Physics-Guided Random Forest Classification of Marine Sediments Using Frequency-Dependent Acoustic Reflection Spectra

Traditional acoustic seabed classification methods, which are often sensitive to survey geometry and environmental conditions, have limitations in reliability and reproducibility. This study presents a novel physics-guided machine learning framework for automated sediment classification that leverages frequency-dependent acoustic reflection spectra. The framework, tested on two representative sediment types of poorly graded sand (SP) and poorly graded gravel (GP) in controlled laboratory conditions across a frequency range of 100–400 kHz, corrects water-column attenuation and isolates intrinsic sediment responses. Unlike earlier studies that focused solely on attenuation modeling or demonstrated spectral separability without statistical validation, this study embeds physics-guided corrections into a machine-learning pipeline, enabling automated, statistically validated sediment discrimination. Reflection spectra were acquired from 200 samples (100 per class) at 31 frequencies, forming a dataset for classifier evaluation. Random Forest (RF) and Logistic Regression (LR) were benchmarked under identical protocols. RF outperformed LR, achieving peak accuracy of 90% in optimal frequency windows (180–220, 310–350, and 330–370 kHz) and 84% across the full spectrum, compared to LR’s maxima of 82% and 80%. Feature importance revealed that discriminative bands align with wavelengths approximating grain sizes, indicating resonance-like mechanisms. The physics-guided approach demonstrated in this study offers reliable discrimination of sediments with similar grain sizes but different gradations, overcoming a limitation of intensity-only methods. The improved accuracy and interpretability of the classification results have significant implications for future marine survey methods, suggesting that the proposed framework could be a valuable tool for enhancing the efficiency and reliability of seabed characterization. Looking ahead, the potential practical applications of this research are significant, including field trials with autonomous sonar platforms and integration into remote sensing workflows. These applications will be essential to validate the robustness of the approach under real-world variability, paving the way for scalable, real-time seabed classification with implications for a wide range of marine research and applications.