Authors: Weitian Tong*, Department of Computer Sciences, Georgia Southern University, Lixin Li, Department of Computer Sciences, Georgia Southern University, Xiaolu Zhou, Department of Geology and Geography, Georgia Southern University
Topics: Climatology and Meteorology, Environmental Science, Quantitative Methods
Keywords: Spatiotemporal interpolation, Air pollution, Deep learning, Recurrent neural network
Session Type: Paper
Start / End Time: 2:00 PM / 3:40 PM
Room: Southdown, Sheraton, 4th Floor
Presentation File: No File Uploaded
An accurate understanding of PM2.5 in a continuous space-time domain by spatiotemporal inter- polation is critical for meaningful assessment of the quantitative relationship between the risk of lung cancer and exposure to PM2.5. However, two major challenges exist. First, PM2.5 data are collected at a limited number of monitoring stations and there have not been interpolated PM2.5 data available in public over a continuous space and time domain in a large scale. Second, existing spatiotemporal interpolation algorithms are usually based on unrealistic assumptions by restricting the interpolation models to the ones with explicit and simple mathematical descriptions, thus neglecting plenty of hidden yet critical influence factors. Deep learning methods can extract high- level, complex abstractions as data representations through a hierarchical learning process, and therefore can be perfect candidates as black-box approaches to automatically consider the hidden factors and build the model for air pollution data. Among various deep learning methodologies, recurrent neural network (RNN) is particularly suitable for time series forecasting and modeling, because it not only considers the current input but also takes into account a trace of previously acquired information via recurrent connections that allow a direct processing of temporal dependencies and other hidden correlations. This project aims at developing efficient RNN-based spatiotemporal interpolation algorithm and generating the more accurate estimation of PM2.5 on a large geographic scale and over a long time period. The experimental results demonstrate the efficiency and effectiveness of our novel algorithm.