Enhancing Cross‐Crop Plant Disease Detection on Complex Field Images Through Dataset Merging, Transfer Learning and Ensemble Averaging of Convolutional Neural Networks

Abstract

ABSTRACT Agricultural productivity is increasingly threatened by limited arable land, population growth and climate change, emphasising the need for effective crop protection. Deep learning, particularly convolutional neural networks (CNNs), has shown promise for automatic plant disease detection. However, performance is often limited by insufficient and non‐diverse training data. This study addresses this challenge by reorganising and merging two datasets—PlantVillage (PV) and EdenLibrary (EL)—based on disease common names rather than crop‐disease pairs to improve generalisation. PV provides a large volume of lab images with limited variability, whereas EL offers fewer but more diverse field images. Three lightweight CNNs, MobileNet‐v3, ShuffleNet‐v2 and SqueezeNet, were fine‐tuned via transfer learning on the reorganised datasets. Results show that disease‐common‐name‐based labelling enhances generalisation to unseen plant species while maintaining performance on known species. Moreover, combining PV and EL improved cross‐domain generalisation on the PlantDoc dataset. Ensemble averaging of the three models further boosted performance, with the ternary ensemble outperforming individual and binary models. The ensemble achieved accuracies of 99.93% and 94.46% on seen and unseen species from the PV test set, respectively, and 98.81% and 91.89% on seen and unseen species from the EL dataset. In the cross‐dataset evaluation, it reached 67.45% accuracy on the PlantDoc dataset, outperforming individual models by up to 13.09%. These findings highlight the value of dataset integration and ensemble techniques in building robust plant disease detection systems adaptable to diverse real‐world scenarios, without requiring additional in‐field data.