Featured Application
This work quantifies how much classification performance peach leaf damage models trained on public images lose when applied in a commercial orchard, with photograph-level confidence intervals and under the explicit assumption that classes absent from the orchard are masked, and how much of that loss a few tens of labelled orchard leaves recover. It gives growers and technical advisers a realistic expectation of model behaviour in the field and an estimate of the annotation effort needed before deployment.
Abstract
Leaf damage classifiers are usually trained and evaluated on public images collected elsewhere, and their behaviour in a new orchard is rarely measured. We quantified this gap for peach leaves. Eleven ImageNet-pretrained convolutional backbones, each with and without a Convolutional Block Attention Module (CBAM), were trained on 1415 public leaves (six classes). They were evaluated, with photograph-grouped cross-validation and confidence intervals, on 182 leaves from 122 photographs taken in a commercial orchard (four classes). On the public hold-out set, macro-F1 ranged from 0.670 to 0.772, and the ranking of the best models was not resolved; CBAM brought no improvement (pooled change to ). Without retraining, mean orchard macro-F1 was 0.634, against 0.719 on the same four public classes, but only after masking the two classes absent from the orchard; without masking, one-third of the leaves were assigned to them. Label-free normalisation-statistics alignment did not help, whereas fine-tuning raised the mean macro-F1 to 0.72, and a few tens of labelled leaves sufficed to match zero-shot transfer. On the shared public classes, full fine-tuning lost more source performance than linear probing, and joint training preserved it. Light models and a modest site-specific annotation effort are sufficient for this task.