Identifying tree species from bark photographs matters for forest inventory and the wood-processing industry, yet the design choices behind deep-learning approaches have not been studied systematically, and some reported results may be inflated by images of the same trunk appearing in both training and test sets. We present a systematic ablation on BarkNet 1.0 (20 species) under a tree-level, leakage-safe five-fold cross-validation, using ConvNeXt-V2 backbones with attention-based multiple instance learning (AMIL). We vary patch size, aggregation strategy, training regime, backbone capacity, the whole-image baseline, and test-time bag size, and we measure leakage directly with a matched image-level split. Per-patch accuracy rises with patch size (75.2% at 160 px to 89.3% at 512 px), and image-level accuracy plateaus near 94.9% at 512 px. AMIL exceeds majority voting by at most 0.7 pp; soft voting and mean pooling do no better than hard voting, so the gain comes from the learned attention weights and vanishes at 512 px, where attention is near-uniform. Backbone capacity changes accuracy by less than a point, and two-stage training gives no advantage. Whole-image classification (91.9%) trails patch aggregation, and accuracy saturates around 64 patches per image. Under an image-level split, the same configuration reaches 99.9% (+5.0 pp) and the whole-image baseline 99.2% (+7.3 pp), an inflation as large as the entire effect of patch size that also erases the differences between design choices. We recommend patches as large as possible while leaving enough to aggregate (about 512 px here), a nano backbone, and tree-level splits for bark benchmarks.
翻译:暂无翻译