Detection of Cocoa Leaf Diseases Using the CNN-Based Feature Extractor and XGBOOST Classifier  

作　　者：Kouassi Simeon Kouassi Mamadou Diarra Kouassi Hilaire Edi Brou Jean-Claude Koua Kouassi Simeon Kouassi;Mamadou Diarra;Kouassi Hilaire Edi;Brou Jean-Claude Koua(Mechanics and Computer Science Laboratory, Felix Houphout-Boigny University, Abidjan, Ivory Coast;Mathematics and Computer Science Laboratory, Nangui Abrogoua University, Abidjan, Ivory Coast)

机构地区：[1]Mechanics and Computer Science Laboratory, Felix Houphout-Boigny University, Abidjan, Ivory Coast [2]Mathematics and Computer Science Laboratory, Nangui Abrogoua University, Abidjan, Ivory Coast

出　　处：《Open Journal of Applied Sciences》2024年第10期2955-2972,共18页应用科学（英文）

摘　　要：Among all the plagues threatening cocoa cultivation in general, and particularly in West Africa, the swollen shoot viral disease is currently the most dangerous. The greatest challenge in the fight to eradicate this pandemic remains its early detection. Traditional methods of swollen shoot detection are mostly based on visual observations, leading to late detection and/or diagnostic errors. The use of machine learning algorithms is now an alternative for effective plant disease detection. It is therefore crucial to provide efficient solutions to farmers’ cooperatives. In our study, we built a database of healthy and diseased cocoa leaves. We then explored the power of feature extractors based on convolutional neural networks such as VGG 19, Inception V3, DenseNet 201, and a custom CNN, combining their strengths with the XGBOOST classifier. The results of our experiments showed that this fusion of methods with XGBOOST yielded highly promising scores, outperforming the results of algorithms using the sigmoid function. These results were further consolidated by the use of evaluation metrics such as accuracy, mean squared error, F score, recall, and Matthews’s correlation coefficient. The proposed approach, combining state of the art feature extractors and the XGBOOST classifier, offers an efficient and reliable solution for the early detection of swollen shoot. Its implementation could significantly assist West African cocoa farmers in combating this devastating disease and preserving their crops.Among all the plagues threatening cocoa cultivation in general, and particularly in West Africa, the swollen shoot viral disease is currently the most dangerous. The greatest challenge in the fight to eradicate this pandemic remains its early detection. Traditional methods of swollen shoot detection are mostly based on visual observations, leading to late detection and/or diagnostic errors. The use of machine learning algorithms is now an alternative for effective plant disease detection. It is therefore crucial to provide efficient solutions to farmers’ cooperatives. In our study, we built a database of healthy and diseased cocoa leaves. We then explored the power of feature extractors based on convolutional neural networks such as VGG 19, Inception V3, DenseNet 201, and a custom CNN, combining their strengths with the XGBOOST classifier. The results of our experiments showed that this fusion of methods with XGBOOST yielded highly promising scores, outperforming the results of algorithms using the sigmoid function. These results were further consolidated by the use of evaluation metrics such as accuracy, mean squared error, F score, recall, and Matthews’s correlation coefficient. The proposed approach, combining state of the art feature extractors and the XGBOOST classifier, offers an efficient and reliable solution for the early detection of swollen shoot. Its implementation could significantly assist West African cocoa farmers in combating this devastating disease and preserving their crops.

关 键 词：Machine Learning Cocoa Leaf Diseases Deep Learning Convolutional Neural Network Feature Extraction Image Recognition XGBOOST 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]