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 11 / 34 a binary target variable. The data is divided into training and test sets based on the DT strategy, with thirty percent of the data left for testing and seventy percent left for training. Treebag is an ensemble learning method that takes advantage of the benefits of decision trees and bagging in order to improve prediction accuracy that is achieved when multiple decision trees are made by constructing bootstrapped subsets of the training data, each of which was trained separately [114,115]. Through the mean of error averaging over many trees, Treebag improves the stability and robustness of decision tree implementations and this technology has been applied with high frequency to vast topics, including issues of classification and regression, and has achieved better per- formance than single decision tree models [116–119]. With a thirty percent test split and seventy percent train split, A Tree- bag classifier with limited complexity (maximum depth of 5, minimum samples for split and leaf) was employed to mitigate overfitting. The Bagging Classifier was subsequently trained using 100 estimators, each utilizing the complete sample and feature set, while employing parallel processing for enhanced efficiency. Treebag is utilized in the present research. Averaged Neural Network, or avNNet, is an ensemble learning method utilized for increasing predictive accuracy by averaging the output of numerous neural networks [120]. It is a method that does not rely on one trained network but

one that utilizes predictions of multiple models trained independently and then averages the models’ output [121]. When overfitting problems must be resolved and model robustness has to be enhanced for hard learning tasks, such a method is quite helpful. Through the utilization of various neural networks trained on various subsets of the data or initially charged with various weights, avNNet has been found to enhance classification and regression task stability, as demonstrated by research [122,123]. This renders the process of averaging to be a stable method to employ in machine learning settings since it reduces the impact of individual network biases and improves the accuracy of the entire output [124]. The avN- Net classifier was executed with a multilayer perceptron architecture featuring two hidden layers containing 100 and 50 neurons, respectively. The avNNet technique is employed within this study by splitting 70% of the dataset for training purposes and 30% for testing purposes, according to the dataset. For all models, to rectify class imbalance, we utilized the class_weight = ‘balanced’ argument, which modifies weights according to class frequencies. Moreover, stratified sampling was employed in the train-test division to maintain class distribution. Model validation Better performance in machine learning models is associated with greater AUC values [125]. The study graphs ROC curves for all of the models through software in the programming language Python. The ROC curve is a plot of a classi-

fier’s performance in distinguishing between numerous classes [126,127]. The Area Under the Curve (AUC) is a scalar measure of the model’s performance that originated from the Receiver Operating Characteristic (ROC) curve. The data was divided in a training-to-testing proportion of 70% and 30%, so that there was an equilibrium between model training and validation. The AUC score and ROC curve were utilized as primary measures of performance, with significantly larger AUC scores indicating superior classification capacity (Table 2). The best classifier will have a ROC curve that is clumped in the upper-left quadrant, which indicates that it has a high level of both sensitivity and specificity [86,126,127]. According Table 2. Classification of Model Performance According to AUC (Area Under the Curve). Values [86,126,127]. AUC Range Classification Interpretation 0.91–1.00 Excellent Very high classification accuracy, close to perfect performance 0.81–0.90 Very Good Strong predictive ability with a low misclassification rate 0.71–0.80 Good Reliable model performance, though not flawless 0.61–0.70 Satisfactory Significant predictive efficacy, acceptable for general use 0.51–0.60 Unsatisfactory Slightly better than random chance; needs further improvement https://doi.org/10.1371/journal.pone.0334347.t002