Leakage-Aware Transfer Learning with Explainable AI and CPU-Efficient Deployment for Mango Leaf Disease Classification on the MangoLeafBD Benchmark

Background and Aim: Mango (Mangifera indica) is a globally important fruit crop whose productivity is repeatedly threatened by foliar diseases such as anthracnose, bacterial canker, powdery mildew, die-back, and sooty mould. Although deep learning has rapidly advanced automated leaf-disease diagnosis, recently reported accuracies on the MangoLeafBD benchmark are approaching saturation, and many studies still rely on random data splits that may inflate performance through leakage of duplicate or near-duplicate images. This study aimed to develop and rigorously evaluate a leakage-aware, deployment-oriented deep learning framework for the complete eight-class MangoLeafBD task. Methods: Three modern transfer-learning backbones—EfficientNetB0, MobileNetV3Large, and ConvNeXtTiny—were fine-tuned using a two-stage training strategy on a group-aware data partition. Duplicate and near-duplicate images were detected with cleaned filename stems, average hashing (aHash), and difference hashing (dHash) and grouped before splitting to guarantee zero cross-partition overlap. The strongest models were combined through probability-level soft voting. Robustness was further assessed using 5-fold StratifiedGroupKFold cross-validation. Explainability was examined with Gradient-weighted Class Activation Mapping (Grad-CAM), deployment suitability was characterized through CPU latency benchmarking, and the framework was operationalized as a publicly accessible web-based diagnostic system. Results: EfficientNetB0 achieved the highest performance under the leakage-controlled protocol, reaching 99.50% accuracy and 99.50% weighted F1-score on the 599-image group-aware test set. The heterogeneous soft-voting ensemble matched EfficientNetB0 but did not exceed it, indicating that the ensemble gains reported in earlier studies may partly reflect optimistic split conditions. Five-fold grouped cross-validation confirmed stability, yielding a mean accuracy of 99.63% (SD = 0.20%) and a mean weighted F1 of 99.62% (SD = 0.20%). Grad-CAM visualizations showed that the model attended to biologically meaningful lesion regions across all eight classes, and CPU benchmarking produced a mean single-image latency of 25.3 ms (p95 = 27.6 ms) with batch throughput scaling up to 166.5 images per second. Conclusions: The proposed framework demonstrates that a compact, leakage-aware EfficientNetB0 model can match the accuracy of more complex hybrid CNN–transformer architectures while remaining interpretable and deployable on commodity CPU hardware. By coupling group-aware evaluation, explainable AI, latency benchmarking, and a publicly accessible web application, the study advances reproducible, deployment-ready precision agriculture research and offers a more rigorous benchmarking protocol for future mango leaf disease classification studies.