Abstract: Accurate estimation of seedling vegetation coverage for mechanically direct-seeded rapeseed under field conditions is a critical prerequisite for evaluating sowing quality and guiding precise management practices such as gap filling and thinning. However, existing methods remain inadequate due to the small target size, dense distribution, complex soil backgrounds, and high visual similarity between rapeseed seedlings and weeds at early growth stages. To address these challenges, this study proposes an improved semantic segmentation model named RCL-UNet (RCL-U-shaped Network) based on the classical U-Net architecture. A dedicated dataset was constructed from high-resolution RGB images acquired by a DJI Mavic 3 Pro UAV over typical mechanically direct-seeded rapeseed fields in Anhui Province, China. After image cropping, screening, and pixel-level annotation, a total of 3,140 valid image tiles were obtained, including 2,960 training tiles (after data augmentation), 120 validation tiles, and 60 test tiles. Three key improvements were introduced into the U-Net baseline: First, the original encoder was replaced with a ResNet50-PPM-S structure, which integrates a 50-layer deep residual network with a pyramid pooling module to extract multi-scale contextual features and enhance sensitivity to small targets. Second, a convolutional block attention module (CBAM) was embedded into the skip connections to adaptively recalibrate feature responses through the synergistic effect of channel and spatial attention, effectively suppressing background noise while preserving fine boundary details of seedlings. Third, the Dice loss function was adopted instead of cross-entropy loss to alleviate the severe class imbalance in small-target segmentation, thereby improving pixel-wise coverage of seedling regions. Ablation experiments confirmed that all three modules contributed positively to segmentation performance. The RCL-UNet achieved a mean Intersection over Union (mIoU) of 89.88%, a mean Pixel Accuracy (mPA) of 94.98%, a mean Precision (mP) of 93.96%, and a mean Recall (mR) of 93.92% on the test set, representing absolute improvements of 16.75, 15.33, 8.21, and 14.69 percentage points over the original U-Net, respectively. Compared with PSPNet and DeepLabv3+, RCL-UNet outperformed both with mIoU gains of 10.78 and 3.33 percentage points, respectively. Visual assessments under varying planting densities (300,000–600,000 plants·ha^-1) and imaging distances (0.8 m and 2.5 m) showed that RCL-UNet produced more complete and boundary-accurate segmentation masks with significantly fewer omissions and adhesions, especially in distant-view and high-density scenes. The improved model was deployed on a laptop equipped with an RTX 4060 GPU via a PyQt5 interface, achieving an average inference time of 49 ms per image. Field validation was conducted at an independent site under two sowing patterns: mechanical drill sowing and aerial broadcasting, with 100 sampling units of 1 m × 1 m established for each pattern. Vegetation coverage estimation results showed that under mechanical drill sowing, the root mean square error (RMSE) between model predictions and manual ground truth was 0.42% with a coefficient of determination (R²) of 0.98; under aerial broadcasting, the RMSE was 0.50% and R² was 0.94. Across both sowing patterns, the RMSE values were consistently below 0.5% and the R² values remained above 0.94, demonstrating that the improved model achieved high pixel-wise segmentation accuracy and exhibited robust generalization capabilities when deployed in complex and variable field environments. The proposed method can provide reliable technical support for establishing a scientific evaluation system for sowing performance of mechanically direct-seeded rapeseed. Nevertheless, the current model has only been validated at the four-leaf stage and has not yet covered the full growth cycle where canopy overlap becomes more pronounced. Moreover, the slightly lower prediction accuracy under aerial broadcasting compared with mechanical drill sowing indicates that further adaptation to irregular sowing patterns is needed. Future work will extend the dataset to multiple growth stages, incorporate multi-source imagery, and explore lightweight architectures for edge deployment to further enhance model generalization and operational efficiency.