In a study published in Neural Computing and Applications, Miguel Franco-Hernández, Vladimir Mejía-Domínguez, and Yakdiel Rodriguez-Gallo, affiliated with Don Bosco University and the University of El Salvador, systematically benchmarked vision transformers against convolutional neural networks for sugarcane leaf disease classification. Their central contribution goes beyond simply declaring a winner: the team first consolidated three publicly available image datasets into a single corpus of 11,120 leaf photographs, then subjected a field of classic and lightweight modern architectures to controlled, identical training conditions. The MobileViT-v2-100 model, a member of a family designed for mobile-scale computing, emerged on top with an accuracy of 96.45 percent.

The significance of the dataset consolidation cannot be overstated. Most published studies on crop disease detection train and test their models on a single curated dataset, which flatters performance figures because the images tend to share lighting conditions, backgrounds, camera angles, and imaging equipment. By merging the Sugarcane Leaf Image Dataset, the Sugarcane Disease Dataset, and the Sugarcane Leaf Disease Dataset, all hosted on Mendeley Data, the researchers created a more heterogeneous collection that better reflects the messy variability of real-world field photography. They then applied systematic preprocessing and data augmentation to standardize inputs and expand effective training volume, a step that is particularly important when class boundaries, such as the visual differences between early-stage rust and mosaic symptoms, are subtle.

Architecturally, the study pits two competing philosophies of computer vision against each other. Convolutional neural networks, the workhorses of image recognition for the past decade, build understanding of an image through layers of local filters that detect edges, textures, and progressively more abstract patterns. The benchmark included venerable architectures such as VGG, ResNet, and DenseNet, alongside efficiency-focused designs like MobileNetV3, EfficientNet, and ConvNeXt, the latter a modernized convolutional network that incorporates design lessons from transformers. Vision transformers, by contrast, treat an image as a sequence of patches and use self-attention mechanisms, borrowed from natural language processing, to model relationships between any parts of the image regardless of their distance. This global receptive field is theoretically powerful for diseases whose symptoms span large or irregular regions of a leaf.

MobileViT-v2 occupies a middle ground between these philosophies. Originally introduced by Mehta and Rastegari as a light-weight, general-purpose, mobile-friendly vision transformer, the architecture wraps transformer blocks inside a convolutional framework, allowing it to capture both local detail and global context while remaining small enough to run on smartphones and embedded devices. The version-two refinement replaced costly matrix-based attention with a separable self-attention mechanism that computes attention using simple element-wise operations, dramatically reducing computational overhead. In the sugarcane benchmark, this efficiency did not come at the cost of accuracy; the MobileViT-v2-100 variant outperformed both heavier transformers and the full range of convolutional competitors, a result with immediate practical implications for precision agriculture tools that must operate in rural areas with limited connectivity and computing power.

Perhaps the most methodologically important element of the study is its treatment of generalization. The researchers employed a Leave-One-Dataset-Out cross-validation protocol, training the models on two of the three source datasets and testing on the held-out third, rotating through all combinations. This is a far more honest test of whether a model has learned disease features or merely dataset-specific quirks. Recent work in machine learning evaluation has increasingly criticized standard random splits for inflating reported accuracy when images from the same source appear in both training and test sets. By quantifying how performance degrades when the model confronts an entirely unseen data source, the LODO protocol provides a more realistic preview of how such systems would behave when deployed on a farm whose imaging conditions differ from anything in the training corpus.

Interpretability received equal attention, addressing one of the most persistent objections to deploying deep learning in high-stakes agricultural decisions. The team applied Grad-CAM, a technique that generates heatmaps highlighting which regions of an image most influenced the model’s prediction, allowing agronomists to verify that the network is attending to actual lesions and discoloration patterns rather than irrelevant background features. Complementing this, t-SNE and UMAP, two nonlinear dimensionality-reduction methods, were used to project the model’s internal feature representations into two-dimensional visualizations. When healthy and diseased leaves form cleanly separated clusters in these projections, it indicates the network has learned meaningful, discriminative features rather than exploiting spurious correlations. Together, these tools transform the model from an opaque oracle into something closer to an auditable diagnostic assistant.

The disease landscape confronting sugarcane growers justifies this level of engineering effort. Sugarcane smut, caused by the fungus Sporisorium scitamineum, can devastate entire plantings and is the subject of ongoing global surveillance. Sugarcane yellow leaf virus stealthily reduces yields across producing regions, mosaic disease complex spreads through insect vectors and infected planting material, and pokkah boeng, brown spot, brown rust, orange rust, and grassy shoot disease each present distinct diagnostic challenges. Because many of these pathogens produce overlapping foliar symptoms in their early stages, and because misdiagnosis leads to wasted fungicide applications or delayed roguing of infected stools, an accurate automated classifier offers genuine economic and environmental value. Earlier deep learning studies on sugarcane disease detection, including CNN-based classifiers and ensemble approaches, demonstrated feasibility but typically rested on single-dataset evaluations; the present benchmark raises the evidentiary bar for the entire subfield.

The researchers, who received institutional support from Universidad Don Bosco and Universidad de El Salvador without external funding, have also made their work unusually reproducible. The consolidated preprocessing pipeline, training scripts, and visualization tools for Grad-CAM, t-SNE, and UMAP are publicly available on GitHub and archived with a permanent identifier on Zenodo, while all three source datasets remain freely accessible. This transparency allows other groups to replicate the benchmark, extend it to additional crops, or test newer architectures under the same controlled conditions, addressing the reproducibility concerns that have dogged applied machine learning literature.

Looking forward, the authors identify clear next steps: larger and better-balanced datasets, since rare disease classes remain harder to classify than common ones, and practical field implementation, where models must contend with variable illumination, occlusion, motion blur, and the constraints of edge devices. The demonstrated strength of the MobileViT-v2 family suggests that the future of agricultural AI may not belong to the largest models but to the smartest small ones, capable of running offline in a farmer’s pocket. If subsequent field trials confirm the laboratory figures, the humble sugarcane leaf may become one of the early showcase successes for transformer-based precision agriculture, bridging the gap between cutting-edge computer vision research and the daily decisions of growers across the tropics.

Subject of Research: Deep learning benchmarking of vision transformers and convolutional networks for sugarcane leaf disease classification

Article Title: Benchmarking vision transformers and convolutional networks for sugarcane leaf disease classification on a consolidated multi-source dataset

Article References: Franco-Hernández, M., Mejía-Domínguez, V., & Rodriguez-Gallo, Y. (2026). Benchmarking vision transformers and convolutional networks for sugarcane leaf disease classification on a consolidated multi-source dataset. Neural Computing and Applications, 38(19), Article 782. https://doi.org/10.1007/s00521-026-12536-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12536-8

Keywords: sugarcane, leaf disease classification, vision transformer, MobileViT-v2, convolutional neural networks, deep learning, precision agriculture, Leave-One-Dataset-Out validation, Grad-CAM, data augmentation, plant pathology, computer vision