The increasing demand for high-precision navigation in Unmanned Surface Vehicles (USVs) underscores the critical role of image segmentation in maritime environments. Accurate segmentation of maritime images into distinct categories, including sky, water, and obstacles, is essential for ensuring safe operations, mitigating collision risks, and enhancing autonomous decision-making capabilities. To address the challenges, the proposed study presents an advanced image segmentation methodology tailored for maritime applications. The proposed approach leverages a deep learning-based encoder-decoder architecture built upon the ResNet34 framework, incorporating atrous convolution techniques to facilitate multi-scale feature extraction—an essential requirement for effectively capturing the complexity of maritime environments. The architecture further integrates skip connections and upsampling layers within the decoder to enhance segmentation precision and improve predictive accuracy. The implementation framework includes preprocessing techniques such as image resizing, normalization, and standardization to optimize input quality, while post-processing aligns predictions with pre classification categories to enhance interpretability. The model is trained on the LaRS dataset, which provides a diverse and realistic representation of maritime conditions. Experimental evaluations conducted on real-world datasets demonstrate that the proposed model achieves a Mean Intersection over Union (mIoU) of 95.4%, surpassing the performance of existing segmentation approaches. By addressing the limitations of conventional methodologies, the following research contributes to the development of more robust and reliable USV navigation systems capable of operating effectively in diverse and challenging maritime environments.