<p>In the low-altitude domain, the UAV (Unmanned Aerial Vehicle) aerial image is dominated by small targets and the background is complex, which makes segmentation difficult. Segment anything model (SAM) has garnered significant attention for segmentation tasks due to its zero-shot segmentation capability. Despite the outstanding performance of SAM on natural images, it still faces significant performance degradation and limited generalization when dealing with UAV images, especially those involving small targets and blurred boundaries. Therefore, we propose UTS-SAM, a novel SAM-based framework that implements a UAV target segmentation method based on complementary dual encoder. This approach not only improves the segmentation accuracy of SAM in the UAV domain but also pursues lower computational cost. Firstly, we propose a dual-branch architecture comprising a lightweight image encoder (Vision Transformer (ViT) branch) to parallel a CNN (Convolutional Neural Network) branch. By introducing multi-level feature fusion block and the cross-branch attention convolution module, we enhance small-scale feature representation and compensate for the limitations in local information extraction. Secondly, a feature adaptor is referenced to improve the adaptability of the SAM in the UAV image domain. To reduce GPU consumption, we keep the parameters of both the ViT branch and the pretrained SAM while exclusively training the CNN branch and its connected modules. Finally, the issue of unclear target boundaries is mitigated through the use of a comprehensive loss function. Extensive comparative experiments have shown that UTS-SAM significantly outperforms existing methods in terms of balance efficiency and performance, highlighting the potential of SAM for specific downstream tasks in the low-altitude domain of the UAV.</p>