Deblurring-aware semantic segmentation of crops and weeds in UAV sorghum imagery via a UNet-ResNet architecture

Automated monitoring of weed infestation is essential for precision agriculture, where reliable crop-weed discrimination underpins site-specific intervention. This study presents a semantic segmentation framework based on UNet with a ResNet encoder backbone that delineates three semantic categories-background, crop (sorghum), and weed-within a single inference pass. A systematic comparison across four ResNet encoder variants (ResNet-18, ResNet-34, ResNet-50, and ResNet-101) reveals that ResNet-34 achieves the highest mean Dice Score of 0.9198 and that accuracy does not increase with depth-the deeper ResNet-50 and ResNet-101 score slightly lower, consistent with capacity saturation and mild overfitting on the limited training set. To address image quality degradation prevalent in UAV-based agricultural imaging, an explicit deblurring module based on NAFNet is integrated prior to segmentation through a restoration-aware training protocol, improving the mean Dice Score on the motion-blurred test subset by 27.7% in relative terms over the strongest baseline without restoration and yielding the best combined-set accuracy (0.8223 mDS). Comprehensive ablation experiments identify decoder depth (full-resolution progressive upsampling) and encoder-decoder skip connections as the two most critical components, whose removal lowers the mean Dice Score by 5.18 and 3.51 percentage points respectively, whereas ImageNet pre-training and dropout have negligible effect at this dataset scale. A complexity analysis based on parameter count and multiply-accumulate operations (MACs) shows that the segmentation network (24.4 M parameters, 1.97 GMACs) is comparable to standard CNN segmentation baselines, suggesting the plausibility of deployment on resource-constrained UAV platforms; a direct on-device throughput validation on embedded NVIDIA Jetson hardware was not performed and is left for future work.