<p>The increasing prevalence of soybean foliar diseases poses a significant threat to agricultural productivity and food security, particularly in regions characterized by heterogeneous environmental conditions. Traditional centralized deep learning approaches, although effective, are constrained by data privacy concerns, limited data sharing, and poor generalization under non-identically distributed (non-IID) data scenarios. To address these challenges, this study proposes FedEMNet, a novel federated learning framework that integrates EfficientNet-B0 with a modified Model-Contrastive Federated Learning (MOON) strategy for robust and privacy-preserving soybean leaf disease classification. The model is evaluated on a multi-class dataset comprising eight disease categories, partitioned across 36 simulated clients using a Dirichlet distribution to emulate realistic non-IID conditions. The proposed framework employs a hybrid loss function that combines categorical cross-entropy with a contrastive loss mechanism to enforce alignment between local and global representations while mitigating client drift. Extensive experimentation demonstrates that FedEMNet achieves a high classification accuracy of 98.04 percent while significantly reducing communication overhead compared to conventional federated approaches such as FedAvg and FedProx. The integration of EfficientNet-B0 ensures computational efficiency suitable for resource-constrained edge devices, while the modified MOON framework enhances convergence stability and generalization across heterogeneous client distributions. Furthermore, the model incorporates data augmentation and class-weighted loss strategies to address dataset imbalance. The results validate the effectiveness of FedEMNet as a scalable and reliable solution for decentralized agricultural monitoring, offering a practical pathway toward intelligent, privacy-aware crop disease diagnosis in real-world farming environments.</p>