Comparative Analysis of Image Segmentation Methods for Unmanned Surface Vehicles under Varying Illumination Conditions

This paper presents a comprehensive evaluation of eight semantic segmentation models-CGNet, ENet, ERFNet, FSSNet, LinkNet, SegNet, SQNet, and UNet-under varying natural illumination conditions. The models are trained on an augmented dataset designed to simulate diverse environmental scenarios using random zoom and rotation. Performance is assessed through segmentation accuracy, training loss, and robustness across five illumination types: sunshine, sunny, overcast, very dark, and twilight. The results indicate that UNet demonstrates the highest mIoU of 0.9265 under overcast conditions, with consistent performance across all lighting scenarios. CGNet achieves the lowest training loss of 0.0831, also under overcast, and shows strong generalization, especially during twilight and very dark conditions. LinkNet also performs competitively with a high mIoU of 0.9307, highlighting its balance between accuracy and efficiency. In contrast, models like FSSNet and SegNet are more sensitive to low-light conditions, with mIoU values dropping to 0.8553 and 0.8662 respectively in very dark environments. These findings suggest that UNet is best suited for applications requiring high segmentation accuracy, whereas ENet and SQNet are optimal for real-time deployments. Despite not being evaluated in realtime deployments, the offline results suggest their applicability in practical scenarios. The models' segmentation consistency and boundary clarity are visualized in various lighting conditions, demonstrating the impact of environmental variation on semantic segmentation tasks.