our results showcased the effectiveness of using AI for extracting the important coastal features and present high precision shoreline segmentations.