Potato Research Agronomy and Crop Science

Integrating Convolutional Patch Embedding and Graph Attention for Deep Learning–Based Potato Leaf Disease Classification

Vishwakarma, Anish Kumar, Ranjan, Rakesh
Springer Nature · 2026-09-30 · DOI 10.1007/s11540-026-10150-8

Abstract

Reliable and early diagnosis of potato leaf diseases is vital for optimizing agricultural yields and ensuring food security. This paper presents a hybrid deep learning framework that integrates ConvMixer-based feature extraction with a Graph Attention Network (GAT) to automate disease classification. Within this architecture, the ConvMixer backbone generates high-dimensional patch-level feature embeddings from leaf imagery, while the GAT module constructs a $$\varvec{k}$$ k -nearest neighbour graph to explicitly model relational dependencies between feature embeddings. By coupling local spatial filters with global topological context, the model effectively characterizes complex morphological patterns. We evaluated the framework using the PlantVillage dataset across three classes: early blight, late blight, and healthy. The proposed ConvMixer-GAT framework achieved an overall classification accuracy of 99.38%, with micro-averaged precision, recall, and F1-score of 99.38%, 99.38%, and 99.38%, respectively. Comparison with several recent convolutional and hybrid transformer-based methods reported on the PlantVillage dataset demonstrates that the proposed framework achieves competitive performance while maintaining consistently high classification accuracy. To facilitate real-world deployment, the framework is implemented as a web-based diagnostic tool, PotatoGuard, which provides real-time detection and agronomic recommendations. Our findings demonstrate that the integration of convolutional feature extraction and graph-based relational learning provides an effective framework for automated potato leaf disease classification and shows promising potential for practical decision-support applications in precision agriculture.

Read at publisher →