An Integrated Feature Framework for Wetland Mapping Using Multi-Source Imagery

Accurate extraction of land cover information and effective classification strategies are crucial for reliable wetland mapping. Data-driven approaches, such as convolutional neural networks (CNNs), demonstrate strong capability in modeling complex nonlinear relationships and learning hierarchical feature representations. However, these methods typically require large labeled datasets, are prone to overfitting and often lack interpretability. In contrast, knowledge-driven approaches based on physical models and expert-defined indices are characterized by interpretable and stable features, but their dependence on predefined formulations restricts flexibility and limits adaptability in heterogeneous environments. To address these limitations, this paper proposes an integrated framework that combines knowledge-driven and data-driven features from multi-source imagery to form a complementary feature set for wetland mapping. All extracted features are incorporated into a Random Forest (RF) classifier, enabling effective utilization of the high-dimensional and heterogeneous feature set. In addition, knowledge-driven and data-driven features are visualized and their importance is analyzed to verify their roles in classification and improve model interpretability. The Yellow River Delta and the Qilihai Wetland, representing study areas with different scales and data conditions, are selected to assess the robustness of the proposed method. The experimental results demonstrate that the proposed approach achieved the best classification performance among all comparative experiments. In the Yellow River Delta and Qilihai Wetland study areas, the Overall Accuracy (OA), Kappa coefficient, and F1-score reached 90.91%, 0.8898, 0.9136 and 91.31%, 0.8893, 0.9308, respectively. In addition, the integration of knowledge-driven and data-driven features effectively proves effective in enhancing classification robustness and improves the interpretability of feature representations.