pyShore: A deep learning toolkit for shoreline structure mapping with high-resolution orthographic imagery and convolutional neural networks

Authors

Zhonghui Lv, Virginia Institute of Marine ScienceFollow
Karinna Nunez, Virginia Institute of Marine ScienceFollow
Ethan Brewer, Virginia Institute of Marine ScienceFollow
Daniel Runfola, College of William and MaryFollow

Document Type

Article

Department/Program

Virginia Institute of Marine Science

VIMS Department/Program

Center for Coastal Resources Management (CCRM)

Publication Date

1-2-2023

Journal

Computers & Geosciences

Volume

171

Issue

article 105296

Abstract

The process of mapping shoreline structures (i.e., riprap, groins, breakwaters or bulkheads) is heavily reliant on in-situ field surveys and manual delineation using orthoimagery or aerial imagery. These processes are time and resource intensive, resulting in update times of longer than a decade for larger waterbodies. In this study, we explore the effectiveness of a deep learning approach to map shoreline armoring structures from remotely sensed high-resolution imagery. We focus on computationally efficient techniques which can be deployed in desktop environments similar to those used by human coders today, with the goal of providing a semi-automated technique which reduces the total amount of time required to delineate shoreline structures. We test a range of architectures using a dataset of over 10,000 observations of four classes of shoreline structure, finding that a ResNet18 based Pyramid Attention Network (PAN) architecture achieves 72% overall accuracy (60 cm resolution), with 80% and 94% prediction accuracy in breakwater and groins, respectively. This relatively lightweight implementation enabled a 1.5 kilometers of shoreline to be processed in 1.4 s (GPU) to 2.16 s (CPU) in simulated user environments. Finally, we present pyShore, an implementation of this deep learning algorithm made available for human coders to apply as a part of a semi-automated workflow.

DOI

https://doi.org/10.1016/j.cageo.2022.105296

Keywords

Deep learning; Coastal management; Remote sensing; Semantic segmentation

Recommended Citation

Lv Z., Nunez K., Brewer E., Runfola D. pyShore: A deep learning toolkit for shoreline structure mapping with high-resolution orthographic imagery and convolutional neural networks. Computers & Geosciences 171, 105296. doi: https://doi.org/10.1016/j.cageo.2022.105296.