An empirical performance comparision of machine learning algorithms for detecting coastline from satellite images

Coastal areas, which contain valuable resources for humans and other species, need to be monitored and protected against many dangers. One of the most important and necessary steps in the protection and monitoring of the coasts is the coastline extraction process. Despite the difficulties and high costs of traditional methods for this process, developing remote sensing techniques and computer science directed the researchers to study the Machine Learning Classifiers. Although there are many studies on this subject in the literature, there are not enough study depending on the coastal types which known to have an important effect on coastline extraction. In this study, it was tried to study the accuracy of machine learning classifiers on different coast types in detail, which were found successful for coastline extraction from satellite images. The coastal types studied include cliffs with and without shaded areas, clayey-silty and sandy gravel beaches and artificial coasts. 12 classifiers consisting of different functions and algorithms belonging to Support Vector Machines, Multi-layer Perceptron's and Ensemble Learning classifiers were examined on these coast types. As a result of the study, Ensemble Learning classifiers achieved lower accuracy in all studied coast types compared to other classifiers. While Artificial Neural Networks and Support Vector Machines achieved high accuracy in many different coast types, it was found that the Support Vector Machines using the Sigmoid kernel function achieved the lowest accuracy in all coast types.