Intelligent leaf disease diagnosis: image algorithms using Swin Transformer and federated learning

In recent years, the international population has surged which has led to increased pressure on food security. Plant diseases are one of the most essential factors affecting food security. Plant diseases often manifest pathological features on plant leaves. This paper combines deep learning with continual learning and federated learning, proposing a federated continual learning comprehensive model based on the Swin Transformer model (SSPW224-LwF-3). Our datasets come from shared datasets and integrate data from other datasets based on a ratio principle of 6:1:3 for the training set, validation set and test set (https://osf.io/v4qfr/?view_only=38b53e39988c4e1e82363031a921f799). Experimental results show that the SSPW224 model achieves an accuracy, precision, recall and F1-score of 97.20%, 95.25%, 94.85% and 94.71% on the test set. It is evident to find out that the SSPW224 model outperforms the other network models in the identification and classification of plant leaf diseases from the data and visualization perspectives. The average accuracy of the SSPW224-LwF-3 model strategy combination proposed in this study reaches the highest 47.06% among 18 combination strategies, with an accuracy of 49.41% and 33.81% for old and new data. The overall performance of this model is optimal. Therefore, the proposed SSPW224-LwF-3 model can achieve efficient, sustainable, and distributed recognition and classification of plant leaf diseases which provides a tool for plant disease identification in smart agriculture.