Changes in dominant plant communities in reservoir water-level-fluctuation zones are important indicators of river ecosystem evolution and provide a key basis for wetland ecological management. However, efficient methods for rapidly and accurately obtaining information on dominant vegetation remain limited. Therefore, taking the dominant herbaceous communities in the water-level-fluctuation zone of Ertan Reservoir as the research object, this study systematically evaluated the performance of three mainstream deep learning models—U-Net, PSPNet, and DeepLabV3+—combined with different convolutional neural network backbones in the automatic identification of typical herbaceous community types in the water-level-fluctuation zone, and compared them with a feature engineering-based stacking ensemble machine learning method. The results showed that: (1) the U-Net–ResNet50 combination achieved the best overall performance, with an overall accuracy of 0.918 and the highest F1 score, and performed particularly well in identifying dominant herbaceous community types such as the Cynodon dactylon community (F1 = 0.941) and the Abutilon theophrasti community. Its encoder–decoder structure effectively fused multi-level semantic information and spatial details, thereby enhancing recognition stability in complex vegetation scenes. (2) PSPNet and DeepLabV3+ exhibited relatively large fluctuations in validation loss in some combinations. For example, PSPNet–ResNet50 showed pronounced loss fluctuations during the early training stage, which may be related to the sensitivity of its multi-scale module to network depth. Nevertheless, both models showed higher training efficiency and shorter inference time, making them suitable for large-scale vegetation monitoring. (3) Compared with traditional methods, deep learning showed comparable performance in identifying dominant herbaceous community types. For example, the stacking ensemble model achieved an overall accuracy of 0.931, which was close to that of U-Net–ResNet50. However, for the “other herbaceous communities” category with relatively few samples, the F1 scores of traditional methods were generally higher than those of deep learning models; even the lowest value among the traditional methods (0.672) was still higher than the highest value among the deep learning models (0.539). Despite this, deep learning has the advantage of automatic feature extraction, significantly reducing dependence on manual feature engineering and showing strong potential for cross-regional transfer. Overall, deep learning models have promising application prospects for herbaceous community classification in reservoir water-level-fluctuation zones. Future studies should further expand sample size, integrate multi-source remote sensing data, and incorporate transfer learning methods to build an intelligent recognition framework that synergistically optimizes “data–model–application”, thereby supporting refined monitoring and management of wetland ecosystems.