VEdge_Detector: automated coastal vegetation edge detection using a convolutional neural network

Abstract

ABSTRACT Coastal communities, land covers, and intertidal habitats are vulnerable receptors of erosion, flooding or both in combination. This vulnerability is likely to increase with sea level rise and greater storminess over future decadal-scale time periods. The accurate, rapid, and wide-scale determination of shoreline position, and its migration, is therefore imperative for future coastal risk adaptation and management. This paper develops and applies an automated tool, VEdge_Detector, to extract the coastal vegetation line from high spatial resolution (Planet’s 3 to 5 m) remote-sensing imagery, training a very deep convolutional neural network (holistically nested edge detection), to predict sequential vegetation line locations on annual to decadal timescales. Red, green, and near-infrared (RG-NIR) was found to be the optimum image spectral band combination during neural network training and validation. The VEdge_Detector outputs were compared with vegetation lines derived from ground-referenced positional measurements and manually digitized aerial photographs, which were used to ascertain a mean distance error of <6 m (two image pixels) and >84% producer accuracy (P A) at six out of the seven sites. Extracting vegetation lines from Planet imagery of the rapidly retreating cliffed coastline at Covehithe, Suffolk, United Kingdom, has identified a landward retreat rate >3 m year−1 (2010–2020). Plausible vegetation lines were successfully retrieved from images in The Netherlands and Australia, which were not used to train the neural network, although significant areas of exposed rocky coastline proved to be less well recovered by VEdge_Detector. The method therefore promises the possibility of generalizing to estimate retreat of sandy coastlines from Planet imagery in otherwise data-poor areas, which lack ground-referenced measurements. Vegetation line outputs derived from VEdge_Detector are produced rapidly and efficiently compared to more traditional non-automated methods. These outputs also have the potential to inform upon a range of future coastal risk management decisions, incorporating future shoreline change.