Performance prediction of tomato leaf disease by a series of parallel convolutional neural networks

We present a new approach of classification decision of tomato (Solanum lycopersicum) leaf disease based on the summarization of the output from a series of parallel convolutional neural networks with different configurations. The use of Swish, LeakyReLU-Swish, ReLU-Swish, Elu-Swish, and ClippedReLU-Swish activation layers, as well as the Batch Normalization-Instance Normalization layer significantly improves the network performance, achieving classification accuracy over 99.0% with training, 97.5% with validation and 98.0% with testing dataset. Despite various performance metrics observed, none of the proposed networks overfitted on the validation dataset. Furthermore, we use various techniques to visualize the network performance. This demonstrates how the networks (Network 1, Network 2, Network 3, Network 4, Network 5) learn from the training dataset and can show diseased areas of leaves with high confident scores under real conditions. Network 1 shows the best performance in terms of network stability and visualization of the disease location. By computing, summarizing, and scoring the output of a series of parallel convolutional neural networks, the weakness of Network 4 and Network 5 in predicting the Healthy class can be overcome. This research will inspire and encourage further use of deep learning techniques to automatically detect and classify plant diseases under real conditions and improve financial condition of the farmers worldwide.