Efficiency of different machine learning methods for shoreline extraction from UAV images

Shorelines are valuable assets of coastal areas that possess unique features. A huge user community is associated with the shorelines, which may be affected by both natural and human-induced factors. Monitoring of such areas is crucial for environmental protection, sustainable development and planning. Shoreline extraction is a fundamental step to determine their dynamics that is crucial for conservation of various natural habitats. The recent developments in Unmanned Aerial Vehicle (UAV) systems have made it possible to acquire high-resolution images at low cost, which can be exploited for shoreline mapping. This study aims to investigate the efficiency of Support Vector Machines (SVM) and Random Forest (RF) classifiers to obtain accurate and reliable water body and land classes from a very high-resolution digital orthophoto achieved using UAV-images. In the study, a 3.3 km shoreline located on the Black Sea coastline was selected as the study area. The digital orthophoto was generated using commercial software based on the SfM method with 1.96 cm ground sample distance. The SVM and RF methods were used to achieve land, water body and wave classes to generate the shoreline of the study area. The same training data were used for both the SVM and RF methods. Rapid illumination changes during flights caused unbalanced color on the entire image. Considering this effect, nine patches were created from the orthophoto image. Each patch was handled as an individual image and training samples had to be selected for each class from each patch. The shorelines have been generated from binary water body and land classes. Accuracy ssessment was performed by comparing results with Global Navigation Satellite System (GNSS) measurements. The average errors were calculated as 0.935 m and 0.939 m for the SVM and RF methods, respectively. The differences of 71.86% and 71.61% are in the 0-1 m interval for SVM and RF, respectively. Although the SVM and RF methods have been used in various studies for image classification purposes, they have not been utilized for shoreline extraction. This study shows that these two machine-learning methods also provide remarkable results for this purpose. The proposed framework can be used to create an essential data source for many coastal-related studies such as monitoring, determining temporal changes and modelling sediment transport. This study has been supported by TUBITAK (The Scientific and Technological Research Council of Turkey) with project number 115Y718.