SYMCON-VIT: A SYMPTOM-CONSISTENCY VISION TRANSFORMER FOR CONTEXT-ROBUST APPLE LEAF DISEASE RECOGNITION

There is a method to the difficulty of recognising disease in an apple leaf that raw accuracy does not tell. The lesions are often tiny and in no particular pattern, one disease is easily confused with another, and a photograph will be full of soil, sky, shadow and other foliage of no diagnostic value. Put a network to work on a well put together benchmark and it will make good scores by virtue of that context, but out in the orchard it is another matter. To put some distance between the two we have SymCon-ViT, our Vision Transformer. It has a Symptom Relevance Estimator to put a score on the evidence for disease in each patch token; a sparse top-k Symptom Token Pool to focus its capacity where the scores are highest; and Symptom Relational Attention so those tokens can put the whole leaf to the query. We also have a cross-view consistency objective that will penalise a change in the prediction or relevance map if the background is put in its place with something else. Our evaluation was done on 1,943 images from a test set we have locked down, with five seeds apiece for each of the seven baselines (CNNs such as ResNet-50, EfficientNet-B2 and ConvNeXt-T, and Transformers including ViT-S/16, DeiT-III-S, Swin-T and MaxViT-T) and all comparisons made at run and image level. In-domain macro-F1 for SymCon-ViT is 0.980, a marginal +0.006 (95% CI [0.003, 0.010]) on the best baseline. Where the gains are to be found is in the Plant Pathology 2020 external data: 0.882 compared to MaxViT-T’s 0.820, which narrows the domain gap to 0.097 from 0.154 while 90.1% of in-domain performance is held. Mean corrupted macro-F1 comes in at 0.966, calibration error is 0.012 and open-set AUROC 0.940; predictions hold up to back-ground replacement 4.2 points more frequently. An ablation study reveals the two consistency terms at the representation level do little for accuracy but reduce context sensitivity by two thirds, making a case for judging symptom grounding on its stability rather than what the leaderboard says.