<![CDATA[This study addressed leaf segmentation in open nursery environments for Eucalyptus pellita seedlings, where fluctuating illumination, cluttered backgrounds, and overlapping foliage had hindered reliable monitoring at operational scale. We proposed a Modified U-Net that integrated a ResNet-50 encoder for high-resolution feature extraction, L2 regularization in the decoder to improve generalization, and a composite binary cross-entropy plus Dice loss to balance pixel-level accuracy with shape conformity. We assembled 2,424 RGB images from an operational nursery and evaluated three architectures (Modified U-Net as the primary model, SegNet, and DeepLabv3+) under cloudy, sunny, and scorching illumination. We conducted inference at native resolution and summarized per-image metrics using medians with interquartile ranges, followed by nonparametric significance testing. The Modified U-Net consistently outperformed the baselines across all scenarios, achieving median Dice coefficients of 0.872 (cloudy), 0.841 (sunny), and 0.854 (scorching), with corresponding Intersection over Union values of 0.773, 0.725, and 0.745. A Kruskal-Wallis test on per-image Dice and Intersection over Union yielded no significant differences across lighting conditions (H = 4.012, p = 0.1345), indicating stable performance under natural illumination variability. Qualitative overlays revealed localized errors, including glare-induced false positives in sunny scenes and shadow-related artifacts under scorching light, which did not materially shift global overlap distributions. We concluded that the proposed architecture delivered robust, high-fidelity segmentation in realistic nursery conditions and provided a practical basis for field deployment, with further gains expected from glare- and shadow-aware augmentation and lightweight optimization for near real-time inference on edge devices.]]>
