FedDyMNv4++: Lightweight Dynamic MobileNetV4 with Adaptive Federated Learning for Remote Sensing Image Change Detection

Facing the challenges of model efficiency, client device resource constraints, and data volume imbalance in federated learning scenarios for remote sensing change detection, this paper proposes the FedDyMNv4++ method. The method comprises a lightweight local model DyMNv4++ and an Adaptive Two-Phase Data-Aware Learning Rate Scheduler (ATP-DALR).DyMNv4++ achieves cross-layer feature fusion and efficient change modeling through the Multi-Scale Fusion & Enhancement Network (MSFENet), Cosine-Aware Channel Difference Module (CACD), and Depthwise Dynamic Kernel Fusion ($\mathrm{D}^{2} \mathrm{KF}$). The ATP-DALR scheduler dynamically adjusts client learning rates to balance model convergence speed with stability while mitigating client data imbalance. Experiments on the LEVIR-CD dataset demonstrate that FedDyMNv4++ achieves comparable performance to centrally trained DyMNv4++ under federated training, with Recall, F1, and IoU improving by $16.33 \%, 7.61 \%$, and 8.03% respectively compared to centralized ResNet18_S5, while reducing parameters by $\mathbf{5 8. 8 \%}$ ($\mathbf{4. 8 8 M}$ vs $\mathbf{1 1. 8 5 M}$). This work provides an effective federated learning solution for remote sensing image change detection.