Abstract
Under the unique advantage of integrating spectral and spatial information, hyperspectral imaging technology has shown great application potential in marine target detection. This study aims to systematically evaluate and compare the performance of seven representative anomaly detection algorithms under different water depths (0m,2m,3m,5m) using aerial hyperspectral data. The experimental results show that all algorithms’ performance generally declines with increasing water depth. In shallow water (0m and 2m), Gated Transformer for Hyperspectral Anomaly Detection (KIFD) achieves a detection rate of up to 1.0 at 0m and Low-Rank and Sparse Representation-based Detector (LRASR) also has high detection rates, but both have a False Positive Rate (FPR) higher than 0.45; Reed-Xiaoli Detector (RX) and Collaborative Representation-based Detector (CRD) perform stably with detection rates above 0.9 and controlled FPR. Local Sparse Matrix Anomoly Detector (LSMAD) has the lowest FPR in shallow water but nearly fails in deeper water (Detection Rate (Pt) = 0 at 3m and 5m). Local Sparse Covariance estimator with Total Variation regularization (LSC_TV) shows strong capability at 3m with a detection rate of 0.9273 but is the most parameter-sensitive and computationally intensive. Gated Transformer for Hyperspectral Anomaly Detection (GT-HAD) performs well across all depths, with an average Pt of 0.9664 and an average FPR of 0.1570. The study concludes that water depth significantly affects algorithm performance; KIFD and LRASR are suitable for high detection rates in shallow water, LSMAD for low false alarm requirements, LSC_TV for moderate depths with careful parameter tuning, and GT-HAD is the most reliable across all depths despite relying on extensive training data, providing a practical reference for algorithm selection and improvement in real marine detection scenarios.