The Fog-enabled UAV-as-a-Service (FU-Serve) platform leverages distributed fog nodes to enable real-time data processing for multiple concurrent applications. However, the computational limitations of these fog nodes significantly hamper the execution of resource-intensive deep learning (DL) algorithms, compromising both operational performance and energy sustainability. To address these challenges, we integrated Federated Learning (FL) within the FU-Serve platform, coupled with the development of three specialized FL-based models tailored for on-device image classification. First, we introduced a sustainable adaptation of MobileNetV2 that synergizes Transfer Learning (TL) with FL principles. This model achieves 97.68% accuracy with an 8.64 MB footprint by distributing pre-trained weights to optimize bandwidth efficiency. To further address resource constraints of fog nodes, we designed FUSERNet—a lightweight DL architecture employing separable convolutions and skip connections, which reduces computational overhead while preserving critical feature representations. This model achieves 97.47% accuracy with an ultra-compact footprint of 237 KB, demonstrating a 98.59% reduction in size compared to state-of-the-art models. Finally, our third model, FusionNet, combines the strengths of MobileNetV2 and FUSERNet to deliver a balanced solution, achieving 97.75% accuracy with moderate resource requirements (8.86 MB). We evaluated our models on the AIDER and NDD disaster response datasets, our models demonstrate superior performance in classifying critical natural disaster scenarios. Notably, FusionNet matches SOTA accuracy levels while reducing memory consumption by 50%, and FUSERNet’s 0.23 MB size enables deployment on even the most resource-constrained UAVs. These contributions enhance the FU-Serve platform’s real-time decision-making capabilities, balancing computational efficiency and mission-critical accuracy for sustainable disaster response.