Multi-Crop Disease Pretraining for Black Pepper Leaf Disease Classification: A Leaf-Grouped Evaluation of CNN Encoders

Abstract

Published work on black pepper (Piper nigrum L.) leaf disease reports accuracies close to 100%, yet the public image sets behind such figures are small and full of repeated photographs of the same leaf. We asked a narrower question that matters more to a grower with few labelled images: does a convolutional network trained on many other crops transfer to black pepper better than an ordinary ImageNet network? Two public datasets were used (1,500 and 819 images). SIFT matching with RANSAC verification traced them to about 590 and 121 distinct leaves, and every evaluation kept all photographs of a leaf on one side of the split. We compared two frozen EfficientNet-B3 encoders with identical architecture: the ImageNet initialisation and a multi-task model trained on 38 PlantVillage classes from 14 crops. With all training leaves available, both reached 97% to 99% accuracy and did not differ. With one leaf per class, the multi-crop encoder raised macro F1 from 0.51 to 0.73 on the first dataset (p < 0.001, 30 paired draws) and from 0.65 to 0.71 on the second, where the gain was not consistent across training sizes. Used without any black pepper training, its source classifier necessarily named a wrong crop, yet did so with a mean confidence of 0.95. Control inputs showed that image corners with no leaf pixels predicted the disease class of the first dataset with 87.1% accuracy. Error level analysis, a JPEG forensics transform sometimes used as CNN preprocessing, lowered leaf-grouped accuracy on both datasets. Multi-crop pretraining is worth using when labels are scarce, but accuracy on current black pepper benchmarks says little about field performance until acquisition shortcuts are removed.