Incorporation of neighborhood information improves performance of SDB models

Mapping of water depth in nearshore waters is important for safe navigation and other human uses of coastal areas. Optical satellite imagery offers a cost-efficient source of data for this purpose. Several methods have been developed for satellite-derived bathymetry (SDB), but the details of their implementation, as well as their relative performance and its dependence on the data and environmental context, is unclear. We compared five SDB models, calibrated and tested against airborne lidar data from southern Corsica. We used a blocked strategy to split data into calibration, validation, and test data, because randomly splitting the data resulted in spatial autocorrelation between these data sets and underestimation of model prediction errors. In general, the non-parametric random forest and neural network models outperformed the parametric multi-band and band-ratio models. The convolutional neural network (CNN) model performed the best, especially in relatively deep (10–20 m) areas, and it could be near-optimally tuned with only ∼1000 data points. However, developing this CNN model was an iterative trial-and-error process that took several months. Using the random forest model, we demonstrated that using information from more than two bands in the visible and near-infrared spectrum contributed to improving model performance, as did incorporation of information from the local neighborhood. We suggest that the use of more than two bands, and the inclusion of spatial contextual information in addition to the values of the individual pixel, should become standard practice in SDB.