Transformers Outperform CNNs at Spotting Diseased Cotton Leaves in Pakistani Field Study

Researchers in Pakistan have compared a dozen deep learning architectures on thousands of cotton leaf images collected from fields in Bahawalpur, Punjab, and found that transformer-based models, particularly the Swin Transformer and Vision Transformer (ViT), classify cotton leaf diseases more accurately than most convolutional neural networks. The study, published on 25 September 2026 in the open-access journal Agriculture & Food Security, evaluated models on both annotated and non-annotated versions of a self-collected dataset of 2,500 images spanning five categories: Caterpillar damage, Herbicide Burn, Leaf Variegation, Verticillium Wilt, and Healthy leaves.

The motivation is rooted in the economic weight of cotton in Pakistan, where the crop is often called ‘white gold’ and supports large numbers of farmers and laborers in Punjab and Sindh. According to the paper, cotton production reached around 6.85 million bales in 2023-24, recovering from the flood-affected 2022-23 season, but predicted production for 2024-25 is 6.59 million bales, roughly a 4 percent decrease attributed to unusually low temperatures in mid-March that hampered germination and early development in Sindh and Punjab. Environmental change, pest infestation, and rising input costs such as fertilizers, pesticides, and seeds add further pressure, and diseases like caterpillar infestations, herbicide burns, leaf variegation, and verticillium wilt reduce both the quality and the saleable yield of cotton.

Traditional disease monitoring relies on manual inspection and expert recommendation, which the authors describe as time-consuming, labor-intensive, and prone to inaccuracy, and which does not scale easily to large farming operations. Deep learning offers an alternative: models that learn to recognize disease symptoms directly from photographs. The research team photographed cotton leaves in natural field conditions between August and October, capturing varied lighting, leaf orientations, and disease severity levels. The original images had a resolution of 1280 by 960 pixels and were converted to a square 640 by 640 format after annotation, then resized to 224 by 224 pixels to match the input requirements of architectures including VGGNet, ViT variants, ResNet, MobileNet, GoogleNet, and AlexNet. Annotations were made manually using the Roboflow platform to verify disease classifications.

Preprocessing included normalizing pixel values to the range of minus one to one, which stabilizes the input distribution and speeds gradient descent, and applying data augmentation with geometric and photometric transformations such as rotations, horizontal and vertical flipping, shearing, and brightness variation. For the annotated dataset, transformations were applied in a controlled way to preserve annotation consistency, while the non-annotated dataset received randomized image-level augmentation to enlarge sample sizes and maintain balanced class distributions. The dataset was split into disjoint training and test partitions, and models were trained for 100 epochs.

The architectures tested spanned two families. On the convolutional side were a simple CNN trained from scratch, GoogleNet with its parallel inception modules, ResNet with residual connections that ease the training of very deep networks, the lightweight depthwise-separable MobileNet, the eight-layer AlexNet, and VGG16 and VGG19 with their stacks of small 3 by 3 filters. On the transformer side were ViT, which treats images as sequences of patches and uses self-attention to capture global dependencies; the Swin Transformer, which restricts self-attention to non-overlapping local windows with a shifted-window mechanism to cut computational cost; the hybrid CNN Transformer; DeiT, a data-efficient vision transformer that learns from a pretrained CNN teacher through distillation; and PVT, a pyramid-structured transformer designed for high-resolution inputs. ViT and Swin used pretrained backbones fine-tuned on the dataset.

On the non-annotated dataset, ViT achieved the highest single-split accuracy at 78.40 percent, followed by DeiT at 77.00 percent and the Swin Transformer at 75.40 percent, with balanced precision and recall for the leading models. ResNet offered the highest recall among all models, while MobileNet, ResNet, VGG16, and VGG19 showed moderate accuracy with high precision, recall, and F1 scores. Confusion matrices revealed that models struggled most to distinguish between closely related categories, particularly Herbicide Burn and Verticillium Wilt, and ROC analysis supported the ranking of the leading transformers by area under the curve.

The picture shifted under k-fold cross-validation with k equal to 5, 10, and 15, a strategy the authors used to reduce the risk of data-partition bias. On the non-annotated data, the Swin Transformer consistently outperformed all other models across every fold, reaching a maximum accuracy of 90.83 percent with 15 folds, while ViT averaged around 79.91 percent with stable F1 scores. DeiT also delivered consistent results across folds. MobileNet exceeded 70 percent accuracy across all folds, whereas ResNet and AlexNet struggled notably, which the authors attribute to vulnerability to non-annotated input and potential overfitting. The CNN Transformer and PVT performed better than AlexNet and ResNet despite unimpressive overall figures.

Annotation improved most models. On the annotated dataset, ViT again led the single-split evaluation at 79.20 percent accuracy with balanced precision, recall, and F1 score, followed by the Swin Transformer at 77.20 percent and DeiT at 74.80 percent. Under cross-validation the Swin Transformer again dominated, achieving its maximum accuracy of 98.45 percent with 15 folds, while ViT averaged around 82.87 percent and DeiT averaged above 76.63 percent. MobileNet surpassed 67 percent accuracy across all annotated folds, and ResNet and AlexNet again lagged. The authors report that annotations reduced variability and ambiguous features within classes, and that ROC curves showed increased AUC values on annotated data, with the CNN in particular making a substantial improvement that narrowed the gap between convolutional and transformer-based approaches.

The authors frame these findings as a step toward AI-based diagnostic tools in precision and smart agriculture, arguing that accurate, efficient, and reproducible disease classification could help farmers distinguish biotic stressors such as caterpillar infestation from abiotic ones such as herbicide burn, enabling more targeted pesticide and herbicide application. They suggest that integrating such models with drones, mobile applications, and IoT-enabled field cameras could reduce reliance on manual inspection, and they connect the work to Sustainable Development Goals on zero hunger, responsible consumption and production, and life on land. The paper also proposes policy measures including subsidies for intelligence-assisted disease tracking systems, collaborative data-sharing frameworks for annotated agricultural datasets, and training programs to help farmers adopt AI-based diagnostic tools.

The study’s scope carries clear limitations. The dataset of 2,500 images was self-collected from a single region over one growing season, the five disease categories represent a narrow slice of cotton pathology, and the reported accuracies describe model performance on this specific dataset rather than validated performance in deployed farm settings. The authors themselves note that real-world implementation factors such as scalability and environmental variability remain underexplored in the wider literature, and they identify future work in model fine-tuning, hyperparameter tuning, ensemble learning, advanced augmentation, and analysis of confusion matrices to resolve the disease categories that models most often confuse. They also flag gaps in existing research, including the scarcity of comprehensive cotton leaf disease studies in Pakistan’s agricultural sector, the dominance of prior work on rice, wheat, and maize, and the limited assessment of transformer-based and fusion architectures for cotton disease detection. Within those boundaries, the consistent cross-validation advantage of the Swin Transformer on both annotated and non-annotated data, together with ViT’s strong showing, marks transformer architectures as the standout performers of this comparison and suggests that labeled data quality remains a decisive factor in how reliably such systems can identify diseased cotton leaves.

Subject of Research: Deep learning classification of cotton leaf diseases from field images collected in Punjab, Pakistan

Article Title: Cotton leaf disease classification using deep learning models for smart agriculture

Article References: Abbas, H. K., Shahid, M. F., Ashari, R. B., Mashat, A. A., Khanzada, T. J. S., & Tanveer, M. H. (2026). Cotton leaf disease classification using deep learning models for smart agriculture. Agriculture & Food Security, 15(1), Article 41. https://doi.org/10.1186/s40066-026-00610-2

Image Credits: AI Generated

DOI: 10.1186/s40066-026-00610-2

Keywords: cotton, deep learning, Vision Transformer, Swin Transformer, plant disease detection, precision agriculture, Pakistan, computer vision, machine learning, crop health, agriculture, Bahawalpur