Estimation of the coverage for mechanically direct sowing rapeseed seedlings using RCL-UNet

Mechanical direct seeding is one of the most important steps during rapeseed cultivation. Sowing quality can also be evaluated for precise practices, such as gap filling and thinning. It is often required to accurately estimate the coverage of rapeseed seedlings under field conditions. However, existing estimation remains a great challenge due to the small target size, dense distribution, complex soil backgrounds, and high visual similarity between rapeseed seedlings and weeds at early growth stages. This study aims to propose an improved semantic segmentation model, named RCL UNet (RCL U-shaped Network), using the classical U-Net architecture. A dataset was constructed during mechanical direct seeding. High-resolution RGB images were acquired by a DJI Mavic 3 Pro UAV. Rapeseed fields were also taken from Anhui Province, China. A total of 3 140 image tiles were obtained after image cropping, screening, and pixel-level annotation, including 2 960 training tiles (after data augmentation), 120 validation tiles, and 60 test tiles. Three improvements were introduced into the U-Net baseline: 1) The original encoder was replaced with a ResNet50 PPM S structure. A 50-layer deep residual network was integrated with a pyramid pooling module to extract multi-scale contextual features for high sensitivity to small targets. 2) A convolutional block attention module (CBAM) was embedded into the skip connections to adaptively recalibrate feature responses using channel and spatial attention, effectively suppressing background noise for the fine boundary features of seedlings. 3) The Dice loss function was adopted instead of cross-entropy loss to alleviate the severe class imbalance in small target segmentation, particularly for pixel-wise coverage of seedling regions. Ablation experiments confirmed that all three modules contributed positively to segmentation performance. The results showed that 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, indicating absolute improvements of 16.75, 15.33, 8.21, and 14.69 percentage points over the original UNet, respectively. Furthermore, RCL UNet outperformed both PSPNet and DeepLabv3+, with mIoU gains of 10.78 and 3.33 percentage points, respectively. Visual assessments were also conducted under varying planting densities (300 000-600 000 plants/hm²) and imaging distances (0.8 and 2.5 m). Segmentation masks were obtained with high boundary precision, 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, with 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. 100 sampling units of 1 m × 1 m were selected for each pattern. coverage estimation showed that under mechanical drill sowing, the root mean square error (RMSE) between predictions and ground truth data was 0.42% with a coefficient of determination (R²) of 0.98; The RMSE and (R²) were 0.50% and 0.94, respectively, under aerial broadcasting. The RMSE values were below 0.5% under sowing patterns, whereas the (R²) values remained above 0.94. The high accuracy of pixel-wise segmentation was achieved for robust generalization under complex and variable field environments. Reliable technical support can help evaluate the sowing performance of mechanical direct seeding. Nevertheless, the current model has only been validated at the four-leaf stage. It is also required to cover the full growth cycle with canopy overlap. Moreover, the higher prediction accuracy can be expected for the irregular sowing patterns under aerial broadcasting, compared with mechanical drill sowing. The dataset can also be extended into multiple growth stages and source imagery for the edge deployment of lightweight architectures. These findings can provide a strong reference to further enhance model generalization and operational efficiency.