Abstract: Wetland landscape components extracted from remote sensing images provide essential information for investigating the spatio-temporal evolution, ecological processes, and dynamic changes of wetland ecosystems. Accurate and efficient extraction of these components is particularly important for plateau wetlands, where complex terrain, heterogeneous vegetation, seasonal hydrological fluctuations, and fragmented landscape patterns often lead to high spectral similarity among different surface features and irregular spatial boundaries. However, existing object-oriented semantic segmentation methods usually require labor-intensive sample preparation and complicated object-level annotation, which limits their applicability to large-scale or long time-series wetland monitoring. In contrast, conventional pixel-based methods are relatively easy to implement, but they often lack sufficient capability to learn spatial contextual information, resulting in limited classification accuracy, especially in complex and highly variable plateau wetland environments. To address these limitations, this study proposes a spectral-spatial double-branch interaction model for extracting landscape components in plateau wetlands from remote sensing imagery. The proposed model is designed to fully integrate spatial contextual features and spectral representation features through two complementary branches. In the spatial branch, multi-scale convolution is first applied to neighboring pixels to capture spatial information at different receptive fields, enabling the model to better represent landscape components with varying sizes, shapes, and spatial distributions. Deformable convolution is then introduced to extract spatial-spectral features, thereby enhancing the model’s ability to characterize irregular boundaries and complex edge structures, which are common in plateau wetland landscapes. In the spectral branch, the original pixel features are first mapped into a higher-dimensional feature space. A weighted-averaging Vision Transformer architecture is then employed to extract fused spectral features from both low-dimensional and high-dimensional representations of pixels. This design improves the model’s capacity to capture subtle spectral differences among wetland landscape components with similar spectral responses. Finally, a cross-attention mechanism is used to adaptively align spatial features and spectral features, allowing the model to strengthen the complementary interaction between the two types of information and improve the overall robustness of feature representation. In this study, a remote sensing dataset was constructed for the Sichuan Changsha Gongma Nature Reserve, and nine categories of wetland landscape components were extracted. Experimental results show that the proposed model achieved an overall accuracy of 95.86%, a macro-averaged precision of 97.64%, an average recall of 94.67%, a mean intersection over union of 92.45%, and a macro-averaged F1-score of 95.89%. These results demonstrate that the proposed model can effectively extract wetland landscape components while maintaining favorable training efficiency. Compared with existing advanced network models, the proposed spectral-spatial double-branch interaction model shows superior performance in complex plateau wetland scenarios, indicating its potential for accurate wetland landscape monitoring and long time-series ecological change analysis.