A Deep Recurrent Neural Network for Plant Disease Classification

Abstract

Agriculture is important in emerging nations like India, but food security is still a serious problem. Plant diseases, inadequate storage facilities, and poor transportation cause the majority of harvests to be squandered. Since illnesses cause almost 15% of India’s crop yield to be lost, this is a big issue that needs to be addressed. This proposed model is an automated system that can identify the diseases and assist farmers to take the necessary action to cure the crop losses. Farmers have been using the traditional method of using their own eyes to detect plant illnesses, but not all farmers can detect these diseases in the same way. Computer vision capabilities must be incorporated into agriculture given the advancements in artificial intelligence. The proposed model uses a convolutional neural network (CNN) with Recurrent Neural Network (RNN) for PlantVillage dataset, the greatest publicly accessible dataset. The proposed model has a 99.37% prediction accuracy for the condition. The proposed approach can identify 14 different plant classes out of the 38 and other moderate in the Plant Village dataset shows how versatile it is. Farmers may decrease crop loss and enhance crop quality and output using this automated and user-friendly technique. In this study, we present the use of a deep recurrent neural network to automatically detect plant diseases. The resulting algorithm is used to identify the bacterial blight of rice during the growing season with a detection accuracy of 99.16%, a classification accuracy of 99.17%, and a sensor-based detection accuracy of 98.98%. Recurrent networks have made great advances in various sequence modeling, such as speech recognition, language modeling, image captioning, and many other applications in recent years. We detect the bacterial blight of rice leaves in this study with a deep recurrent network. We use a stacked LSTM-CNN network to train representations for the radio signal data collected during the lifespan of the rice plant precision agriculture experiment. Our proposed system works 99.16% of the time on rice in the non-padded dataset and 98.98% within ± 10 days from the actual disease detection in the padded dataset, with a detection accuracy of 99.17% from only the signals transmitted between rice and a 135 MHz antenna.