A hybrid CNN-transformer model with adaptive activation function for potato leaf disease classification

Tidsskrift: Scientific Reports, vol. 16, 4282, 2026

Doi: doi.org/10.1038/s41598-025-34406-4
Arkiv: hdl.handle.net/11250/5360461

Sammendrag:
Abstract Potato plants are highly vulnerable to numerous diseases that can substantially affect both yield and quality. Conventional approaches for detecting these diseases are often labor-intensive, slow, and prone to inaccuracies, particularly under variable environmental conditions. This study presents a hybrid deep learning architecture, termed potato leaf diseases DenseNet (PLDNet) , which integrates a DenseNet-based convolutional neural network with a Transformer-based attention module to accurately classify potato leaf diseases. Furthermore, an adaptive parametric activation function, referred to as Adaptive Flatten p-Mish (AFpM) , is proposed to enhance the model’s learning flexibility and representational capacity. When evaluated on the PlantVillage and Mendeley datasets, PLDNet attains classification accuracies of 99.54% and 87.50%, respectively, surpassing contemporary state-of-the-art models and activation techniques. The proposed framework exhibits strong generalization performance and offers a scalable, efficient approach for automated plant disease identification. To highlight the novelty, the proposed AFpM activation function introduces a learnable parameter enabling adaptive nonlinearity, improving over Mish, Swish, and PFpM activation functions through dynamic gradient control. AFpM improves accuracy by 2.52% on Mendeley dataset, and 1.93% on PlantVillage dataset compared to PFpM, and by more than 3% compared to Swish and Mish.