Robust Plant Disease Recognition Using a Neutrosophic-Enhanced, RBF-Based Stacked Ensemble of ConvNeXt and Classical CNN Models
Accurate and timely recognition of plant diseases is crucial to prevent crop loss and ensure global food security. This paper presents a robust ensemble-based framework that combines six classical and state-of-the-art deep convolutional neural networks (DCNNs), including a ConvNeXt architecture, and integrates Neutrosophic Science to better handle uncertainty in leaf images. The proposed approach features three main components: (1) transfer learning with pre-trained DCNNs, (2) a model-averaging strategy to stabilise individual predictions, and (3) a stacked ensemble design that employs a radial basis function (RBF) meta-learner to refine the classification outputs. Experiments on the Plant Village dataset, comprising 54,305 segmented images of 38 plant diseases, included 10-fold cross-validation. The results show that the final stacking ensemble achieved near-perfect performance with 99.97% accuracy and an F1 score of 99.55% on an unseen test set of 27,160 images. Compared with standalone models, the ensemble demonstrated greater robustness in distinguishing visually similar diseases, benefiting from the complementary strengths of multiple DCNN architectures. The Neutrosophic component further enhances reliability by modelling uncertainties due to noise, occlusions, and environmental variations. Although a higher computational overhead and modest misclassifications remain, especially in certain visually overlapping classes, this study demonstrates the effectiveness of an ensembledriven, uncertainty-aware strategy. These findings hold considerable promise for real-world agricultural applications, where rapid and accurate disease diagnosis is paramount.
Volume & Issue
Vol. Volume 26 / Iss. Issue 2