A Loss-Modified Densely Connected Atrous Convolutional U-Shape Encoder-Decoder Style Network for Ship Segmentation via Satellite Images

The segmentation of ships in satellite imagery is vital for maritime surveillance but presents unique challenges such as imbalanced datasets, small-scale object detection, and complex backgrounds. To address these, we propose an advanced U-Net-based architecture enhanced with depthwise separable convolutions for efficient feature extraction, dropout regularization, and batch normalization for improved generalization. Furthermore, a custom combined loss function integrating Dice Loss, Binary Cross-Entropy (BCE), and Focal Loss is introduced to tackle class imbalance and improve segmentation precision. The proposed model achieves superior performance across multiple metrics, including a Dice Coefficient of 0.6325 and a precision of 0.6678, outperforming baseline models such as standard U-Net, Res-U-Net, Dense-U-Net, and MultiRes-U-Net. Qualitative results further demonstrate the models ability to detect small and complex ship structures, although limitations in noisy environments and misclassification of non-ship objects are observed. Our findings underscore the importance of architectural enhancements and loss optimization in segmentation tasks. Future work could focus on incorporating attention mechanisms, advanced denoising techniques, and multimodal data integration to further improve accuracy in challenging conditions. This study highlights the potential of the enhanced U-Net for maritime applications and broader segmentation tasks in satellite imagery analysis, offering a robust solution for global maritime monitoring.