Although easily discernible by the naked eye, a robust and consistent methodology to identify the spatio-temporal occurrence of wave breaking in the nearshore on a wave-by-wave basis has been elusive to date. In this work, a Convolutional Neural Network (U-Net) is trained and its performance evaluated using a large number of images in the electro-optical range, and its performance is compared against an existing sensor-fusion methodology. The results show a good performance of the resulting U-Net model, matching nearly 71% of the breaking instances detected by the sensor-fusion approach. Although this value can be seen as low, qualitative comparisons show that in many cases, wave breaking identification is improved by the U-Net model. Moreover, a sample application to a different surfzone showed good qualitative performance, suggesting its applicability to other wave conditions, with short processing times. This offers the possibility of implementing automated wave breaking detection that could enhance our understanding of nearshore processes. The resulting the U-Net model is made available to the community for future testing and continuous development.

虽然用肉眼很容易辨别，但迄今为止，一种稳健且一致的方法来识别近岸波浪破碎的时空发生一直难以捉摸。在这项工作中，训练卷积神经网络 (U-Net) 并使用电光范围内的大量图像评估其性能，并将其性能与现有的传感器融合方法进行比较。结果表明，生成的 U-Net 模型具有良好的性能，匹配了传感器融合方法检测到的近 71% 的破坏实例。尽管该值可以被视为很低，但定性比较表明，在许多情况下，U-Net 模型改进了波浪破碎识别。此外，对不同冲浪区的示例应用显示出良好的定性性能，表明它适用于其他波浪条件，处理时间短。这提供了实施自动破浪检测的可能性，可以增强我们对近岸过程的理解。由此产生的 U-Net 模型可供社区用于未来的测试和持续开发。