SVIADF: Small Vessel Identification and Anomaly Detection Based on Wide-Area Remote Sensing Imagery and AIS Data Fusion

Small target ship detection and anomaly analysis play a pivotal role in ocean remote sensing technologies, offering critical capabilities for maritime surveillance, enhancing maritime safety, and improving traffic management. However, existing methodologies in the field of detection are predominantly based on deep learning models with complex network architectures, which may fail to accurately detect smaller targets. In the classification domain, most studies focus on synthetic aperture radar (SAR) images combined with Automatic Identification System (AIS) data, but these approaches have significant limitations: first, they often overlook further analysis of anomalies arising from mismatched data; second, there is a lack of research on small target ship classification using wide-area optical remote sensing imagery. In this paper, we develop SVIADF, a multi-source information fusion framework for small vessel identification and anomaly detection. The framework consists of two main steps: detection and classification. To address challenges in the detection domain, we introduce the YOLOv8x-CA-CFAR framework. In this approach, YOLOv8x is first utilized to detect suspicious objects and generate image patches, which are then subjected to secondary analysis using CA-CFAR. Experimental results demonstrate that this method achieves improvements in Recall and F1-score by 2.9% and 1.13%, respectively, compared to using YOLOv8x alone. By integrating structural and pixel-based approaches, this method effectively mitigates the limitations of traditional deep learning techniques in small target detection, providing more practical and reliable support for real-time maritime monitoring and situational assessment. In the classification domain, this study addresses two critical challenges. First, it investigates and resolves anomalies arising from mismatched data. Second, it introduces an unsupervised domain adaptation model, Multi-CDT, for heterogeneous multi-source data. This model effectively transfers knowledge from SAR–AIS data to optical remote sensing imagery, thereby enabling the development of a small target ship classification model tailored for optical imagery. Experimental results reveal that, compared to the CDTrans method, Multi-CDT not only retains a broader range of classification categories but also improves target domain accuracy by 0.32%. The model extracts more discriminative and robust features, making it well suited for complex and dynamic real-world scenarios. This study offers a novel perspective for future research on domain adaptation and its application in maritime scenarios.