Early Stress Detection in Plant Phenotyping using CNN and LSTM Architecture

In order to ensure the availability of food in the future, plant stress identification is one of the crucial tasks used in plant phenotyping to develop better crops. In this research, we use a convolution neural network (CNN) combined with LSTM to identify the early state of plant stress caused by a deficiency of nutrients. We use a treatment study dataset of sorghum (S. bicolor) which consists of more than 40,000 images of growing sorghum images captured in the phenotyping facility in 3 views. The experiment studies plant growing under 3 treatment conditions: 100/100 (100% ammonium/100% nitrate), 50/10, and 10/10. The network is divided into two parts: the features extraction and classification network. VGG16 with pre-trained weights from the ImageNet dataset is used as the feature extractor. LSTM cell with multi-layer perceptron (MLP) is used to classify extracted features to determine the stress of the plants after subjected to the stressor. The result revealed that the network can detect the stress at the accuracy of more than 85% at 2 days after plants subjected to the stressor treatment.