Ensemble of Deep Convolutional Network for Citrus Disease Classification Using Leaf Images

Crop disease is a major threat to agricultural production. Reduced yield due to crop diseases can lead to immeasurable economic losses. Therefore, the detection and classification of crop diseases are of great significance. Based on AlexNet, VggNet, ResNet, and DenseNet, this paper presents an Ensemble Network improve the recognition accuracy. In this study, firstly training four CNN models by randomly sampling the dataset. Before that, these CNN were improved by batch normalization and global average pooling to accelerate convergence, designed a dynamic concatenated ReLU, an improved activation function on ReLU, to improved detection performance. We use focal loss to solve data imbalance. Then weighted voting is used to fuse the four CNN models. Finally, we verify that the EnsembleNet can effectively improve the recognition performance compared with a single network. Verified improvement in recognition performance on our datasets, obtained from PlantVillag.org, containing 4577 citrus leaf images of three categories. The maximum test accuracy in identifying citrus leaf diseases was as high as 93.58 \(\%\) . Compared with single network, our EnsembleNet can significantly performance. The experimental results showed that this method can be practically applied to the identification of citrus leaf diseases and provides a basis for the identification of other plant leaf diseases.