What Improves UAV RGB Segmentation of Maize and Weeds? An Ablation Study of U-Net Components

Abstract

The use of RGB images acquired by unmanned aerial vehicles (UAVs) is increasingly common in precision agriculture, for monitoring crops, weeds, and field conditions. However, large variations in plant size, illumination, shadows, and background appearance make automatic segmentation difficult. A major challenge is discriminating crop plants from weeds that may compete for resources and identifying conditions that may favor the development of plant diseases. In this study, we evaluated U-Net-based semantic segmentation models on 142 UAV RGB images from two maize fields using grouped 5-fold cross-validation to segment maize, weeds, and bare soil. The full baseline and seven one-component ablations were each evaluated with three output modes, yielding 24 run-level configurations and 120 fold-level evaluations. Across the five folds, no_deep_supervision with pure_dl achieved the highest mean F1 (0.7680 ± 0.0250), mIoU_no_bg (0.6436 ± 0.0267), aggregate score ($\mathbf{0 . 9 5 9 3} \pm \mathbf{0 . 0 2 4 5}$), and weed IoU ($\mathbf{0 . 4 6 5 7} \pm \mathbf{0 . 0 5 3 1}$), while also yielding the lowest mean HD95 ($\mathbf{6 0 . 5 5 8 8} \boldsymbol{\pm} \mathbf{1 2 . 6 2 1 0}$).ExG-guided refinement remained close to pure_dl but did not improve mean performance consistently, whereas hardthreshold reconstruction sharply degraded both overlap and distance metrics. The highest individual fold-level weed IoU was 0.5222 for no_sigreg with pure_dl in fold 2; however, no_deep_supervision with pure_dl remained the best configuration when performance was averaged across all folds. Annotation examples also showed that isolated weed-label outliers near image borders can disproportionately increase distancebased metrics even when the predicted mask is visually plausible. Overall, the results favor a simpler pipeline for this dataset, direct multiclass inference, and cautious interpretation of weed HD95. The study does not propose a new architecture; instead, it provides a controlled ablation protocol and practical evidence that additional U-Net components should be validated rather than assumed beneficial in small-data UAV RGB segmentation.