<p>In recent years, UAV-assisted federated learning (FL) has emerged as a promising paradigm for disaster detection by combining the mobility of unmanned aerial vehicles with decentralized model training. However, in realistic disaster scenarios, FL systems face stringent spectrum and latency constraints, highly heterogeneous data distributions, and unfair client selection, which may result in incomplete disaster coverage and biased global models. To address these challenges, we propose a fairness-aware federated learning framework for UAV-assisted disaster detection that explicitly balances training latency, model accuracy, and client participation. Specifically, a multi-criteria client selection scheme based on an upper confidence bound (UCB) strategy is designed, in which a dynamic reward function jointly accounts for normalized latency, local accuracy, and selection fairness, with adaptive weight adjustment across training stages. Extensive experiments on the real-world aerial disaster dataset AIDER demonstrate that the proposed MCCS framework consistently achieves a better trade-off among global test accuracy, cumulative training latency, and performance fairness, as quantified by the Jain fairness index, compared with multiple representative client selection baselines. These results indicate that the proposed framework enhances the comprehensive performance of federated learning and is well suited for latency- and fairness-sensitive UAV-assisted disaster detection scenarios.</p>