The semantic segmentation of UAV remote sensing imagery is of great importance for scene understanding and decision-making. However, existing state-of-the-art models incur high computational costs, making them unsuitable for real-time deployment on UAV platforms.To address this issue, this paper proposes an improved lightweight framework, HWD-Deeplab v3+, in which the backbone is replaced with a modified MobileNetV2 to reduce the number of parameters and FLOPs. A Haar Wavelet-based Downsampling (HWD) module is integrated into MobileNetV2 to enhance edge preservation and mitigate information loss. In addition, the Atrous Spatial Pyramid Pooling (ASPP) module of DeepLab v3+ is replaced with a Pyramid Pooling Module (PPM) to further decrease computational overhead and improve multi-scale feature fusion.Experimental evaluations on the UDD6 dataset demonstrate that the proposed HWD-Deeplab v3+ achieves an mIoU of 76.12%, with only 7.3M parameters, 52.0G FLOPs, and an inference speed of 31.9 FPS on an RTX 2070 SUPER GPU. Compared with the original DeepLab v3+ (42M parameters, 144G FLOPs, and 8.5 FPS), this represents a reduction of 82.6% in parameters and 63.9% in FLOPs, while improving segmentation accuracy by 3.56% mIoU, 1.90% MPA, and 2.53% overall accuracy. These results confirm that the proposed model not only achieves a superior trade-off between accuracy and efficiency but also exhibits real-time inference capability, providing a practical and effective solution for UAV-based semantic segmentation tasks.