Deep Learning for the Classification of Cassava Leaf Diseases in Unbalanced Field Data Set

Cassava is one of the main sources of carbohydrates in the world. However, the diagnosis of diseases in cassava crops is laborious, time-consuming and requires specialised personnel. In addition, very little research is available on images of cassava leaves taken with mobile phones and under field conditions. Therefore, the study designs deep learning models for the detection of diseases in cassava leaves from photos taken with mobile phones in the field. This study used a dataset of 21'397 images of cassava bacterial blight, cassava brown streak disease, cassava green mottle and cassava mosaic disease from a Kaggle competition. Twelve CNN models have been evaluated by applying transfer learning and data augmentation. Each of the models was trained with uniform samples and class-weighted samples. The results showed that the use of weighted samples reduced F1 score and accuracy in all cases. Furthermore, the DenseNet169 model was outstanding with an accuracy and F1 score of 74.77% and 0.59 respectively. Finally, the causes that hinder correct classification have been identified. The results reveal that it is still necessary to work on creating a balanced and refined database.