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

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.