Crop-disease classifiers that pool raw images from farms and IoT devices violate the data-control and privacy expectations of individual growers. We propose a federated framework combining FedProx and FedPer that integrates (i) a ResNet-18 backbone with attention-gated multi-feature fusion of CNN, color-histogram, and texture features; (ii) FedProx proximal regularization for stable convergence under non-IID client data; and (iii) per-client personalization heads that never leave the device. The tomato-leaf subset of PlantVillage is split across five clients by Dirichlet allocation (α = 0.5), which makes the data strongly non-IID. On it, the framework attains 90.74% accuracy and 90.54% F1-score, which is 20.78 and 25.74 percentage points (pp) above local-only training and within 5.26 and 5.44 pp of a privacy-violating pooled baseline that gathers all client data on one server. In the ablations, removing FedProx regularization costs the most final-round accuracy (11.32 pp), followed by adaptive fusion (8.32 pp) and personalization (4.32 pp). Federation moves 5.6 times the network traffic of pooling (10.55 GB versus 1.88 GB) because the shared encoder is exchanged every round; arithmetic projections, not experiments, put this at 3.5× with 8-bit uploads. Results cover one crop and one lab-photographed dataset.