Detection of cocoa pod diseases using a hybrid feature extractor combining CNN and vision transformer with dual classifier

Ivory Coast is the world’s leading cocoa producer, with a harvest of over two million tons in 2023. This dominance is severely threatened by several diseases, including Swollen Shoot, which was first detected in 1943. The epidemic, particularly severe in 2003, destroyed over 77,000 hectares of plantations. The prevalence of the disease has continued to rise, affecting new farms and causing substantial economic losses. In response to this situation, research is focusing on innovative solutions, such as artificial intelligence, to automate the detection of these diseases. In this study, a computer vision system was developed using a hybrid algorithm that combines convolutional neural networks and Transformers for feature extraction, along with dual classification through SVM and LightGBM for Swollen Shoot symptom detection. Our algorithm analyzes images of cocoa pods to identify symptoms such as reddish or black spots, brown or black lesions, or pod malformation. This model achieved an accuracy rate of 99.24%, surpassing other similar methods mentioned. Promising results for the classification of cocoa pod images, pave the way for practical solutions in the management of the Swollen Shoot pandemic. Through mobile applications or embedded systems, this technique will also contribute to the needs of early detection and intervention for precision agriculture.