Probabilistic coastal wetland mapping with integration of optical, SAR and hydro-geomorphic data through stacking ensemble machine learning model

The present study focuses on preparing the wetland map using earth observation data and applying a novel ensemble model. Eight advanced machine learning algorithms were applied to determine the probability of the occurrence of wetlands. The random forest (RF), support vector machine (SVM), and multivariate adaptive regression spline (MARS) models outperformed others. So, these models were further used to create a stacking ensemble for improving precision. The RF-SVM-MARS ensemble model was run with seven various parameters. The results show that the integrated parameter has the highest area under the curve (0.960) followed by opticalhydrogeomorphic (0.953), SAR-hydrogeomorphic (0.940), hydrogeomorphic (0.896), SAR-optical (0.892), optical (0.881), and SAR (0.702). The hydrogeomorphic variables exhibited greater influence than the optical and SAR variables. However, the integration of all selected key variables proved to be the most effective approach for probabilistic wetland mapping. The RF-SVM-MARS ensemble model can improve classification accuracy as compared to a single model. An accuracy of 96% was obtained when the ensemble model output was crossvalidated with field data. The ground-penetrating radar (GPR) tool was also successfully employed to study the sub-surface strata and water level conditions in wetland and non-wetland areas. The successful application of a new ensemble approach along with the GPR technique for coastal wetland mapping in this study could encourage researchers to choose it as an appropriate methodology. Moreover, the outcomes of the work will help land-use planners and government agencies for better planning, monitoring, and promoting sustainable development of the coastal area.