A Hybrid Feature Fusion and Machine Learning Approach for Cocoa Pod Disease Detection

Early detection of diseases affecting cocoa pods is vital for preserving harvest quality and minimizing agricultural losses. Thanks to advances in image analysis and artificial intelligence, it is now possible to automate these diagnoses. Convolutional neural networks (CNNs) are effective at extracting deep visual features, although they can exhibit some instability during training. Additionally, traditional descriptors such as color, texture, and shape provide valuable information to enhance classification. In this context, we propose a multi-feature fusion approach designed to improve the robustness and accuracy of diagnosis. CNNs stabilized using the exponential moving average method were used to extract deep features, which were then fused. This fusion was combined with classical descriptors, and then a dimension reduction by principal component analysis (PCA) was applied before the classification phase. We evaluated the performance of several models, including SVM, CatBoost, and a stacking-based ensemble model, under three different feature extraction strategies: the fusion of stabilized convolutional neural networks (FCNN), the fusion of CNNs with classical descriptors (CNN + CD), and the fusion of CNNs with classical descriptors reduced by principal component analysis (CNN + CD + PCA). The results show that the combination of CNN + CD + PCA associated with an SVM classifier significantly improves accuracy, reaching 91.48%. Furthermore, applying stacking to data from this same strategy achieved the best overall performance with an accuracy of 93.73%, while reducing error and stabilizing predictions. Thus, combining multiple types of features with a set-theoretic approach represents a promising solution for the automatic detection of cocoa diseases. This work paves the way for integration into mobile or embedded tools for producers, while encouraging future research on optimizing visual information fusion methods.