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 21 / 34 Discussion This study analytically evaluated coastal erosion susceptibility in Bangladesh using sophisticated machine learning algo- rithms that combined topographical, hydrological, climatic, geological, and human elements. The Normalized Difference Vegetation Index (NDVI) was identified as the most significant variable by feature selection procedures, specifically Pear- son correlation, variance inflation factor (VIF), and p-values. The significance of vegetation cover in reducing the danger of erosion was highlighted by the NDVI with up to 28% contribution in the Random Forest (RF) model and 26% in XGBoost. Another important component was the Aridity Index, which made up 20% of avNNet and 12–18% of ensemble configura- tions. With contributions of 15% in XGBoost and 14–16% in other models, Distance to Coast (DtC) continuously placed highly across all models. On the other hand, geological factors including soil texture and geology had lower significance ratings, usually less than 6%, suggesting that they had little bearing on the variability of erosion in space. These results are reliable with earlier research by R. Bera and R. Maiti [137] and Islam et al. [138], who also highlighted the supremacy of NDVI and coastal proximity in determining coastal susceptibility. The results, however, are in contrast to those of Gao et al. [139], whose modeling approach based on RUSLE gave geomorphological variables more weight. The deterministic

nature of RUSLE models, which lack the adaptive learning capability found in the machine learning approaches employed here, is probably the cause of the discrepancy. Ensemble machine learning methods consistently beat individual classifiers in terms of model performance. With an area under the curve (AUC) score of 0.95, XGBoost outperformed CatBoost and Bagging, which were closely followed by GBDT, RF, and LightGBM (0.93). With an AUC of only 0.78, simpler models such as Decision Tree performed poorly. High-risk zones ranged from 7.9% (RF) to 13.23% (DT), whereas moderate-risk zones, which ranged from 71.82% (Bag- ging) to 79.36% (GBDT), spanned the largest areas across all models. Particularly, the AUC values in this study surpass those reported by Islam et al. [138] for distinctive between high and low erosion susceptibility zones in the Chattogram coastal region, excellent model performance is indicated by the AUC value of 0.909. Additionally, Bammou et al. [140] documented an AUC of 0.96 for the XGBoost model in their soil erosion evaluation, corroborating the dependability of superior performance results in analogous situations. This suggests that the existing ensemble configurations for coastal erosion scenarios have a greater level of predictive dependability. A different result was reported by Rahmati et al. [141], for susceptibility mapping in Iran, Random Forest (AUC 0.91) performed somewhat better than XGBoost (AUC 0.89). This demonstrates how dataset properties and regional conditions can affect model performance.

According to the high-resolution susceptibility maps created for this study, Bhola (19.41%), Cox’s Bazar (26.20%), and Patuakhali (21.47%) are some of the districts that are most susceptible to erosion at the moment. Conversely, districts with large percentages of low-risk zones include Jashore (86.94%), Narail (71.24%), and Gopalganj (52.71%). This is Table 4. Coastal Erosion Susceptibility Zones Distribution Based on ML Models. Low Moderate High Percentage Area Percentage Area Percentage Area AdaBoost 9.28% 4380.12 78.69% 37143.66 12.03% 5676.21 CatBoost 18.50% 8733.79 72.68% 34306.06 8.81% 4160.16 GBDT 10.65% 5025.94 79.36% 37457.76 9.99% 4716.31 LightGBM 12.60% 5948.73 77.81% 36727.76 9.58% 4523.51 XGBoost 18.30% 8636.01 73.58% 34729.97 8.12% 3834.02 avNNet 10.02% 4730.51 79.20% 37381.93 10.78% 5087.56 Bagging 19.03% 8981.01 71.82% 33900.16 9.15% 4318.82 DT 12.41% 5855.37 74.36% 35100.18 13.23% 6244.46 RF 15.96% 7533.13 76.14% 35939.47 7.90% 3727.40 Treebag 14.37% 6784.35 76.59% 36148.45 9.04% 4267.21 https://doi.org/10.1371/journal.pone.0334347.t004