How can we map coastal seabeds quickly and without in‑situ data?

Shallow coastal waters are vital for beach management, ecosystem protection and navigation safety. To know the depth, scientists often use airborne hyperspectral images that show the colour of water and seafloor in great detail. Turning those images into reliable bathymetric maps, however, requires complex models and, crucially, in‑situ reference data, which are frequently missing. Researchers have therefore refined a method called HOPE‑Pure Water (HOPE‑PW), already fast but less precise in capturing fine bottom details. To improve it they added an artificial‑intelligence step that classifies the seafloor type using three machine‑learning algorithms: support vector machine, random forest and maximum likelihood. Tested on images from Wenchang coast, China, the random forest proved best, allowing HOPE‑PW to match the accuracy of the original HOPE model while cutting runtime by 56 % and memory usage by 68 %. In deeper zones the method slightly underestimates depth, likely because the simplification reduces sensitivity to water optical properties. This approach provides a practical tool for monitoring coastal areas lacking direct measurements, supporting more effective marine‑resource management.

The HOPE (Hyperspectral Optimization Process Exemplar) model performs an analytical inversion of seafloor reflectance (ρ) from observed hyperspectral spectra, accounting for water optical properties and solar illumination. Although HOPE delivers high accuracy, its computational cost rises with spatial resolution and number of bands, making it impractical for large datasets. HOPE‑PW simplifies the problem by assuming pure water (no suspended sediments or dissolved organic matter), reducing the number of parameters to estimate and enabling a closed‑form solution that is much faster. However, this assumption limits the ability to capture spatial variations of ρ linked to different bottom types (sand, algae, coral). In this study, the authors introduced a supervised classification step of the substrate using three machine‑learning algorithms: SVM, which seeks the maximum‑margin hyperplane separating classes; Random Forest, an ensemble of decision trees that reduces variance and improves robustness; and Maximum Likelihood, which assigns each pixel to the class with the highest likelihood under Gaussian assumptions. The classification was trained on in‑situ bottom samples collected along the Wenchang coast and then applied to the full hyperspectral scene to produce a probability map for each substrate type. These probabilities were used as constraints in the HOPE‑PW inversion, forcing the ρ estimate to stay within realistic intervals for each class. Validation was performed by comparing depths derived from the model with bottom elevation measurements from the ICESat‑2 laser altimeter, which provides sparse but highly accurate reference points. Results show that, with Random Forest as the optimal classifier, the mean absolute error (MAE) drops from 0.42 m (HOPE‑PW without constraints) to 0.21 m, matching the accuracy of the full HOPE model, while processing time falls from 12.4 s to 5.5 s per tile and RAM usage from 1.8 GB to 0.6 GB. The slight underestimation observed in deeper zones (>5 m) is attributed to the reduced sensitivity of the simplified model to variations in the water attenuation coefficient, which becomes more significant with longer optical paths. Authors plan to extend the approach by incorporating more detailed water optical models (e.g., chlorophyll and particulate matter variables) and testing the technique in other coastal settings such as lagoons and estuaries, as well as exploring convolutional neural networks for substrate classification under cloudy or non‑ideal illumination conditions.