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 18 / 34 Coastal erosion susceptibility zones distribution based on machine learning models Table 4 and Fig 11 illustrate the coastal erosion vulnerability zones categorized into low, moderate, and high impact levels. Table 4 demonstrates that there are comparable patterns among AdaBoost, GBDT, avNNet, and LightGBM, with a high percentage of moderate-impact areas (77.81%–79.36%), low-impact areas (around 10%–12.6%), and high-impact areas (9.58%–12.03%). CatBoost and XGBoost likewise exhibit similar outcomes, with high impact coming in slightly lower at 8.12%–8.81%, low impact areas at about 18%, and moderate impact predominating at 72.68% and 73.58%, respectively. Between 71.82% and 74.36%, the Bagging and DT models have moderate impact; Bagging has the largest low-impact area (19.03%), while DT has a noteworthy high-impact proportion (13.23%). With a moderate impact of almost 76%, RF and TreeBag exhibit comparable patterns; however, RF’s high-impact area is smaller (7.9%) than TreeBag’s (9.04%). Validation of models by ROC and AUC Fig 12 depicts the results of model validation through ROC curves, confusion matrices, F1-scores, precision, and recall, highlighting significant performance disparities. XGBoost proved to be the most effective model, attaining a maximum AUC of 0.95 and an F1-score of 0.81, with precision and recall metrics surpassing 0.68. Close competitors included Bagging and CatBoost, both attaining robust results with AUCs about 0.94 and F1-scores of 0.79. Models including

GBDT, Random Forest, and LightGBM demonstrated comparable efficacy, attaining AUCs of 0.93 and F1-scores of 0.78 and 0.80. Both ANN and Treebag demonstrated somewhat suboptimal performance, with AUCs of 0.92 and F1-scores between 0.73 and 0.78. AdaBoost demonstrated marginally inferior performance, achieving an AUC of 0.91 and an F1-score of 0.72. The Decision Tree model had the lowest evaluation, attaining an AUC of 0.78 and an F1-score slightly under 0.70. Table 3. Multicollinearity and Significance Testing for Model Variable Selection. Feature VIF Tolerance Pearson Correlation P-Value Aridity Index 3.48 0.29 0.11 0.001 Aspect 2.97 0.34 −0.03 0.016 Curvature 1.01 0.99 −0.01 0.603** DEM 6.15 0.16 0.09 0.001 DtC 1.34 0.75 −0.30 0.001 DtR 1.60 0.63 0.01 0.562** DtF 4.96 0.20 −0.15 0.001 Geology 7.84 0.13 0.10 0.001 Geomorphology 9.26 0.11 −0.16 0.001 LULC 9.06 0.11 0.02 0.389** Mean Wind Speed 3.13 0.32 0.12 0.001 MSL 459.99*** 0.00 −0.13 0.001 NDVI 9.31 0.11 −0.57 0.001 Precipitation 240.81*** 0.00 0.22 0.001 Slope 3.74 0.27 0.18 0.001 Soil Texture 7.96 0.13 0.05 0.021 TPI 1.53 0.65 −0.02 0.371** TRI 2.12 0.47 −0.07 0.001 TWI 1.33 0.75 −0.01 0.779** VDCN 2.89 0.35 −0.21 0.001 Note: The model excludes variables with p-values above 0.05 (**) and VIFs above 10 (***) because they are statistically insignificant. https://doi.org/10.1371/journal.pone.0334347.t003