U-shaped split federated learning (U-SFL) has been widely applied in image coding, as it can effectively balance parallel training, model privacy, and local computational cost. The performance of U-SFL largely depends on the selection of split points and the aggregation frequency, and thus, model splitting (MS) and model aggregation (MA) strategies are critical. In uncrewed aerial vehicle scenarios, fluctuating communication links and computational resources can significantly impact U-SFL performance. To address this challenge, we propose a resource-adaptive U-SFL (AU-SFL) framework, which adaptively selects optimal MS and MA strategies based on the available computational and communication resources. Specifically, we first conduct a theoretical convergence analysis that systematically quantifies the individual and joint impacts of MS and MA strategies on convergence behavior. Then, we formulate an optimization problem aimed at minimizing training latency, grounded in the theoretical convergence analysis. Subsequently, an alternating MS–MA algorithm is proposed to solve the problem, which is decomposed into two subproblems and solved alternatively. Extensive experiments on multiple benchmark datasets demonstrate that AU-SFL achieves a $1.61\times $ reduction in convergence time while improving multiscale structural similarity index (MS-SSIM) by 4.4%, conclusively validating our adaptive optimization strategy’s effectiveness.

U-shaped Split Federated Learning (U-SFL) is a promising paradigm for distributed image coding, offering parallel training capabilities and privacy preservation while mitigating computational burdens on edge devices. However, the frequent bidirectional transmission of intermediate features between dual-split points incurs substantial communication overhead. To mitigate this issue, we propose a compact-feature U-shaped split federated learning framework (CoF U-SFL), which reduces communication overhead and improves training efficiency while maintaining low image distortion. We introduce a feature entropy estimation network to model the distribution of split-layer features, enabling effective compression during transmission. Furthermore, we formulate a joint optimization objective incorporating entropy constraints to guide the end-to-end training. Experimental results demonstrate that CoF U-SFL reduces communication overhead by 104.6 times while maintaining reconstruction performance.