Autonomous classification of wave breaker type in a large wave flume

This dataset accompanies the article "Autonomous Classification of Wave Breaker Type in a Large Wave Flume." It contains the Python code used to train, test, and implement a You Only Look Once-random forest (YOLO-RF) machine learning (ML) model for classifying breaking waves (plunging or spilling) from GoPro videos collected in a wave flume. In addition to the Python code, it contains supplemental files, including the training and testing data sets for the YOLO and RF models, full-length input videos, an example of the model applied to one set of wave conditions, and examples of all files (including intermediary files) generated while training and testing the model. The YOLO model, which classifies five wave features (e.g., prebreaking, curling, splashing, whitewash, crumbling) in a set of video frames, is coupled to an RF model that takes normalized feature counts over multiple frames as inputs and outputs a wave-breaking type for each detected wave. The model, trained and validated with data from a large-scale wave-flume experiment, identifies breaker type with 94% accuracy.

Methods
Data Collection: Laboratory testing was conducted at the O.H. Hinsdale Wave Laboratory large wave flume at Oregon State University. The usable length of the wave flume is about 90m, and the width is 3.7m. The bathymetry was adjusted using concrete slabs, and the height of the flume at the piston-type generator is 4.6 m. The primary objective of these tests was to measure the energy loss and change in breaking induced by the submerged breakwater structures, the latter of which is facilitated by the YOLO-RF approach. A variety of 1/6 scale tests, relative to conditions at a planned deployment site, were conducted with different wave periods and wave heights at a mean water level of 1.44m. Both regular and irregular wave fields were generated from a piston-type wavemaker, but mostly irregular cases are included in this dataset. A JONSWAP spectrum was used to generate irregular waves with specified significant wave heights and peak periods such that the random spectrum, for each set of wave conditions, was identical for each structure layout. In addition to a variety of wave bulk parameters, four different layouts of submerged structures are included in this dataset. Videos ranging from 10-15 minutes in length were collected using a GoPro HERO 5 camera mounted on the sidewall of the flume (left side when facing oncoming waves), approximately 38m from the wave generator, recording at a 60-fps capture rate and 1080x1920 pixel resolution.

Data processing: See the README for a description of the different stages of processed videos and other files generated throughout the machine learning algorithm.