Multiscale feature fusion analysis of multibeam and side-scan sonar for seafloor sediment classification in shallow reef environments

Shallow marine environments are critical transition zones between land and sea and exhibit dynamic, spatially heterogeneous seafloor sediments. Conventional single-source acoustic mapping using multibeam echosounder (MBES) or side-scan sonar (SSS) often fails to fully capture this complexity because of survey-orientation effects and instrument-related limitations. To address these challenges, this study proposes a multisource acoustic seafloor mapping framework that integrates MBES bathymetry and SSS intensity data for improved sediment classification in shallow reef environments. A spatial coregistration approach is applied to align MBES and SSS datasets by extracting common feature points and using a high-precision transformation model. A multiscale–multidirectional (MS–MD) feature extraction strategy is then developed by combining bathymetric and acoustic texture features. The Relief-F algorithm is employed to optimize feature selection and reduce redundancy. Five supervised classifiers—Random Forest (RF), K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Random Under-Sampling Boosting (RUS Boost), and Broad Learning System (BLS)—are evaluated. Results show that under the MS–MD feature strategy, the RF model achieves an overall accuracy of 98.26% using fused multisource data, outperforming the MBES-only RF baseline (96.96%). Other classifiers show consistent improvements under the same strategy. In addition to accuracy gains, multisource fusion produces sediment maps with improved spatial coherence and clearer transitional boundaries. Overall, the proposed framework demonstrates the potential of multisource acoustic integration and machine-learning-based classification for high-resolution benthic habitat mapping in complex shallow-water environments.