Object detection in high-resolution satellite imagery (HRSI) is critical for various applications, including maritime surveillance, natural disaster management, and environmental mapping. Accurately detecting ships in HRSI images presents challenges due to the variability in ship appearances, changing lighting conditions, and complex backgrounds. In this paper, we introduce ShipFPN, an innovative architecture that integrates meta-learning and data generation techniques to enhance model performance and robustness, particularly for under-represented ship categories. Our approach utilizes spectral convolutions to leverage the rich spectral information in satellite imagery for improved vessel discrimination. We also revise the formula for calculating the levels of the feature pyramid to better manage objects of varying sizes, especially large ships, and adopt an oriented bounding box generation technique to accurately localize vessels in the images. The effectiveness of ShipFPN is demonstrated through extensive evaluations on the ShipRSImageNet dataset, showcasing significant improvements in detection performance over existing FPN models.