Environmental Monitoring

Featured Study: Air Pollution Monitoring in California

Environmental monitoring employs space-time analysis techniques to describe the evolution and fate of natural attributes. Monitoring networks provide observations of the main attribute, and possibly of contributing covariates. On the side of geostatistical challenges, analysis often involves large amounts of information that might be difficult or unfeasible to represent in traditional map forms.

With this study, we focused on tackling both issues. First, we analyzed 15 years of nitrogen dioxide and sulfate observations to predict their space-time distributions across California. Then, we combined this output with a spatialization technique to reduce dimensionality and explore patterns and relationships across the attributes.

 

About the Analysis

Our approach involves two stages. The first stage is a space-time statistical analysis with Bayesian Maximum Entropy (BME) to yield predicted nitrogen dioxide (NO2) distributions on the basis of the monitor observations. In the following figure, the left hand side animation shows the predicted mean monthly-averaged NOconcentration in the state of California for 2002. The units are in ppb, and the concentration scale ranges from 0 (white) to 76 ppb (dark brown) inside the state borders. The animation suggests increased NOconcentrations around the Los Angeles area that become moderately high in the first months and the fall of 2002. However, they remain overall lower than the absolute high of 76 ppb that was predicted for November 1995, earlier in the 15-year analysis period 1988-2002. Predictions are made for a regular grid of 504 nodes that covers the state area.

Space-Time Animation of Predicted Nitrogen Dioxide in California in 2002Cell-Normalized SOM by Using 15 Years of Predicted Nitrogen Dioxide Monthly Values

 

In the second stage we start with the predicted values, which are available at all nodes of the spatial grid and over the entire period of 15 years. This complete space-time coverage is essential to continue with an analysis of clustering and dimensionality reduction. Using the self-organizing maps (SOM) method, the predicted values undergo a spatialization transformation into a lower-dimensional geometric representation. In the above figure, the plot on the right illustrates an instance of this analysis, where predicted monthly NOconcentrations have been normalized by node cell. In cell normalization, it is the smallest and largest values for each cell over the 15-year span that drive the normalization. This perspective leads to a more direct comparison of temporal NOsignatures. It enables identification of broad regional patterns that affect the NOconcentrations in a similar manner within the same local neighborhood, regardless of the actual concentration magnitudes. In this context, different cell colors in the plot indicate the level of similarity to neighboring cells with respect to regional pattern behavior.

The SOM analysis enables additional types of normalization. Each of them can help reveal different attribute characteristics and patterns. For example, one can apply global or time normalization to the attribute. It is also possible to cross-examine different attributes as potential covariates. As proof-of-concept, our study includes a separate analysis of sulfate (SO4) in California in the same time period. SO4 is known to be a pollutant with different characteristics from NO2, and its concentration changes in different time scales from those of NO2. Despite the fact that SO4 makes an unlikely covariate of NO2, our study concludes that in their joint analysis SOM can still reveal patterns and relationships that exist within the different attributes.  

As a result, the combination of BME and spatialization techniques (BME-S) led us to a powerful, versatile spatiotemporal framework that is capable of studying a variety of attributes with a cognitively informed visualization. The proposed framework is also able to handle high-dimensional data and reduce their dimensionality in an adjustable manner to explore different perspectives; and it can be used as a tool to investigate potential covariate relationship among different attributes.

 

Reference

Kolovos A., Skupin A., Jerrett M., and G. Christakos. 2010. Multi-Perspective Analysis and Spatiotemporal Mapping of Air Pollution Monitoring Data. Envir Sci and Tech, 44(17), pp.6738-6744, doi: http://dx.doi.org/10.1021/es1013328