Plant disease recognition in a low data scenario using few-shot learning

Abstract

Plant disease is one of the major problems in agriculture. Diseases damage plants, reduce yields and lower the quality of the produce. Traditional approaches to detecting plant diseases are usually based on visual inspection and laboratory testing, which can be expensive and time-consuming. They require trained plant pathologists as well as specialised equipment. Several studies demonstrate that artificial intelligence (AI) methods can produce promising results. However, AI methods are generally data-hungry and require large annotated datasets, and the collection and annotation of such datasets can be a limiting factor. It often appears that only a small amount of data is available for certain disease types. Whereas the performance of typical AI methods drops significantly when they are trained with inadequate data. This paper proposes a novel few-shot learning (FSL) method to detect plant diseases and alleviate the data scarcity problem. The proposed method uses as few as five images per class in the machine learning process. Our method is based on a state-of-the-art FSL pipeline called pre-training, meta-learning, and fine-tuning (PMF), integrated with a novel feature attention (FA) module; we call the overall method PMF+FA. The FA module emphasises the discriminative parts in the image and reduces the impact of complicated backgrounds and undesired objects. We used ResNet50 and Vision Transformers (ViT) as the feature learner. Two publicly available plant disease datasets were repurposed to meet the FSL requirements. We thoroughly evaluated the proposed method on the PlantDoc dataset, which contains disease samples in field environments with complex backgrounds and unwanted objects. The PMF+FA method with ViT achieved an average accuracy of 90.12% in disease recognition. The results demonstrate that the PMF+FA pipeline consistently outperforms the baseline PMF. The results also highlight that the method using ViT generates better results than ResNet50 for diagnosing complex data. ViT and ResNet50 implementations are computationally efficient, taking 1.11 and 0.57 ms on average per image to evaluate the test set respectively. The high throughput and high-quality performance with only a small training dataset indicate that the proposed technique can be used for real-time disease detection in digital farming systems.