ADSRP: A DBSCAN-SNN Framework for AIS-Based Ship Route Planning Using Spatiotemporal Feature Fusion

Abstract

Addressing the issues of noise interference, inadequate modeling of nonlinear characteristics, and computational inefficiency in ship trajectory planning, this study introduces a multi-stage joint optimization model. The model is built upon Automatic Identification System (AIS) data cleaning, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and a Siamese Neural Network (SNN). The operation of the AIS Data and DBSCAN-Based Ship Route Planning Model (ADSRP) involves the following steps: First, AIS data is cleansed by employing a dynamic neighborhood radius and linear interpolation with a sliding window. Key steering points are then extracted by integrating the Douglas-Peucker (DP) algorithm, resulting in an 85.4% reduction in trajectory redundancy. Subsequently, DBSCAN is utilized for density-based clustering of trajectory endpoints, achieving a 93.6% filtering accuracy for noise points. Finally, a symmetric-weight SNN architecture (comprising a 4-layer Transformer encoder and multi-head attention) is designed to filter high-density routes based on cosine similarity.Experimental results demonstrate that, in comparison to the traditional genetic algorithm-based Whole Process Route Planning (WPRP), ADSRP enhances trajectory fitting in the simulation environment by 21% (with an average cosine similarity of 0.86 for ADSRP and 0.71 for WPRP) and shortens the planning time by 67.8% (8.11s for ADSRP and 25.24s for WPRP). In real-world port scenarios, ADSRP reduces voyage deviation by 36.8% (0.98nmi for ADSRP and 1.55nmi for WPRP), cuts fuel consumption by 20.8% (362.58L for ADSRP and 457.89L for WPRP), and optimizes memory usage to 27.5% (compared to the benchmark's 42.5%). Parameter sensitivity analysis verifies the significant impact of key parameters on clustering fragmentation and port identification accuracy (F1-score difference of 22%). The model is co-optimized by data-driven clustering and deep metric learning, providing a high-accuracy, low-energy solution for dynamic path planning in complex sea areas and supporting edge device deployment.