Wheat Disease Classification Using Continual Learning

As wheat is one of the major crops worldwide, therefore, accurate disease detection in wheat plants is critical for mitigating effects and halting disease spread. Nowadays, the detection of diseases through images using machine learning and deep learning has achieved state-of-art results. Yet these models suffer from two shortcomings, one is the data-hungry models of deep learning, and the second is the adaption of new classes with previous ones, i.e., if a new disease arrived, one must retrain the entire model with a new dataset which has led to the origination of few-shot learning that solves both problems. A wheat disease classification network is proposed based on a few-shot learning method for the classification of 18 wheat diseases using EfficientNet as a backbone. Also, the attention mechanism is incorporated to facilitate efficient feature selection. Our proposed network has achieved 93.19% accuracy on the 40 images of 18 disease classes manually collected from the internet while that of our backbone network EfficientNet has gained an accuracy of about 98.5% accuracy on CGIAR Dataset. The evaluation demonstrates that despite the presence of only a few images in both the training and query sets, the proposed model has achieved superior performance in terms of accuracy and computation cost. Furthermore, the evaluation also highlights the effectiveness of our model in terms of data and computation cost. The proposed method can be used for disease detection in wheat crops requiring less amount of data.