ABSTRACT
Wetlands remain the least well-known and most threatened ecosystems, which are important for the well-being of humans as well as flora and fauna. Wetland maps are of the essence for the conservation and management of terrestrial ecosystems. Since single classifiers fail to obtain wetland maps with high accuracy, we proposed an object-based stacked generalization (Stacking) method for wetland mapping using the combination of Sentinel-1and Sentinel-2 data. Additionally, to guarantee a satisfactory prediction capacity of the “base learner” (BL), we utilize the genetic algorithm automatically for hyper-parameter tuning to ensure mapping accuracy. Firstly, we segment Sentinel-2 multispectral images to delineate objects in this study. Then, we extract the object-oriented features from Sentinel-1 and Sentinel-2 image objects. To reduce the data redundancy, we calculate the feature importance using the random forest algorithm to optimize the feature set. Thirdly, four well-known classifiers, namely K-Nearest Neighbor (KNN), Random Forest (RF), Support Vector Machine (SVM), and Maximum likelihood Classification (MLC), are used as the base classifier. Finally, the stacked generalization algorithm is utilized to perform wetland classification based on the optimal feature set in the Yellow River delta wetland. Experimental results demonstrate that the proposed object-based stacked generalization approach achieves better classification results.
