Enhancing crop sustainability: GSEBNet for precise tomato leaf disease classification

Introduction Plant diseases pose a major threat to food-supply sustainability. Early identification of crop diseases is crucial for agricultural production and crop yield. Classifying tomato leaves helps determine the treatments needed to keep crops healthy. Although many deep learning approaches have been proposed for plant disease classification, most rely on laboratory-captured images and struggle with real-field conditions. Methods This study introduces an extensive class-based dataset collected in natural environments, visually distinguishing 13 infectious diseases and healthy tomato leaves. We propose the Gamma-preprocessed Squeeze-Excitation Block Network (GSEBNet), a lightweight CNN architecture that integrates Squeeze-and-Excitation (SE) blocks with carefully configured Mobile Inverted Bottleneck Convolution (MBConv) layers and a domain-specific Gamma-Unsharp Brightness Reduction (GUBR) preprocessing pipeline. The model was trained and evaluated on a combined dataset of field-captured (PlantCity) and laboratory (PlantVillage) images covering 14 classes. Results The proposed GSEBNet achieved a test accuracy of 99.48% on the 14-class dataset, with precision, recall, F1-score, and Cohen’s Kappa of 99.49%, 99.49%, 0.99, and 0.9944, respectively. Discussion The obtained results are competitive with those reported in previous studies, while acknowledging differences in datasets and experimental settings. The synergistic co-design of GUBR preprocessing with SE-enhanced MBConv layers significantly improves feature discriminability in low-quality field-captured images.