Deformable DETR for Submarine Cable Burial-Depth Estimation from Shallow Strata Profile Images

The engineering parameters for the burial depth of the submarine cable need to be determined for route maintenance, exposure-risk assessment and post-lay verification; however, when the cable response is weak or partially covered by sediment reflectors, these cannot be accurately obtained from shallow strata profile images. A Deformable DETR-based model is proposed in this paper to locate cable signatures and determine burial depth directly from calibrated sub-bottom profile images. A multi-scale convolutional encoder with deformable transformer attention, a cable-query decoder, and a depth-aware regression head that converts image-space offsets into metric burial depth via survey geometry is used. A dataset of 4,860 profile frames from muddy sand, silty clay and mixed shallow deposits was divided into line-disjoint training, validation and test sets. Compared with Faster R-CNN, YOLOv8 and a CNN-LSTM depth regressor, the proposed model achieved a detection F1-score of 92.6%, a mean absolute burial-depth error of 5.8 cm, and a 95th-percentile absolute error of 14.7 cm. Performance remained stable for low-contrast cable echoes and burial depths between 0.35 m and 2.10 m, and the route-level shallow-burial warnings closely matched manual interpretation. Based on the results, deformable attention can focus the sampling on cable-scale responses and near sediment interfaces to provide a feasible route for repeatable image-based cable burial assessment.