A computationally efficient hybrid framework combining deep feature extraction and gradient boosting for early...

Olive production holds a significant position in the global agricultural trade. In addition to being influenced by seasonal climatic conditions, agricultural diseases are another factor that affects olive yield. Peacock spot disease and olive bud mites are the primary agricultural diseases affecting olive production. These two diseases cause specific lesions in the leaves of olive trees. It has been observed that artificial intelligence approaches such as deep learning and machine learning are used for early detection of such adverse conditions. However, the need for high computational processing in the classification and detection processes of deep learning models limits the accessibility of these algorithms to all businesses. To address this challenge, this study proposes a hybrid framework that combines the robust feature extraction capabilities of deep learning models with the computational efficiency of machine-learning classifiers. Specifically, this study analyzes the performance of this combined approach and compares the results with those of existing deep learning studies in the literature. As feature extraction deep learning models, MobileNetV2, DenseNet121, EfficientNetV2B0, and ConvNext Tiny were selected, while AdaBoost, XGBoost, LightGBM, CatBoost, and Gradient Boosting algorithms from the Boosting family were included as classifiers. In the model training, a dataset consisting of 3,400 images of olive leaves belonging to three classes-healthy, olive_peack_spot, and aculus_olearius-was used. The experimental results showed that the DenseNet121 + XGBoost combination achieved a baseline accuracy of 92%. Following a data augmentation phase to enrich the training data, the model performance was significantly enhanced, reaching a final accuracy of 94% and a macro average F1-Score of 94%. The Wilcoxon Signed-Rank test revealed that the DenseNet121 + XGBoost combination statistically outperformed the second-best model (p < 0.05). This performance is attributed to the dense connectivity of DenseNet, which promotes effective feature reuse and improves gradient flow. Furthermore, the study demonstrates that a higher number of parameters does not always guarantee better performance; rather, architectural efficiency plays a crucial role in avoiding overfitting and ensuring model robustness, as evidenced by DenseNet121 outperforming the larger ConvNeXtTiny model.