Privacy-Preserving Federated Land-Cover Segmentation Under Non-IID Data: Region-Aware Contrastive Regularization and DP-SGD

Federated learning offers a practical route for collaborative land-cover mapping when satellite imagery cannot be pooled centrally, yet two issues remain: non-IID client data can weaken segmentation and shared model updates can still leak information. This study evaluates a privacy-preserving federated semantic-segmentation framework on DeepGlobe. The 803 labeled images were split at source-image level into 562/120/121 train/validation/internal-test images; each image was padded to 2560 × 2560 and divided into 25 non-overlapping 512 × 512 patches. Five configurations were compared: centralized U-Net, vanilla FedAvg, FedAvg with client-local region-aware contrastive regularization, FedAvg with client-side DP-SGD, and a combined private-contrastive variant using fixed data-independent prototypes. Under one fixed five-client Dirichlet partition (α = 0.5), contrastive FedAvg achieved the highest observed non-private validation performance (mIoU = 0.4062; Macro F1 = 0.5344) versus vanilla FedAvg (0.3712; 0.5011). DP-SGD provided formal patch-level (ε, δ)-Differential Privacy under the stated assumptions, with maximum ε = 4.8201 at δ = 10-5, but mIoU fell to 0.1300 and training became substantially slower. The combined private variant reached mIoU = 0.1362 at ε = 4.9958. These single-run validation results reveal a clear privacy-utility-computation trade-off.