Architecture-oriented benchmarking and weighted ensemble learning for multi-class plant disease classification
Array · 2026-12 · 32 · 101348
DOI: 10.1016/j.array.2026.101348 ↗Kurz gefasst
Fachartikel in der Zeitschrift Array, 2026-12. Verfasst von Md Tarek Hasan und 6 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Accurate identification of plant diseases from leaf images is a crucial issue in enhancing agrarian productivity and confirming food security. Recent advances in deep learning have demonstrated strong performance for automated plant disease classification; however, most prevailing studies depend on a single model architecture or provide comparative benchmarks without introducing a unified learning mechanism. In this paper, we present a comprehensive evaluation of 16 state-of-the-art deep learning models, including hybrid architectures, convolutional neural networks (CNNs), vision transformers (ViTs), and mobile-optimized networks, using an augmented PlantVillage dataset comprising 38 disease and healthy classes. To address the limitations of single-model reliance, we further propose an Architecture-Oriented Weighted Ensemble (AOWE) framework that integrates complementary model predictions using validation-guided weighting. Under the commonly used image-level protocol, fine-tuned models reach 99.6%–99.9% accuracy. We show that this protocol is optimistic: the augmented release and the original PlantVillage photographs contain near-duplicates of the same leaf, and keeping all photographs of a leaf in the same fold lowers the accuracy of a frozen-backbone linear probe from 97.17% to 91.62%, and that of a memorising 1-NN classifier from 94.48% to 84.24% (MobileNetV3-Large features). We therefore re-evaluate all conclusions with a leakage-controlled protocol that uses leaf-group-aware, nested and repeated cross-validation, in which ensemble weights are estimated only from out-of-fold predictions. We compare AOWE against eight alternative combination rules using paired statistical tests and confidence intervals, profile the computational cost of all 16 architectures and of the ensemble, and validate externally on the PlantDoc field dataset. On unseen PlantVillage leaves, AOWE reaches 95.54% accuracy versus 92.71% for the best single model (CoaT-Lite-S), a statistically significant improvement over every individual member; on PlantDoc field images, AOWE reaches 65.61% accuracy versus 60.15% for the best single model (CoaT-Lite-S), a statistically significant improvement over every individual member. Among the combination rules, AOWE is statistically indistinguishable from equal-weight soft voting, and it significantly outperforms pseudo-BMA, optimised weights on PlantVillage and pseudo-BMA, stacking on PlantDoc. Models trained on laboratory images transfer poorly to field images (10.6%–23.4% accuracy on PlantDoc). Laboratory benchmark accuracy should therefore not be equated with readiness for field deployment.
Zitieren
APA
Hasan, M. T., Rayhan, F., Samiha, M. T., Bashir, M., Hossain, M. S., Ahmmed, S., & Ohidujjaman (2026). Architecture-oriented benchmarking and weighted ensemble learning for multi-class plant disease classification. Array, 32, 101348. https://doi.org/10.1016/j.array.2026.101348
Vancouver
Hasan MT, Rayhan F, Samiha MT, Bashir M, Hossain MS, Ahmmed S, et al. Architecture-oriented benchmarking and weighted ensemble learning for multi-class plant disease classification. Array. 2026;32:101348. doi:10.1016/j.array.2026.101348
BibTeX
@article{Hasan2026,
author = {Hasan, Md Tarek and Rayhan, Faiyaz and Samiha, Myisha Tun and Bashir, Minhajul and Hossain, Md Sakhawat and Ahmmed, Suman and Ohidujjaman},
title = {Architecture-oriented benchmarking and weighted ensemble learning for multi-class plant disease classification},
journal = {Array},
year = {2026},
volume = {32},
pages = {101348},
doi = {10.1016/j.array.2026.101348}
}
EUVIMED