Aiming at the problem of insufficient classification accuracy of traditional methods in complex wetland environments, this paper proposes an improved DeeplabV3+ model (Ct-Deeplabv3+) integrating a comparative learning encoder and a dual attention mechanism: a clustering-based comparative learning mechanism is designed at the encoder side to optimize the feature representation in an unsupervised manner, and an improved inflated convolution strategy is combined to expand the sensory field. The decoder side introduces the dual modules of gated context transformation and channel-coordinate attention to strengthen the ability of multi-scale feature fusion and spectral weight assignment. The results of the ablation experiments show that the comparative learning encoder and the dual attention mechanism module can effectively improve the benchmark model accuracy. The Ct-Deeplabv3+ model is applied to the task of semantic segmentation of high-resolution multispectral UAV remote sensing images, and the experimental results show that the accuracy of Ct-Deeplabv3+ on the test set reaches 93.5%, and the recall and Kappa coefficients are 88.4% and 0.9318, which is better than that of the mainstream model. This paper provides a new technical model for high-precision dynamic monitoring of wetland vegetation, which has important application value for ecological restoration and sustainable development.