<p>Modelling changes in nearshore bathymetry (<10m depth) is complicated by the nonlinear interactions between sediment, waves, and currents that can cause complex flow and transport patterns such as rip currents. Rip currents are of particular interest because of their implications for both sediment transport and beach-goer safety. An active area of research is using remote sensing (e.g., radar, video imagery) to estimate the existence and location of rip currents. Radar actively measures surface flow directions at high resolutions, however, the equipment can be expensive and difficult to set up. In contrast, video cameras are less expensive and more accessible, but can only provide passive observations that estimate derived surface quantities such as current speed and direction, and wave runup. Time exposure (timex) images from video cameras also provide information about the location of bright pixels (indications of breaking waves). Previous research has relied on the appearance of elongated, shore-normal regions of dark pixels (intersecting bright white regions) as a clear indicator of rip current presence, making timex images a prime candidate for automated detection of rip currents on beaches with video cameras installed. However, it is also known that rip currents vary widely in appearance, and that a better understanding of these parameters is necessary for automated rip current detection.</p><p>In this study, radar data and Argus camera imagery from the United States Army Corps of Engineers Field Research Facility at Duck, NC, USA were evaluated to determine how often radar measured offshore flow indicative of a rip current spatially correlates with dark, shore-normal features in the camera imagery. Radar data for two different times were processed to obtain surface current directions. Timex imagery from the video cameras on the same dates were evaluated with a machine learning algorithm &#160; (Maryan et al. 2019) to objectively define the dark shore-normal features previously assumed to indicate rip currents&#8217; existence within the imagery. A confusion matrix between these two datasets (surface flow direction and machine-identified rip current regions) confirms that dark, shore-normal features in the timex images are not always rip currents, and that offshore directed surface currents are not always visible as dark features in timex images. These results provide the first quantitative evaluation of how often rip current detections are missed and show that additional information is required for accurate automated rip current detection from camera imagery.</p><p>Further analysis will include using wind and wave data from field instruments at the site to reveal which conditions produce (1) offshore flow that is correlated with dark, shore-normal features in the timex imagery, (2) offshore flow that is not correlated with dark, shore-normal features in the timex imagery, and (3) dark, shore-normal features without focused offshore flow. This ongoing study could lead to the clarification of specific conditions under which the existence of rip currents can be correlated with a particular feature that machine learning techniques can be trained to recognize in camera imagery, thereby improving the accuracy of automated rip current detection.&#160;</p>
Quantitative Evaluation of Confidence Measures in a Machine Learning World
