Using deep learning system-based for tide and surge measurement

In recent years, the development of artificial intelligence and machine and deep learning techniques have attracted the attention of various research fields. Among them, the field of land and environmental study has seen some development in applications of this type by placing some interest in remote data acquisition, measurement of marine parameters and their automated analysis (Scardino et al., 2022; Sabato et al., 2023). This work focuses on the implementation of a deep learnig model based on Inception V3 capable of automatically measuring, after appropriate training, the height of tides and storm surge at a given site using fixed camera images. This solution is useful at sites where tide gauges are far from the measurement point and especially for calculating storm surge at the site of interest. The study was carried out on two different Italian locations, the first in southeastern Sicily, in the center of the Mediterranean and called Santa Lucia, and the second in the north of the country called Lignano Sabbiadoro. For convenience and better performance, the development environment was created on Google Colab. To train the deep learning model, a dataset of images was created for each site, and each frame was axed with the corresponding value measured by an in situ instrument. The dataset partitioning was done according to the literature, which as the best partitioning has 70% of the images used for Convolutional Neural Network (CNN) training and the remaining 30% for validation (Götz et al., 2022). Once the model was trained, prediction was performed on images of the site. The algorithm performed well with accuracy above 90% and Categorical Cross Entropy Loss less than 1. Confusion matrices also show good results and the calculated F1 score is above 0.9 (Huang et al., 2015). Finally, from the comparison of the actual values and those processed by CNN, it was possible to see that the values are very similar to each other and the corresponding time-sheets could be processed. In conclusion, the incorporation of systems such as the one presented in this paper could bring many advantages, such as having almost instantaneous feedback on the consequences from intense weather events and eliminating the need for in-person inspection by the operator. These findings underscore its potential to fill the data collection gap in challenging coastal environments, offering valuable insights for coastal management and hazard assessment. This study makes an important contribution to the rapidly growing field of remote sensing and machine learning applications in environmental monitoring, facilitating greater comprehension and decision-making in coastal areas.