FedAD: Federated learning with generative data augmentation and adaptive parameter decomposition for rice pest and disease identification

Rice is a primary global food source, but its production is severely threatened by pests and diseases, leading to significant yield losses and economic instability. Traditional manual identification is labor-intensive and error-prone. Deep Learning (DL) offers automated and accurate diagnosis, but its centralized deployment raises data privacy concerns and incurs high communication costs. Federated Learning (FL) provides a privacy-preserving alternative, but it faces challenges such as statistical heterogeneity (Non-IID data) and stringent resource constraints in the Internet of Agricultural Things (IoAT). To address these issues, we propose a personalized framework called Adaptive Parameter Decomposition Federated Learning (FedAD). FedAD employs a strategy based on Singular Value Decomposition (SVD) to decouple model weights into a shared full-rank component and a personalized low-rank component, and we introduce a dynamic rank allocation mechanism to adaptively adjust the personalized capacity for each client without requiring explicit SVD decomposition. In addition, we design a resource-efficient backbone model (SepResNet) to balance model performance with overall resource constraints in federated environments. To mitigate data scarcity and improve robustness, we construct a novel dataset (RP15) by employing a hybrid augmentation strategy that combines traditional transformations with generative AI synthesis. Experiments on both the RP15 dataset and the public PlantVillage dataset demonstrate the superior performance and generalization ability of FedAD. On the RP15 dataset, FedAD achieves 90.40% Accuracy, an 86.79% F1-Score, and 87.10% Recall, outperforming baseline methods by over 10% and doubling the convergence speed. On the PlantVillage dataset, FedAD achieves 99.45% Accuracy, a 99.04% F1-Score, and 99.12% Recall, significantly outperforming baseline models. Finally, we provide both experimental and theoretical validation of the framework's convergence, which ensures training stability and offers guidance for hyperparameter selection. The source code and RP15 dataset are available at https://github.com/DongfuHou/FedAD.