Comparative Analysis of the Effectiveness of Petroleum-Product Contamination Detection in Aquatic Systems Using Pre-Trained YOLO-Family Neural Network Models and the SMDNetV4 and PSPNet Architectures

Source abstract

Introduction . The PSPNet, SMD-Net, and YOLO26-Seg semantic segmentation architectures are considered for the task of detecting oil spills in satellite imagery. Particular attention is paid to comparing segmentation accuracy under identical training conditions. Materials and Methods . PSPNet uses a ResNet101 backbone with a pyramid pooling module for semantic segmentation. SMD-Net is a lightweight shared-backbone architecture with two decoders and CTGM, CoordAtt, and ASPP modules. YOLO26-Seg combines a YOLO26 encoder backbone with fixed weights and a lightweight decoder. The following metrics are compared: IoU (Intersection over Union), defined as the ratio of the intersection of the predicted and ground-truth regions to their union; Dice (Dice coefficient), a measure of segment overlap based on the intersection area; accuracy, the proportion of correct predictions among all predictions; inference time, the time required by the model to process one input; and the number of trainable parameters. Results . After 50 epochs of training on the existing dataset masks without manual annotation, PSPNet significantly outperformed SMD-Net in binary oil-spill segmentation (IoU = 0.704, Dice = 0.770, Acc = 0.896 versus IoU = 0.467, Dice = 0.575, Acc = 0.671 for SMD-Net). The additionally fine-tuned YOLO26l-seg model demonstrated moderate pixellevel metrics (IoU = 0.320) at a reduced detection threshold, which limits its applicability to this task. Discussion . The choice of architecture for oil-spill segmentation is governed by a trade-off between accuracy and speed: PSPNet provides the highest segmentation quality, SMD-Net performs worse across all metrics, and YOLO26-SemSeg cannot reliably detect oil spills without task-specific training. Conclusion . PSPNet significantly outperforms SMD-Net and YOLO26-Seg in binary oil-spill segmentation under identical training conditions (IoU = 0.704 versus 0.467 for SMD-Net), confirming the effectiveness of the ResNet101 backbone combined with pyramid pooling.