Generalized model for wet/dry shoreline detection and total water level elevation using deep learning

Relative sea-level rise is driving increasingly frequent and impactful coastal inundations. It is important to document these events and their progression through coastal imagery. The wet/dry shoreline is the intersection of water and land surfaces. Precisely georeferencing its location has many positive implications for geological, biological, and societal applications, such as beach and biodiversity management. Current methods in the literature are limited in that they only perform well at specific locations. The present work proposes to use a modified HED (Holistically-Nested Edge Detection) architecture to create a general model for detecting wet/dry shorelines and total water level elevation in coastal imagery, able to be implemented at a variety of different locations and conditions.

The imagery in this study was collected using UAV (Unmanned Aerial Vehicles) during nine field surveys between 2017 and 2020. The area of study is the Gulf of Mexico, having seven datasets in Texas and two in Florida. Unmanned Aircraft Systems (UAS) simplify the acquisition of remote sensing imagery and allow for higher resolutions depending on the flight height. Additionally, it is easier to quickly collect large amounts of data when conditions of interest arise.

We propose the use of a modified HED architecture that allows for a generalized wet/dry detection model. The proposed architecture shows an improvement of 9.4% in the average precision and 8.5% on the F1-Score with respect to the original HED architecture. Additionally, an innovation of the proposed work is the calculation of the total water elevation of the wet/dry shoreline. Calculating the total water elevation is essential for predicting inundation, where the value for the z-coordinate is more important than the x and y-coordinates. To further prove the model's generalization performance, the model was trained in eight locations and tested in an unseen location. When we compare the results of the labeled and the AI wet/dry shoreline elevation we obtained an absolute mean error of 2.1 cm, while an absolute mean standard deviation of 2.2 cm. These results prove that the model is able to generalize and accurately predict the total water shoreline and elevation.