Coast type based accuracy assessment for coastline extraction from satellite image with machine learning classifiers

Abstract

Machine learning (ML) classifiers provide convenience and accuracy in coastline extraction compared to traditional methods and image processing techniques. In literature, the studies about coastline extraction with machine learning classifiers are not focused adequately on the coast types that affect the process. To eliminate this gap, machine learning classifiers were examined in terms of their accuracies on different coastal morphologies. ML classifiers were divided into 3 main groups: Support Vector Machines (SVMs), Multi-Layer Perceptron (MLP) and Ensemble Learning (EL) Classifiers. Within the groups, coastlines were estimated by utilizing different formulas and/or classifiers and their accuracies were examined considering different coast types. Most frequently encountered coastal types, including bedrock, beaches and artificial coasts are included in the study. Bedrock and beach type of coasts were investigated by dividing into sub-groups as shaded, unshaded bedrock coasts and silty-sandy, sandy-gravel beaches. Classifiers were observed as accurate on unshaded bedrock coasts and their results were similar. In spite of that, extraction errors were incurred on the bedrock coasts due to shadows. MLP classifiers with Linear, Logarithmic, and Tanh activation functions were the most accurate in these areas. The challenge was shallow depths and suspended solids in beach type coasts. EL classifiers and SVMs with sigmoidal kernel function were adversely affected on these areas whilst the best results were obtained by utilizing the other SVMs and MLP classifiers. On artificial coasts, successful results were obtained with all classifiers.