Automation of the Van de Casteele Test via Deep Learning: Feasibility and Analysis of Instantaneous Sampling

Abstract

The Van de Casteele Test is an important step in analyzing the quality of observations from tide gauge stations, but its manual execution is laborious. To automate this process, this research evaluated the use of a deep learning model based on YOLO, applied to images of the tide staff at the Arraial do Cabo tide gauge station, captured by low-cost digital cameras. The study also analyzed the limits of instantaneous inference, characterized by direct image measurement without the common time-averaging filters found in electronic sensors. In tide staff detection, the model showed excellent performance across various scenarios, with mAP50 of 0.995 and mAP50-95 ranging from 0.844 to 0.913. The metrological validation compared five experiments, using the conventional method as a reference. Regarding accuracy, while the Radar and Encoder sensors showed standard deviations (SD) of 0.02 m when compared to tide staff readings taken by an observer, the computer vision model converged to a SD of 0.03 m throughout the experiments. This slightly higher variability, together with an identified systematic bias, is attributed to the instantaneous capture of the sea surface, which records high-frequency noise (waves) that is filtered out by the temporal integration of level sensors and by the observer. Therefore, the results demonstrate that, given the instantaneous nature of image capture, the technique is robust and feasible for automating the Van de Casteele Test.