An Image-Based Deep Learning Model for Cannabis Diseases, Nutrient Deficiencies and Pests Identification

摘要

In this work, a deep learning system for cannabis plants disease, nutrient deficiencies and pests identification is developed, based on image data processed by convolutional neural network models. Training of the models was performed using image data available on the Internet, while database development included data cleansing by expert agronomists, basic image editing, and data augmentation techniques commonly used in deep learning applications in order to expand the rather limited amount of available data. Three fungi diseases, two pests and three nutrient deficiencies were included in the identification system, together with healthy plants identification. The final model reached a performance of 90.79% in successfully identifying cannabis diseases (or healthy plants) in previously 'unseen' plant images. The most difficult cannabis problems to be identified were powdery mildew and potassium deficiency. Results showed that transfer learning from existing models specialized in similar tasks to the one under development, is more successful than using transfer learning from more general models. Finally, even though the amount of training images in some of the considered problems was significantly small, no correlation between model performance and the size of the training dataset for each category was found.