LXViT: a hybrid hierarchical CNN-ViT framework for lemon leaf disease classification with multi-method explainable AI integration

Abstract Lemon leaf diseases threaten agricultural productivity, yet early detection remains difficult due to subtle lesion patterns and environmental variability. Conventional deep learning models often fail to balance local feature extraction with global context. To address this, we introduce LXViT, a hybrid hierarchical CNN-Vision Transformer (CNN–ViT) model. The architecture synergizes an Xception backbone for localized spatial features with a Vision Transformer (ViT) module to capture long-range dependencies. Utilizing a newly curated dataset, LemonLENS (3,266 high-resolution images, six classes), the pipeline incorporated rigorous preprocessing (denoising, CLAHE and gamma correction) and SMOTE for class imbalance. The model achieved a high-performing accuracy of 99.59% on the internal LemonLENS test set, significantly outperforming benchmarks like VGG19, ResNet152V2, ConvNeXt Tiny, and MobileNetV2. Statistical stability was confirmed via stratified 5-fold 99.75% and 10-fold 99.21% cross-validation. Technically, the framework ensures edge efficiency, requiring only 9.652 GFLOPs with 145.98 ms inference latency. Generalizability was validated on three external datasets: PlantCity 98.27%, the 6-class Lemon leaf dataset 88.49%, and the 9-class Lemon leaf dataset 97.57%. Furthermore, a multi-method Explainable AI (XAI) framework employing LIME, Grad-CAM, and Occlusion Sensitivity bridged the “black-box” gap, highlighting biologically relevant lesion regions aligned with expert assessment. These findings demonstrate LXViT as a promising architecture for lemon leaf disease classification, warranting further validation across diverse geographic and environmental conditions before large-scale field deployment.