An edge-cloud collaborative model evolution and communication paradigm in Internet of low-altitude UAV

The low-altitude Internet of things (IoT), based on an air-ground integrated network, combines communication and computing functions. This allows it to efficiently collect, transmit, and analyze data in low-altitude scenarios, continuously empowering the development of the low-altitude economy. In this network, aerial platforms such as unmanned aerial vehicle (UAV) uses onboard sensors to gather multimodal perception data and perform AI-based data processing to support various low-altitude applications, such as agricultural monitoring and environmental modeling. Executing multimodal data inference and content generation tasks requires large AI models. To meet these demands, UAV needs powerful computing resources and vast data support, making efficient model training and optimization essential. However, this poses significant challenges to the current low-altitude IoT network. To address this, an integrated air-ground edge-cloud collaborative framework was proposed, where UAV function as edge nodes, collecting data and performing small-scale computations. Through wireless channels, cloud servers provide large-scale computations and update models for the UAV, enabling efficient collaborations. Given limited wireless communication bandwidth, the framework faces challenges in scheduling information exchange between the UAV and the cloud servers. To solve this, joint optimizations for task allocation, transmission resource management, data quantization, and edge model updates were presented, to improve inference accuracy by maximizing the mean average precision (mAP) of the proposed framework. A closed-form lower bound for the mAP based on the performance of the edge and cloud models were derived and a solution to mAP maximization was proposed. Simulations, based on visual classification experiments, show that the mAP of proposed framework under IoLoUA consistently outperforms centralized and distributed frameworks across various bandwidth and data conditions.