Assessment of Machine Learning Approaches for Bathymetry Mapping in Shallow Water Environments using Multispectral Satellite Images

Abstract

This paper evaluates the performance of two proposed empirical approaches—random forest (RF) and multi-adaptive regression spline (MARS)—for bathymetry calculations in three diverse areas: the Alexandria harbor shallow coastal area, Egypt, as an example of a low-turbidity, silt-sand bottom water area with depths ranging from 4 m to 10.5 m; the Lake Nubia entrance zone, Sudan, which is considered a high-turbidity, unstable, clay bottom area with a depth of 6 m; and Shiraho, Ishigaki Island, Japan, a coral reef area with a depth of 14 m. Data from Landsat 8 and Spot 6 satellite images were used to evaluate the performance of the proposed models. The bathymetry results of the proposed models were compared with the corresponding results yielded from two conventional empirical methods: the neural network (NN) model and the Lyzenga generalized linear model (GLM). When compared with echo sounder data, the RF and MARS results outperformed Lyzenga GLM results. Moreover, the RF method produced more accurate results with average 0.25 m RMSE improvements range than the NN model. The RF algorithm produced the most accurate results proved to be a preferable algorithm for bathymetry mapping in the shallow water context.