City-Focused Air Quality Inference Learning
Board Location: #86
Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Session: 2
Benjamin Barrera-Altuna - University of South Florida
Co-Author(s): n/a
Knowledge of real-time air quality levels, as indicated by particulate matter (PM2.5), plays a crucial role in alerting the public about potential health issues from the surrounding environment. Although information on air quality levels is vital, only a few air quality monitoring stations have been deployed throughout the United States. To combat the lack of air quality data, auxiliary features such as temporal and spatial data within an area have been used in previous machine learning research to forecast it. In our work, we aim to infer rather than forecast air quality in unknown regions via a semi-supervised graph-based approach.In our initial study, our main focus is Dallas, Texas. To extract existing air quality data, we used data from PurpleAir (private company) sensors. Half a year’s worth of data (temperature, humidity, pressure, and PM2.5) in hourly intervals was collected. For spatial data, we used OpenStreetMap, an open database on city features, to collect land use areas, road networks, and residential housing data. Once the data was collected, we focused on the center of Dallas and created a 3 by 3 grid-based representation in which each square (35.36 km^2) had associated spatial and temporal data (for squares that were missing temporal data, we averaged the data from the 3 nearest squares). Within our approach, the model has two main processing phases: temporal and spatial. To process the temporal data (excluding PM2.5), we used a gated recurrent unit (GRU) in which each node’s data was batched (size 8) and sequenced (size 3). Afterwards, the final hidden state of each node’s batch was appended with the spatial data. Finally, for all processed nodes, a graph convolutional network was used to output an air quality (PM2.5) prediction at each time step. Only nodes that had data of their own (not averaged) were considered labeled (5 out of 9 nodes). For the optimization/loss function (mean squared error), all but one of the labeled nodes were used. The single node, which had labeled data but was not used in the loss optimization, was used as a control validation. To train the model, we used 70% of the data for 300 epochs. Using the remaining 30% of data as testing, the model achieved RMSE 12.1256 and MSE 9.6305. Upon a graph comparison of the model’s prediction and real data, our model is able to capture the general trends of PM2.5; however, it does underestimate most spikes in PM2.5. Future steps now include further increasing the nodes (regions) in the graph and understanding which auxiliary features are most important to our model’s prediction.Sources:Lin, Y. et al. 2018. Spatiotemporal patterns for air quality forecasting. Proc. 26th ACM SIGSPATIAL. 359-368.Iskandaryan, D., Ramos, F. & Trilles, S. 2023. Graph Neural Network for Air Quality. IEEE Access, 11:2729-2742.Zheng, Y., Liu, F. & Hsieh, H.P. 2013. U-air: Urban air quality & big data. Proc. 19th ACM SIGKDD. 1436-1444.
Funder Acknowledgement(s): Lehigh University: Accelerator Grant
Faculty Advisor: Yu Yang, yuy421@lehigh.edu
Role: Being the lead undergraduate researcher, my work for this project encapsulates all aspects of it. My contributions are as follows, in chronological order: I conducted an extensive literature review on the subject, created data collection scripts for PurpleAir and OpenStreetMap, verified and cleaned all data collected, and implemented and developed all of the deep learning model.It should be noted that a significant amount of time was spent on other data collection ventures that were not usable/implemented.