AQNet: Fine-Grained 3D Spatio-Temporal Air Quality Monitoring by Aerial-Ground WSN Yuzhe Yang, Zixuan Bai, Zhiwen Hu, Zijie Zheng, Kaigui Bian, and Lingyang Song School of Electrical Engineering and Computer Science, Peking University, Beijing, China Email: {yuzhe.yang, zixuan.bai, zhiwen.hu, zijie.zheng, bkg, lingyang.song}@pku.edu.cn Abstract—This demo presents AQNet, an aerial-ground wire- less sensor network (WSN) system, for fine-grained air quality monitoring and forecasting in urban three-dimensional (3D) area. AQNet contains 200 programmable on-ground PM 2.5 sensors for 2D baseline monitoring, and an unmanned-aerial-vehicle (UAV) with the same sensor for air quality profiling at different heights. These low-cost sensors are programmed to wake up between adjustable time intervals, record and send real-time PM 2.5 data back to the central server for data fusion. A learning model is proposed to utilize the data in both spatio-temporal perspectives to estimate PM 2.5 at unmeasured locations and forecast the air quality distribution in the near future. Further, the collected data is also used to control and optimize the UAV’s monitoring operation. For the convenience of user queries, we present the PM 2.5 map by a website-based GUI for real-time visualization. AQNet has been realized and deployed on campus of Peking University, and is scalable and energy-efficient to be extended to larger and more dedicated areas. I. I NTRODUCTION According to the World Health Organization (WHO), fine particulate matter (PM 2.5 ) has been identified as the most harmful particles to public health. A higher PM 2.5 can cause harmful health effects for people. Thus, monitoring PM 2.5 becomes a critical issue. The more accurate PM 2.5 distribution that can be obtained in a region, the more effective methods we can find to deal with the air pollution. PM 2.5 monitoring can be summarized into two categories: • Ground stations: A few monitoring stations are set up on dedicated sites in a city [1]. However, these fixed stations only provide coarse-grained 2D monitoring, with several kilometers between two measurement spots. Existing study shows that PM 2.5 distribution has intrinsic variation within meters, and it is preferred to monitor PM 2.5 in a specific 3D space (e.g., surrounding an office building), rather than city-wide [2]. The “meter-sliced” fine-grained PM 2.5 distri- butions would be more desirable for urban residents. • Mobile devices: Recent studies proposed to use phones, cars, balloons or drones [3] to monitor the 3D fine-grained air quality. However, they cannot do long-term monitoring due to the high energy consumption. Moreover, without 2D ground baseline, the control and operation becomes hard. In this demo, we present AQNet, an aerial-ground wireless sensor network (WSN) system, to monitor and forecast fine- grained PM 2.5 distribution in spatio-temporal perspectives. AQNet contains 200 on-ground devices for 2D baseline moni- toring, and a UAV for vertical distribution profiling. The aerial- ground WSN is a hybrid network that (1) leverages the data collected from on-ground sensors to control and optimize the UAV’s operation; and meanwhile (2) extends the on-ground monitoring to 3D space by aerial onboard sensors. To process real-time data, we set up a central server, where a learning- based model is designed for PM 2.5 estimation at unmeasured locations and future air quality prediction. Further, we also design a website-based GUI to present the PM 2.5 map, which visualizes historical, present and future air quality distribution. AQNet has been deployed on campus of Peking University, and can be easily extended to larger and more dedicated areas owing to the high scalability and energy-efficiency. Sensor Nodes on UAV and Ground Aerial-Ground WSN Monitoring System Air Quality Data Raw Database Other Features Collection (e.g. wind) Air Pollution Data Collection Layer Select training data in most recent time stamps S-kNN T-kNN Training Set Spatial Temporal Euclidean Distance Dynamic Time Warping Top k at current time stamp Top k at previous Tmin time stamps Feature Normalization Filling in Missing data Data Visualization Layer Website Based GUI UAV Control & Operation Deep Neural Networks Outputs Saliency Analysis Data Analysis Layer Fig. 1. The overall architecture of AQNet. II. SYSTEM OVERVIEW Fig. 1 illustrates the three-layer architecture of AQNet: data collection, data analysis, and data visualization. A. Data Collection Layer On-ground Nodes: On-ground PM 2.5 monitoring is based on 200 programmable monitoring devices, each of which contains a low-cost laser-based digital PM 2.5 sensor, a two-layer circuit board and a fixed shell structure. These devices can capture real-time PM 2.5 value within ±3% monitor error, and send them back to the central server for further data analysis. To realize high energy-efficiency, the devices are programmed to sleep during most of the time and wake up for data collection based on adjustable time intervals that controlled by the server. Aerial Nodes: A UAV carrying sensors is used for vertical profiling that extends 2D baseline monitoring into 3D space.