Evaluating gradient descent variations for artificial neural network bathymetry modeling and sensitivity analysis

Abstract

Artificial intelligence has been widely applied to water depth retrieval across various environments, deemed essential for habitat modeling, hydraulic structure design, and watershed management. However, most of these models have been developed for deep waters, with the critical impact of the gradient descent algorithm often not evaluated. To address this gap in current research, this study adopted the artificial neural network with seven gradient descent methods, including step, momentum, quick propagation, delta-bar-delta, conjugate gradient, Levenberg-Marquardt, and resilient backpropagation (RProp), for shallow water depth modeling. Shallow water depths in Taiwan's mountainous rivers were then modeled using multispectral imagery taken by drone and vegetation indices. From our results, it was revealed that methods optimizing weight updates were outperformed by those based on gradient information, such as RProp. The selection of gradient descent algorithm was identified as pivotal; an inappropriate selection might even result in performance inferior to a traditional linear regression model. In the sensitivity analysis, near-infrared and normalized difference water index were classified as highly sensitive. By leveraging multispectral data and vegetation indices with ANN, the optimal gradient descent algorithm and the critical model input for shallow water modeling were successfully identified, offering invaluable insights for future studies.