LAUSR

In the last decades, environmental pollution has grown up to be a major problem that influences people's health. Providing accurate environmental sensing services is of great significance. To realize environmental sensing, distributed monitoring sites are used to collect comprehensive long-term environmental data. However, sparse sensory data caused by insufficient monitoring sites and their incomplete records become the main challenge of fine-grained environment sensing. At the same time, due to the limitations of network bandwidth and storage space, traditional centralized training is difficult to meet the task training requirements based on big data. In this paper, we develop a novel distributed inference framework, named Federated Region-Learning (FRL) for urban environment sensing. It inherits the basic idea of federated learning avoiding transmission and centralized storage of data, and also considers the regional characteristics of each monitoring site. Through an elaborate designed edge computing system, a local regional model is customized for the micro cloud to improve the inference accuracy. Moreover, we develop two types of global model aggregation strategies to better target different bandwidth requirements. Extensive experiments based on two real-world datasets are performed to prove universality and effectiveness.