Benchmarking deep vision architectures for fruit-tree sapling bark classification on SapBark-64

Accurate identification of fruit-tree saplings is important for nursery management and varietal authentication, but related species and cultivars can be visually similar at early growth stages. This study benchmarks seven ImageNet-pretrained deep vision architectures for bark-based classification on the public SapBark-64 dataset: ResNet18, ResNet50, DenseNet121, EfficientNet-B0, MobileNetV3-Large, ConvNeXt-Tiny and Swin-Tiny. All models were evaluated using a unified stratified 10-fold cross-validation protocol with identical preprocessing, training, model-selection and evaluation procedures. ConvNeXt-Tiny achieved the highest accuracy (0.9527 ± 0.0087) and macro-F1 (0.9499 ± 0.0084), significantly outperforming all other models after multiple-comparison correction. Swin-Tiny achieved 0.9188 ± 0.0136 accuracy and 0.9146 ± 0.0147 macro-F1. ImageNet-pretrained full fine-tuning performed best in the three-fold diagnostic analysis of DenseNet121 and EfficientNet-B0. ConvNeXt-Tiny was the most stable of the three models evaluated under synthetic perturbations, although blur and especially Gaussian noise caused substantial degradation. Grad-CAM provided qualitative evidence of attention to trunk and bark-texture regions. Because no external validation was performed and SapBark-64 lacks sapling and capture-session identifiers, possible identity-level leakage may make image-level estimates optimistic. Results should therefore be interpreted as a reproducible within-dataset benchmark rather than evidence of generalization across nurseries, devices or environments.