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 10 / 34 Boosting classifier was trained with 100 estimators, a maximum tree depth of 5, a learning rate of 0.03, and subsample and feature sampling rates of 0.8 to enhance generalization. It is thus the best solution. 70% of the time is spent in train- ing and 30% in testing here. LightGBM is an open-source gradient boosting library founded by Microsoft that utilizes the histogram-based method to optimize training efficiency and reduce computational complexity [97,98]. In addition, it has enhanced network commu- nication to enable additional parallel learning, for example, the parallel decision tree voting technique and grows trees by managing variance gain and sorting training examples in decreasing order based on absolute gradient values [99,100]. The LightGBM classifier was trained using 100 estimators, a maximum tree depth of 5, a learning rate of 0.03, and complete subsampling of both data and features (subsample = 1, colsample_bytree = 1). Robust regularization was imple- mented through L1 (reg_alpha = 8) and L2 (reg_lambda = 8) penalties, while class imbalance was mitigated by employing balanced class weights. Feature significance was derived and standardized to ascertain the most significant predictors. Thirty percent of the data was used for testing, and seventy percent was used for training for this study. CatBoost is a very advanced machine learning method that has been effective in capturing a variety of factors influenc-

ing water potential as well as properly handling noisy and heterogeneous information [101]. As a Gradient Boosting Deci- sion Tree-based technique, it is particularly well-suited to model complex relationships and heterogeneous feature sets [101,102]. Sample weights were calculated to address class imbalance, with class weights explicitly designated according to the imbalance ratio. A CatBoost classifier was trained with 100 iterations, a maximum depth of 5, a learning rate of 0.03, and an L2 regularization value of 10. Feature subsampling was established at 70% (rsm = 0.7) to improve generalization. According to this study, a statistically target-oriented approach is employed wherein seventy percent of the training data is randomly chosen, and thirty percent is reserved for testing. Random Forest (RF) is a statistical ensemble machine learning method employed most of the time to forecast the target variable [24,103,104]. The method employs bootstrapping techniques along with stochastic binary trees for the purpose of using some portion of the data that provides a better representation. The model is constructed by the random selection of training from the initial set of observations [105]. The bagging classifier was trained with 100 decision trees, each with a maximum depth of 5, and minimum sample split and leaf sizes adjusted to mitigate overfitting. Probabilities were computed for predictions, and feature importance was averaged and normalized across all base estimators to find

significant contributors. Within the experiment, the information is divided into two sets: thirty percent is held out for testing, and seventy percent for training. The bagging strategy for an ensemble entails repeatedly training the same algorithm on numerous distinct subsets of the training set [105,106]. The average of all the projections from the sub-models forms the basis of the final forecast. This process minimizes heterogeneity in classification, which ultimately increases overall accuracy [107]. Bagging is an approach that can be used to improve accuracy in situations where big fluctuations in the training set significantly affect the subsequent prediction, where the outcome of a test sample is determined through the vote of the ensemble that received the highest vote [108,109]. A Bagging Regressor consisting of 100 decision trees, each with a maximum depth of 5, was trained, utilizing 1% of the training data per tree to mitigate overfitting while employing all features. Despite being intended for regression, the model’s continuous predictions were utilized as probabilities for binary classification. The significance of features was averaged throughout the ensemble and normalized to emphasize principal predictors. Two data sets were established for the study, thirty percent of which were reserved for testing and seventy percent reserved for training. In order to execute the decision tree (DT) method, the data are first split into various regions according to the inde-

pendent variables, and rules for each region are subsequently developed [110,111]. The primary application of decision tree models is to obtain the development of a model for predicting results on the basis of input characteristics. Decision trees are significantly effective in classification, regression, and prediction tasks [112,113]. This code trains a Decision Tree classifier on the specified dataset, utilizing features associated with environmental and geographic indices to predict