This paper introduces a two-stage method, combining K-means++ and a modified genetic algorithm (T-KMGA), aimed at simplifying collaborative task allocation complexity and enhancing operational efficiency within heterogeneous autonomous task allocation unmanned systems. The proposed T-KMGA constructs a solution for task allocation through two stages: task clustering and path planning. In the first stage, we develop a modified K-means++ clustering model sensitive to unmanned aerial vehicle (UAV) endurance radii to tackle the increasing complexity of task allocation as the scale of tasks grows. In the second stage, we propose a modified genetic algorithm to plan paths for multiple UAVs and a single unmanned surface vehicle (USV) to minimize the execution time of cooperative tasks. We adopt a virtual node encoding method to solve the multiple UAV task allocation problem in the second stage. The effectiveness of the proposed approach is assessed through simulation experiments. The results demonstrate that the proposed method can significantly reduce task execution time and computational consumption, especially in large-scale task scenarios.