Coastal Wetland Mapping Using Ensemble Learning Algorithms: A Comparative Study of Bagging, Boosting and Stacking Techniques

Abstract

Coastal wetlands are a critical component of the coastal landscape that are increasingly threatened by sea level rise and other human disturbance. Periodically mapping wetland distribution is crucial to coastal ecosystem management. Ensemble algorithms (EL), such as random forest (RF) and gradient boosting machine (GBM) algorithms, are now commonly applied in the field of remote sensing. However, the performance and potential of other EL methods, such as extreme gradient boosting (XGBoost) and bagged trees, are rarely compared and tested for coastal wetland mapping. In this study, we applied the three most widely used EL techniques (i.e., bagging, boosting and stacking) to map wetland distribution in a highly modified coastal catchment, the Manning River Estuary, Australia. Our results demonstrated the advantages of using ensemble classifiers to accurately map wetland types in a coastal landscape. Enhanced bagging decision trees, i.e., classifiers with additional methods to increasing ensemble diversity such as RF and weighted subspace random forest, had comparably high predictive power. For the stacking method evaluated in this study, our results are inconclusive, and further comprehensive quantitative study is encouraged. Our findings also suggested that the ensemble methods were less effective at discriminating minority classes in comparison with more common classes. Finally, the variable importance results indicated that hydro-geomorphic factors, such as tidal depth and distance to water edge, were among the most influential variables across the top classifiers. However, vegetation indices derived from longer time series of remote sensing data that arrest the full features of land phenology are likely to improve wetland type separation in coastal areas.