Plant disease detection models trained on centralized datasets often fail to generalize across regions due to variations in climate, crop varieties, and imaging conditions. This paper introduces an Adaptive Federated Learning (AFL) framework that enables cross-regional plant disease detection without centralized data collection. Using lightweight Convolutional Neural Networks (CNNs) as local models, the framework adaptively adjusts to regional data heterogeneity by incorporating personalization layers and dynamic aggregation strategies. Experiments on a partitioned PlantVillage dataset simulated across five agro-climatic regions show that our method improves average classification accuracy by 12.6% over standard FedAvg and reduces communication overhead by 38%. The proposed AFL framework demonstrates scalable and privacy-preserving plant disease diagnosis across geographically distributed farms.