MultiBranchFusionNet: A Lightweight Attention-Guided Dual-Branch Framework for Tomato Leaf Disease Classification Using Heterogeneous Field and Public Images

Abstract

Tomato foliar diseases reduce yield, degrade fruit quality, and cause major economic losses, especially in regions such as Nigeria where diagnosis still relies largely on manual visual inspection. Although deep learning has improved automated plant disease recognition, many models are trained on controlled datasets that do not capture field variability. This study presents an attention-guided multi-branch transfer learning framework for tomato leaf disease classification using a heterogeneous dataset of 20,245 images from South-West Nigerian farms and public sources, spanning ten disease categories. Field-acquired labels were assigned through visual symptom assessment with expert adjudication of uncertain cases and should therefore be interpreted as symptom-based classifications rather than pathogen-confirmed diagnoses. During preliminary backbone screening, MobileNetV2 was selected as the pretrained feature-extraction pathway because it combined strong recorded classification performance with a compact model profile. MultiBranchFusionNet integrates a frozen ImageNet-pretrained MobileNetV2 branch enhanced with Squeeze-and-Excitation attention and a custom convolutional branch, with feature-level fusion used to improve discrimination among visually similar disease classes. Under the evaluated mixed-source image-level protocol, the proposed model achieved 97.28% test accuracy, with macro-precision, macro-recall, and macro-F1-score values of 96.75%, 96.69%, and 96.66%, respectively. Calibration and qualitative Grad-CAM analyses were used to examine model behaviour. The image-level partitioning strategy, together with the absence of systematically retained plant- and acquisition-group identifiers, prevented verification of independence at the plant, farm, acquisition-session, state, collection-phase, or source level. Although combining field-acquired and public images increased acquisition diversity within the assembled dataset, it did not constitute source-independent external validation. The reported performance should therefore be interpreted as internal mixed-source image-level performance rather than evidence of independent field-domain generalisation. The model has a compact size of approximately 15.28 MB, supporting its characterisation as architecturally lightweight; however, practical mobile or edge deployment remains to be established through direct device-level evaluation.