Utilizing Squeeze-and-Excite Mechanisms and Voting Ensembles Technique for Coffee Leaf Disease Detection

Coffee is a crucial global crop, but its production is often hindered by plant diseases such as coffee leaf rust, coffee berry disease, and bacterial blight, leading to significant yield losses. This study aims to develop an accurate disease detection system for coffee plants using an ensemble of deep learning models. A key contribution of this paper is assessing the impact of adding squeeze-and-excite blocks to Convolutional Neural Networks (CNNs) for enhancing coffee disease detection. We employed several pre-trained CNN models, including MobileNetV2, ResNet152, EfficientNetB7, EfficientNetV2B0, InceptionV3, ResNet50, DenseNet121, and DenseNet201, each modified with squeeze-and-excite blocks to improve feature extraction. The models were trained with three optimizers—Adam, RMSProp, and SGD—with Adam consistently delivering the best results. We then implemented a voting ensemble technique, combining the top-performing CNN models, and found that the ensemble of EfficientNetV2B0 and EfficientNetB7 achieved the highest accuracy of 97.75%. This approach demonstrates the effectiveness of deep learning ensembles and the added benefit of squeeze-and-excite blocks in accurately detecting coffee plant diseases, offering a valuable tool for improving crop management and production.