Abstract:
The most significant limitations to commercial mango production are mango leaf diseases and pest infestation which usually lead to a significant loss in yield and poor quality of the fruits. The visual similarity between the symptoms of disease, changes in the environment, and the low generalization capacity of the standard pre-training deep learning (DL) models are the obstacles to timely and accurate identification. In order to overcome these issues, this paper presents MangoXPPNet, a special and lightweight CNN framework that provides highly discriminative feature extraction without instability in various imaging conditions. The model is tested on three benchmark datasetsMangoLeafBD, Mango Pest Classification and MLDID without changing its fundamental structure and shows good flexibility in the disease and pest categories. It has been experimentally shown that MangoXPPNet classifies MangoLeafBD with nearperfect accuracy (99.5%), as well as has a strong cross-dataset generalization on MangoPest (95.14%) and MLDID (98.33%). ROC-AUC experiments indicate that the separability is high in all classes, and the submission of analysis values to AUC values are close to 1.00 even in cross-domain conditions. Confusion matrices also demonstrate that there are consistent class-level precision of the models, especially those hard to see, like Gall Midge, Sooty Mould, and Powdery Mildew. Grad-CAM and saliency maps were combined to improve interpretation, and it was found that MangoXPPNet partially concentrates on biologically significant areas and thus guarantees clear and reliable forecasts. The findings have made MangoXPPNet a strong, interpretable and generalizable solution to automated mango disease and pest classification. Due to its stability, cross-dataset consistency, and XAI-based transparency, MangoXPPNet has great potential to be implemented in a real-world precision agriculture system that can be used to detect early diagnoses, minimize losses on crops, and sustain mango production.
