Stability prediction of muddy submarine channel slope based on sub-bottom profile acoustic images and transfer learning

Abstract

This research addresses the challenging task of predicting the stability of muddy submarine channel slopes, crucial for ensuring safe port operations. Traditional methods falter due to the submerged nature of these channels, impacting navigation and infrastructure maintenance. The proposed approach integrates sub-bottom profile acoustic images and transfer learning to predict slope stability in Lianyungang Port. The study classifies slope stability into four categories: stable, creep, expansion, and unstable based on oscillation amplitude and sound intensity. Utilizing a sub-bottom profiler, acoustic imagery is collected, which is then enhanced through Gabor filtering. This process generates source data to pre-train Visual Geometry Group (VGG)16 neural network. This research further refines the model using targeted data, achieving a 97.92% prediction accuracy. When benchmarked against other models and methods, including VGG19, Inception-v3, Densenet201, Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM), and an unmodified VGG16, this approach exhibits superior accuracy. This model proves highly effective for real-time analysis of submarine channel slope dynamics, offering a significant advancement in marine safety and operational efficiency.