Convolutional Neural Network Based Semantic Segmentation for Coastal Flooding Analysis

Abstract:

The use of remote sensing to gather information through images obtained from UAVs, satellites orthoimagery and fixed camera arrays is becoming increasingly popular in the field of coastal engineering and hydrology. One of the most important question, however, is related to tracing the position and the size of different coastal regions in the image, e.g. water, wet sand, dry sand, vegetation, etc. This would allow a more complex analysis of the bio-geomorphic processes driving coastal changes. Here, we explore the use of convolutional neural network (CNN) for semantic segmentation of small-scale high-temporal resolution coastal imagery, and present some results related to water segmentation. We show that training from scratch with a relatively small number of images could provide reliable segmentation results, without any transfer learning. Furthermore, we propose a simple but powerful, post-processing method for more accurate semantic segmentation, by minimizing errors mostly induced by false positives. We use the trained algorithm to obtain a high-resolution, 6-month long, time series of flooded beach area in a Harvey-impacted region in the Texas coast. We find that beach flooding events are correlated at the sub-hour scale but at higher-scales can be modelled as a Poisson process with exponentially distributed marks. Both the statistical properties of flooding events, and their frequency, about 3 events per month, are consistent with recent predictions using an empirical model of wave run-up based on deep-water wave properties.