Convolutional Neural Network (CNN) Models for Crop Diseases Classification

Abstract

Crop diseases have a significant impact on agricultural production. As a result, early diagnosis of crop diseases is critical. Deep learning approaches are now promising to improve disease detection. Convolutional Neural Network (CNN) models can detect crop disease using images with automatic feature extraction. This study proposes crop disease classification considering ten pre-trained CNN models. Fine-tuning for each model was conducted in the Plant Village dataset. The experimental results show that fine-tuning improves the model’s performance with an average accuracy of 8.85%. The best CNN model was DenseNet121, with 94.48% and 98.97% accuracy for freezing all layers and unfreezing last block convolution layers. Moreover, fine-tuning produces less time-consuming with an average of 2.20 hours. VGG19 is the less time-consuming reduction by 8 hours. On the other hand, MobileNetV2 is the second-best performance model with less time-consuming than DenseNet121, and produces fewer parameters, which is affordable for embedding it to mobile devices.