<p>Water body extraction in remote sensing is critical for environmental monitoring, flood control, and urban planning. However, the complex semantic information and varied morphology in remote sensing images present challenges for accurate extraction. Traditional semantic segmentation algorithms, with limited receptive fields, often fail to capture long-range dependencies, leading to misclassification and omissions. Additionally, most current methods rely on supervised learning, limiting generalization due to insufficient use of unlabeled data. This paper proposes ResDDSCUNet++, an improved UNet++ architecture for remote sensing water body extraction. The model replaces standard convolution and pooling layers with ConvNextV2 blocks, enhancing the capture of long-range dependencies. It also introduces the Attention Modulation Module (AMM) to focus on critical information and a residual depthwise separable double convolution module (ResDoubleDSC) to reduce model parameters and computational load. Furthermore, a hybrid training approach combining supervised and self-supervised learning is employed, leveraging unlabeled data for pretraining and fine-tuning on labeled data to improve generalization. Experimental results demonstrate that the proposed method outperforms other segmentation algorithms, improving the Dice coefficient by 1.6% and the IoU coefficient by 1.22%.</p>