Shallow-Water Bathymetry Retrieval Based on an Improved Deep Learning Method Using GF-6 Multispectral Imagery in Nanshan Port Waters

In a seaport, accurate bathymetric maps are valuable for both environmental and economic reasons. One of the main complementary methods for measuring shallow-water depth is the retrieval of the water depth by satellite. The results of the water depth inversion are greatly influenced by factors related to water quality. The proposed updated quasi-analysis algorithm (UQAA) allows for the calculation of water quality factors, and their spatial distribution characteristic strongly correlates with the trend in water depth distribution. By using satellite-derived bathymetry, these parameters can be used in the model training to extract the underwater terrain. This article proposes the idea of combining the UQAA with a convolutional neural network (CNN) based deep learning framework to retrieve the depth of the water and automatically extract the underwater terrain. We compare four different existing machine learning algorithms as baselines, using GF-6 multispectral remote-sensing images and in situ depth data in Nanshan Port as a priori validation set. We find that the result of the CNN model using the UQAA is better than other baselines, where the root-mean-square error was down to 0.55 m, the mean relative error was 6.63%, and the R² was 0.92. The developed method, which introduces the water quality factors containing geographic information as feature quantities, provides a new direction for further improvement.