Abstract: Accurate and efficient wheat seedling monitoring is a critical prerequisite for precision variable-rate fertilization and refined field management, directly affecting crop yield formation and agricultural production efficiency. Traditional manual field survey methods suffer from low efficiency, limited spatial coverage and high labor costs, while conventional visible-light deep learning models face persistent challenges including severe spectral confusion between seedlings and stubble or weeds in low-coverage seedling scenarios, blurred boundary segmentation under leaf adhesion conditions, and insufficient utilization of vegetation-sensitive spectral bands. To address these issues, this study proposes the Multi-Spectral Hybrid Network (MS-HybridNet), a semantic segmentation framework integrating an improved Swin Transformer with unmanned aerial vehicle (UAV) multispectral imagery, to achieve high-precision wheat seedling identification under complex field conditions. The proposed model takes Swin Transformer as the baseline architecture and is adapted to 5-channel UAV multispectral image input covering blue, green, red, red edge and near-infrared bands. It constructs a visible-red edge-near infrared band synergy mechanism, leveraging the chlorophyll sensitivity of the 730 nm red edge band and the crop-soil differentiation advantage of the 840 nm near-infrared band to enhance spectral discrimination between wheat seedlings and complex backgrounds including stubble, weeds and bare soil. A CNN-Transformer hybrid encoder is designed to replace the original pure Transformer encoding structure, combining the local edge detail extraction capability of convolutional layers and the global spatial distribution modeling advantage of self-attention mechanisms to fully preserve crop morphological features and alleviate segmentation confusion in leaf adhesion areas. An Atrous Spatial Pyramid Pooling module is introduced at the bottleneck layer to expand the receptive field through multi-scale dilated convolutions, strengthening feature adaptability to seedlings of different sizes, and a decoder with deep supervision is adopted to supplement multi-scale auxiliary supervision signals to enhance the boundary segmentation accuracy of small-target seedlings and adherent leaves. Experimental results on the self-built wheat seedling multispectral dataset demonstrate that MS-HybridNet achieves 84.1% mean pixel accuracy (mPA), 73.4% recall, 76.0% F1-Score, 80.5% mean Intersection over Union (mIoU), and 71.0% Boundary IoU (bIoU) on the test set. Compared with the baseline Swin Transformer, mPA, recall, F1-Score, mIoU and bIoU are improved by 5.0, 2.3, 2.1, 6.5 and 5.9 percentage points respectively. Compared with Mask2Former with the closest segmentation accuracy, mIoU increases by 5.2 percentage points, with 39.4% fewer parameters and 28.9% faster inference speed at only 18.2 ms per image. Compared with ResNet-UNet and DeepLabV3+, mIoU is improved by 14.2 and 18.4 percentage points, and bIoU rises by 5.7 and 10.4 percentage points respectively. Spectral ablation experiments further verify that the three-band synergistic input improves mIoU and bIoU by 18.2 and 10.3 percentage points respectively compared with single RGB input. The proposed method maintains controllable computational overhead while significantly improving segmentation accuracy, fully meeting the real-time processing requirements of UAV aerial survey data and providing reliable technical support for variable-rate fertilization and refined field management in precision agriculture.