Comparative analysis of machine learning and deep learning methods for coastal erosion susceptibility mapping

In this study, we describe a comprehensive methodology to assess coastal erosion susceptibility, integrating various input factors, deep learning and machine learning models, and validation metrics. Physical and environmental variables, such as wave height and direction, magnitude of horizontal flow, geology, and slope, were used as inputs, along with coastal erosion inventories, to train and test models, including the Multi-Layer Perceptron (MLP), Functional Trees (FT), Logistic Regression (LR), Na & iuml;ve Bayes (NB), Support Vector Machines (SVM), and Deep Learning (DL). The validation phase employed various metrics for assessing model performance against actual erosion inventories. Factor analysis highlighted wave direction as the most impactful variable, influencing coastal vulnerability significantly. The subsequent model performance evaluation revealed that the MLP model excelled across various criteria (e.g., sensitivity = 94.29%, specificity = 89.93%, accuracy = 92.56%, and area under the curve = 0.99), exhibiting high accuracy and reliability. FT also performed well (e.g., sensitivity = 98.19%, specificity = 97.8%, accuracy = 98.03%, and area under the curve = 0.986), capturing complex nonlinear relationships, while SVM, LR, NB, and DL demonstrated reasonable performance. Comparative advantages of MLP and FT over LR, NB, SVM, and DL are attributed to their ability to handle non-linearity, hierarchical data representation, and flexibility in architecture design. In contrast, limitations of LR, SVM, NB, and DL, such as linearity assumptions, independence assumptions, and data efficiency issues, are acknowledged. Despite variations in model performance depending on the dataset and features, this study underscores the consistent effectiveness of MLP and FT in coastal erosion susceptibility prediction. The findings offer valuable insights for coastal management, guiding resource allocation for mitigation and adaptation strategies. Additionally, the study contributes a nuanced understanding of the interplay between input factors, machine learning models, and validation metrics, enriching the field of coastal vulnerability assessment.