Integration of Remote Sensing and Social Sensing Data in a Deep Learning Framework for Hourly Urban PM2.5 Mapping

Abstract

Fine spatiotemporal mapping of PM2.5 concentration in urban areas is of great significance in epidemiologic research. However, both the diversity and the complex nonlinear relationships of PM2.5 influencing factors pose challenges for accurate mapping. To address these issues, we innovatively combined social sensing data with remote sensing data and other auxiliary variables, which can bring both natural and social factors into the modeling; meanwhile, we used a deep learning method to learn the nonlinear relationships. The geospatial analysis methods were applied to realize effective feature extraction of the social sensing data and a grid matching process was carried out to integrate the spatiotemporal multi-source heterogeneous data. Based on this research strategy, we finally generated hourly PM2.5 concentration data at a spatial resolution of 0.01°. This method was successfully applied to the central urban area of Wuhan in China, which the optimal result of the 10-fold cross-validation R2 was 0.832. Our work indicated that the real-time check-in and traffic index variables can improve both quantitative and mapping results. The mapping results could be potentially applied for urban environmental monitoring, pollution exposure assessment, and health risk research.

Short-Term PM2.5 Concentration Prediction by Combining GNSS and Meteorological Factors

“…So for long-term prediction, adding ZWD cannot perform well. From section I, the accuracy of various hourly PM2.5 concentration monitoring and prediction models is basically between 20μg/m 3 -60μg/m 3 , due to the difference between mathematical models and external factors (such as region, season, climate, time period, PM2.5 source, PM2.5 concentration change range) [21]- [26], [31], [32]. And based on SVMR method with metabolic method, using ZWD and meteorological factors to online predict PM2.5 concentration can be substantially better than 60μg/m 3 within 3 hours and better than 80μg/m 3 within 6 hours in this paper.…”

“…Therefore, geostationary satellite (for example Himawari-8), which can obtain hourly AOD, is used to improve the time resolution of PM2.5 concentrations estimation with satellite remote sensing method [24], [25]. Shen et al [26] use a deep learning method to build nonlinear hourly PM2.5 concentration mapping model by combining social sensing data with remote sensing data and other auxiliary variables, which can bring both natural and social factors into the modeling. However, AOD is easily affected by clouds, and it is difficult to be predicted with high time resolution because of too many influencing factors.…”

Deep learning for processing and analysis of remote sensing big data: a technical review

“…Such methods also respond to the requirements of high precision in the processing and analysis of RSBD. (2) Using multilayer structures, deep learning networks transfer multimodal RSBD into similar feature spaces, presenting great capabilities for data fusion (Shen et al, 2019;Zhou et al, 2020). This provides new ideas for cooperative analysis of multisource, multimodal, multiscale, and heterogeneous RSBDs.…”

“…Taking advantage of the reliability of RS data and the spatial-temporal dynamic characteristics of sensor data, Cao et al (2020) proposed an end-to-end deep-learning-based approach to fuse remote and social sensing data to recognize urban region functions in high-density cities using CNNs and LSTM networks. Shen, Zhou, and Li et al (2019) used a deep belief network to learn the complex relationships among RS data, social sensing data, meteorological data, and the spatial-temporal features of PM 2.5 for finer spatial-temporal mapping of PM 2.5 concentrations in urban areas.…”

Environmental Health Surveillance System for a Population Using Advanced Exposure Assessment

“…Some researchers have investigated the spatial distribution of air pollutant concentrations from the geo-statistics perspective based on actual observations. This is due to the increasing number of fixed air monitoring stations and the greater availability of low-cost sensors for continuous spatio-temporal air quality monitoring [71][72][73]. For example, a novel application based on the optimal linear data fusion method was applied in combination with the kriging interpolation technique for data fusion between different types of PM 2.5 sensors [35].…”