Akademik Veri Yönetim Sistemi
International Journal of Computational Intelligence Systems, cilt.19, sa.1, 2026 (SCI-Expanded, Scopus)
This study focuses on the detection and classification of ships in satellite images under adverse weather conditions (rain, snow, and fog). To mitigate the negative impacts of weather conditions, the Two Stage Knowledge Learning and Multi-stage Progressive Refinement Network models were applied separately, and their effects on ship detection were compared. It was observed that reducing the impact of bad weather resulted in an approximate 8% increase in the mAP value during the ship detection phase. Extremely small ships, appearing tiny due to the satellite’s viewing distance, were successfully identified. The utilization of the Feature Pyramid Network for positioning small ships, combined with YOLOv8’s center point approach to address overlapping situations, seems crucial. To prevent the misclassification of very small ships as land masses or small islets, a new dataset was created. This dataset was used for training an enhanced variational autoencoder for eliminating false negative samples. This dataset also facilitated the elimination of potential land masses that could be erroneously identified as ships. In this study, the Detection, Localization, Recognition, and Identification phases were designed for independent optimization. The proposed model incorporates the Pyramid Residual Attention Inception Blocks architecture for the detection, classification, and identification phases, while YOLOv8 is employed for the positioning phase. The F1 score values achieved independently for the detection, localization, recognition and identification phases were found to be 94.3%, 84.0%, 74.1%, 88.2%, and 63.9%, respectively. Moreover, the overall F1 score of the model was determined to be 96.0%, 85.4%, 65.0%, 63.0%, and 55.0%.