RIVDCB-UNet: A Robust Vessel Detection in Cluttered Backgrounds via UNet for SDGSAT-1 Thermal-Infrared Data

Aiming at complex and variable scenarios with cloud and fog obscuration and diverse interference sources, a robust infrared vessel detection in cluttered backgrounds via the UNet (RIVDCB-UNet) is designed. This model effectively overcomes the limitations of traditional UNet in small target detection and significantly improves its detection performance. The RIVDCB-UNet model takes a single thermal infrared image as input and incorporates the BCE-Dice loss function to ensure stability and accuracy during the training process. The model integrates the EfficientNet backbone network for efficient feature extraction, followed by feature upsampling and fusion through a U-shaped decoder. Finally, an eight-connected neighborhood clustering module is introduced to further refine pixel-level classification results and achieve precise localization. The training of the proposed model leverages the publicly available NUDT-SIRST-Sea Dataset and our self-constructed SDGSAT-1 thermal infrared vessel detection dataset. Data augmentation techniques, including rotation and cropping, are employed to expand the dataset, thereby addressing the issue of target-background imbalance in the original datasets. This imbalance can cause the model to overly focus on information-sparse background regions. The proposed model achieves detection rates, recall rates, and F1 scores of 95.39%, 97.89%, 96.62% on the NUDT-SIRST-Sea Dataset, and 97.81%, 96.78%, 97.29% on the SDGSAT-1 dataset, respectively. These results are significantly higher than those of other models such as ACM (84.81%, 73.17%, 78.56% on NUDT-SIRST-Sea and 90.17%, 72.87%, 80.60% on SDGSAT-1) and DNANet (89.98%, 83.10%, 86.40% on NUDT-SIRST-Sea and 95.64%, 88.21%, 91.77% on SDGSAT-1).