Predicting coastal erosion susceptibility in Bangladesh under climate scenario via machine learning techniques

PLOS One | https://doi.org/10.1371/journal.pone.0334347 November 5, 2025 8 / 34 slopes of agricultural land. The final soil loss map was generated by multiplying the five RUSLE components (R, K, LS, C, and P) on a pixel-by-pixel basis. This yielded estimated yearly soil loss figures articulated in tons per acre. Following the estimation of annual soil loss in tons per hectare per year (t ha ⁻ ¹ year ⁻ ¹) over the study area, the results were visualized and categorized into two classes of erosion severity, employing a threshold-based classification method to establish the coastal erosion inventory. Based on [76–79], all pixels exhibiting soil loss rates exceeding 30 t ha ⁻ ¹ year ⁻ ¹ were classified as erosion-prone locations and assigned a rating of 1 whereas, all other pixels with values equal to or below this threshold were classified as non-erosion zones and awarded a value of 0. This inventory not only offers reliable erosion occurrence data grounded in quantitative soil loss estimates but also acts as a fundamental input for further susceptibility modeling and spatial analysis. Selection of explanatory variables Multicollinearity, a prevalent problem in multiple regression, occurs when independent variables exhibit significant cor- relation, which may inflate the standard errors of coefficients and diminish statistical reliability [80]. Pearson correlation matrices provide independent variable correlation coefficients to discover multicollinearity. Multicollinearity is possible if

two variables have a high correlation near 1 or −1 [81]. However, Variance Inflation Factor (VIF) and Pearson’s correlation coefficient (p-value) are employed to measure multicollinearity more precisely [80]. A VIF below 10 shows a significant correlation that may pose issues, while a VIF over 10 signifies severe multicollinearity, potentially distorting regression estimates [82]. Analytic methods, including boosting and bagging technique This study employes five boosting and five bagging algorithms to predict coastal erosion in Bangladesh (Fig 3). XGBoost is one of the strong and robust methods that is utilized in the machine learning field is referred to as [83–85]. For per- forming user-defined predictive tasks such as regression, classification, ranking, and supervised learning, XGBoost is a flexible and efficient ensemble learning framework that makes use of gradient boosting and tree topologies [86,87]. The dataset was normalized and divided into training and testing subsets using stratified sampling. Sample weights were uti- lized throughout training to address class imbalance. XGBoost classifier was trained with 100 trees (n_estimators = 100), a maximum depth of 5, and a learning rate of 0.03, utilizing complete subsampling and feature utilization (subsample = 1, colsample_bytree = 1). Robust L1 and L2 regularization (reg_alpha = 8, reg_lambda = 8) was employed to mitigate overfit- ting. In the current study, the XGBoost algorithm was used, and seventy percent of the database was allocated to training

the model while thirty percent was allocated to testing the model. AdaBoost employs adaptive boosting for classification trees [88]. Adaboost improves the classification precision of machine learning based on classification, and the non-parametric method efficiently identifies outliers without the need to identify bad learners [88,89]. AdaBoost training starts with the creation of a base decision tree through the equal weighting of the dataset [89,90]. AdaBoost is susceptible to noisy or corrupted data [91]. AdaBoost classifier was executed utilizing a decision stump (a decision tree with a maximum depth of one) as the foundational learner. The model was trained using 100 estimators and a learning rate of 0.03. The issue of class imbalance in the training data was mitigated by utilizing sample weights derived from balanced class weighting. The authors employed the AdaBoost model with a 70% and 30% training to test split. Gradient Boosting Decision Trees, simply known as GBDT, is a sophisticated combination algorithm that iteratively enhances regression decision trees to reduce classification errors at each step along the process [92,93]. Numerous weak classifiers are constructed across different iterations, and learning rates are smaller, so more iterations need to be conducted to secure the same fit [94]. It is the correct tree algorithm on which the GBDT depends that scales up the loss function with the negative gradient as a form of residuals, and whose result results in the gradual diminution of residual

values [95,96]. Because of its use of many weak decision trees, GBDT offers stability and accurate projections. Gradient