Background: Ship detection using satellite imagery is critical for maritime surveillance, security, and environmental monitoring. However, existing approaches frequently encounter issues such as small item detection, occlusions, and false positives in complicated maritime situations.

Objectives: This thesis seeks to create an improved ship recognition algorithm by combining the strengths of You Only Look Once version 8 (YOLOv8) and Faster Region-based Convolutional Neural Network (Faster R-CNN) to achieve high accuracy in recognizing ships in satellite images. It also aims to improve the model’s performance through hyperparameter tuning and compares it to previous methodologies.

Methods: We created a revolutionary deep learning system that combines YOLOv8 and Faster R-CNN. The model was trained and validated using the High-Resolution Satellite Image Dataset (HRSID). The learning rate, batch size, momentum, and optimizer selection were all optimized using hyperparameter tuning. The model’s performance was compared against individual YOLOv8 and Faster R-CNN models, as well as a previously published YOLOv7 + Graph Neural Network (GNN) method.

Results: The combined YOLOv8 + Faster R-CNN model outperformed with 98.2% accuracy, 100% recall, and 87.00% precision. It achieved an Intersection over Union(IoU) of 84.95%, beating both individual models and the YOLOv7 + GNN method. The model performed particularly well at spotting small ships and handling occlusions in congested harbor scenarios.

Conclusions: The combination of YOLOv8 and Faster R-CNN, together with adjusted hyperparameters, yields a reliable solution for detecting ships in satellite imagery. The model effectively tackles critical issues in marine object recognition while retaining real-time processing capabilities. This method shows promise for improving maritime surveillance systems and may lead to increased safety and security in marine areas.