A hybrid deep learning framework for automated mango leaf disease classification

An effective identification of diseases on mango leaves could have important need for orchard health and agricultural efficiency, yet it can be difficult to perform. Some of these diseases look similar because of discolored leaf patterns, spots and distortion of leaf patterns and it overlap between healthy and diseased leaves for various diseases that makes the manual diagnosis ambiguous, difficult, and resource-dependent. Using MangoleafBD dataset (4000 leaf images) from Kaggle, containing both healthy and diseased samples across seven categories including Anthracnose, Bacterial Canker, Powdery Mildew, Sooty Mould, Cutting Weevil, Gall Midge, Die Back and healthy collected from multiple orchards in Bangladesh. The proposed framework integrates the image enhancement using CLAHE with HSV followed by a selection of feature using transfer learning and deep learning models for feature selection models complementary deep feature extraction using three heterogeneous pretrained CNN architectures (EfficientNetB0, ResNet50, and DenseNet121) models from Keras applications which were trained on the ImageNet dataset. The outputs generated by the three pretrained CNN models were combined through feature-level concatenation and then classified using a hybrid XGBoost-Random Forest meta-classifier consisting of XGBoost and Random Forest. The integrated framework combines complementary feature representations with hybrid ensemble classification to improve feature discrimination and classification robustness.The proposed method achieved a 97% accuracy and precision, recall, F1 score, AUC performance.