Phân Tích Ảnh Hưởng Của Phân Phối Dữ Liệu Không Đồng Nhất Trong Học Liên Kết Cho Bài Toán Phân Loại Bệnh Lá Lúa

Tóm tắt

This study analyzes the impact of heterogeneous data distribution on the performance of federated learning in classifying rice leaf diseases in the Mekong Delta, Vietnam. The dataset was constructed from publicly available sources combined with 2,000 field-collected images from various rice-growing regions. A deep learning model was initially used to detect and extract diseased areas to create a dataset refined for the classification task. To simulate the heterogeneity of data across rice-growing regions, a Dirichlet-based partitioning method was applied with different concentration parameters. This study conducts a comprehensive assessment of the impact of data heterogeneity and proposes an improved convolutional learning framework combined with LCSNet a more compact convolutional neural network model to optimize classification performance and reduce system resource overhead. Experimental results show that data imbalance significantly affects convergence and causes performance disparities among clients. However, the proposed method effectively addresses these challenges, and even under conditions of severe non-independence and homogeneity α = 0.5 FedPerGC still achieves reliable detection with F1 scores ranging from approximately 94.95% to 96.69%. These findings provide valuable insights into the behavior of linked learning systems in real-world data environments. Furthermore, they affirm the feasibility of balancing accuracy with computational cost when deploying on edge devices in agricultural settings in the Mekong Delta.