Optical wave gauging using deep neural networks
Optical wave gauging using deep neural networks
复制标题
DOI：
10.1016/j.coastaleng.2019.103593
复制
发表时间：
2020-01-01
影响因子：
4.4
通讯作者：
Warrick, Jonathan A.
中科院分区：
文献类型：
--
作者：
Buscombe, Daniel;Carini, Roxanne J.;Warrick, Jonathan A.
我们开发了一种远程波浪测量技术，可以根据冲浪区波浪的图像来估计波高和周期。在这项概念验证研究中，我们将相同的框架应用于三个数据集：第一个是美国北卡罗来纳州达克的单个近岸波浪的一组近距离单色红外 (IR) 图像；第二个是美国加利福尼亚州圣克鲁斯附近较大近岸区域的一组可见（即 RGB）波段正射马赛克；第三，来自同一站点的一组倾斜（未校正）图像。该网络使用重合图像和原位波测量进行训练。光波规 (OWG) 由一个深度卷积神经网络 (CNN) 组成，用于从图像中提取特征（称为“基本模型”），附加层将特征信息提取到较低维空间，最后一层密集神经元用于预测连续变化的量。比较了四种基本模型。 OWG 针对单个波高和周期以及有效波高和峰值波周期等统计量进行训练。 IR 数据集上表现最佳的 OWG 在高度和周期方面分别实现了 0.14 m 和 0.41 s 的 RMS 误差，捕获了这些量中高达 98% 的方差。在可见光波段校正数据集上表现最佳的 OWG 在高度和周期方面分别实现了 0.08 m 和 0.79 s 的 RMS 误差。倾斜 RGB 图像的高度和周期的值分别为 0.11 m 和 0.81 s。总体而言，波高和周期精度对基础模型的选择很敏感；基于 MobilenetV2 构建的 OWG 往往表现最差，而基于 Inception-ResnetV2 构建的 OWG 具有最小的 RMS 误差。模型中是否存在残余层对最终 OWG 精度几乎没有系统影响。模型训练中使用的批量较小往往会产生更准确的 OWG。使用与训练数据中表示的值范围之外的波高或周期相关的图像进行的校准外验证表明，OWG 预测低波高底部 5% 和高波高顶部 5% 的能力相当不错，但对于波周期而言，情况通常并非如此。尽管电磁波段、视角和尺度存在差异，但未针对任一数据集进行优化的相同框架在图像训练时可以高精度预测这两个数量。 OWG 在中等大小的 CPU 上在不到 100 毫秒的时间内根据图像估计波浪属性，从而可以进行连续实时波浪估计。
We develop a remote wave gauging technique to estimate wave height and period from imagery of waves in the surf zone. In this proof-of-concept study, we apply the same framework to three datasets: the first, a set of close-range monochrome infrared (IR) images of individual nearshore waves at Duck, NC, USA; the second, a set of visible (i.e. RGB) band orthomosaics of a larger nearshore area near Santa Cruz, CA, USA; and the third, a set of oblique (unrectified) images from the same site. The network is trained using coincident images and in situ wave measurements. The optical wave gauge (OWG) consists of a deep convolutional neural network (CNN) to extract features from imagery - called a 'base model', with additional layers to distill the feature information into lower dimensional spaces, and a final layer of dense neurons to predict continuously varying quantities. Four base models are compared. The OWG is trained for both individual wave height and period, and statistical quantities like significant wave height and peak wave period. The best performing OWG on the IR dataset achieved RMS errors of 0.14 m and 0.41 s for height and period, respectively, capturing up to 98% of the variance in these quantities. The best performing OWG on the visible band rectified dataset achieved RMS errors of 0.08 m and 0.79 s, respectively, for height and period. The same values for the oblique RGB imagery were 0.11 m and 0.81 s for height and period, respectively. Overall, wave height and period accuracy is sensitive to choice of base model; OWGs built upon MobilenetV2 tend to perform worst and those built on Inception-ResnetV2 have the smallest RMS error. The presence or otherwise of residual layers in the model makes little systematic difference to the final OWG accuracy. Smaller batch sizes used in model training tend to result in more accurate OWGs. An out-of-calibration validation, using images associated with wave heights or periods outside the range of values represented in the training data, showed that the ability for OWGs to predict the bottom 5% of low wave heights and the top 5% of high wave heights was reasonably good, but the same was not generally true of wave period. The same framework, not optimized for either dataset, predicts both quantities with high accuracy when trained on imagery, despite the differences in electromagnetic band, perspective, and scale. The OWG estimates wave properties from an image in less than 100 ms on a modestly sized CPU, allowing for the possibility of continuous real-time wave estimates.