Using satellite data to develop environmental indicators

Poor air quality is a major concern worldwide. The World Health Organization (WHO) estimates that as many as 1.4 billion urban residents around the world breathe air with pollutant levels exceeding the WHO air quality guidelines. According to Lim et al (2012), outdoor air pollution is a major contributor to the global environmental burden of disease. The WHO indicates that it causes close to one million premature deaths worldwide each year, with particulate matter as a leading contributors (Ostro 2004). As such, air quality has garnered the attention of policymakers worldwide. While several air pollutants have adverse health effects, PM_2.5 (microscopic particles less than 2.5 μm in diameter that lodge deep in the lungs) is widely recognized as the worst, owing to its potential to contribute to respiratory and cardiovascular disease in exposed populations.

Understanding air pollution levels, sources, and impacts is a critical first step in addressing air pollution problems (Hsu et al 2013). However, ambient air quality monitors are relatively sparse in many parts of the developing world, and even where they exist, only represent air quality conditions in the immediate vicinity of a monitor (Gutierrez 2010). There are significantly more pollutant monitors within North America and Western Europe (approximately 4100 monitors) than collectively in the rest of the world (Engel-Cox et al 2012). This disparity represents a significant in situ data gap in much of the developing world that hinders inclusion of air quality in global environmental indicators.

Satellite remote sensing air quality datasets offer the potential to fill that gap (Hsu et al 2013, Martin 2008). Based upon AG consultations, the project team focused principally on PM_2.5, given the health impacts and recent advances in correlation of satellite data with surface level PM_2.5 concentrations. (For a discussion of other satellite-derived air quality indicators proposed to the AG, see the SOM.) A primary satellite-derived dataset relevant to PM_2.5 is Aerosol Optical Depth (AOD), a measurement of scattering of light between the satellite and ground surface. AOD is available from a number of satellites worldwide, including NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) and Multi-angle Imaging Spectroradiometer (MISR) instruments on the Terra satellite. Numerous studies have shown that AOD is proportional to PM_2.5 (e.g., Hoff and Christopher 2009, Weber et al 2010) and can be used as a surrogate dataset to fill the spatial gaps of ground-based monitoring networks.

Two related methodologies were developed for calculation of the air quality indicator. Here we briefly describe the first method, which relied on an existing dataset of satellite-derived surface level PM_2.5 concentrations for 2001–2006 (van Donkelaar et al 2010) to craft a policy-relevant pilot indicator that accounts for population exposures and is geographically aggregated. Details on the second method, which used conversion factors from van Donkelaar et al to process time series particulate matter grids based on MODIS/MISR data, are presented in the SOM.

van Donkelaar et al (2010) generated a global, satellite-derived, annual average PM_2.5 surface for the years 2001–06 based on level 2 (L2), daily MODIS and MISR satellite instrument AOD (figure 1). The results were validated against coincident and non-coincident ground-based measurements of PM_2.5 with high levels of agreement. The geographic coverage of the dataset is nearly global with a horizontal resolution of 0.1° × 0.1°. Coincident aerosol vertical profiles from the GEOS-Chem chemical transport model, validated with CALIPSO space-borne lidar vertical profiles, were used to calculate daily conversion factors that account for the relationship between satellite column AOD and surface PM_2.5 concentrations. The AOD–PM_2.5 relationship varies spatially and temporally due to global and seasonal variations in aerosol size, aerosol type, relative humidity, and boundary layer height. The conversion factors were applied by van Donkelaar et al to the MODIS/MISR AOD data to estimate surface-level PM_2.5 concentrations, measured in micrograms per cubic meter (μg m⁻³).

Zoom In Zoom Out Reset image size

A population-weighting was applied to the surface-level PM_2.5 concentrations in order to create an exposure indicator with human health policy relevance. Population weighting simply gives greater weight to PM_2.5 concentrations in more populated areas than in less populated areas, thereby ensuring that measures tied to administrative units are not biased by large sparsely populated areas with relatively clean air. To implement the population-weighting, the fraction of the total country population within each grid cell of a country was determined using the Global Rural–Urban Mapping Project (GRUMP), v1 population dataset representing the year 2000 (Center for International Earth Science Information Network (CIESIN)/Columbia University, International Food Policy Research Institute (IFPRI), The World Bank, and Centro Internacional de Agricultura Tropical (CIAT) 2011). The population-weighted PM_2.5 indicator was then calculated as the country sum of the product of the estimated PM_2.5 concentration for a grid cell and the fraction of the population within that grid cell⁴ . The result represents an average concentration of PM_2.5 to which a country's population is exposed.

The results show patterns of PM_2.5 concentrations above WHO guidelines of 10 μg m⁻³ annual average PM_2.5, as shown by the colored areas in figure 2. High concentrations are found in several European countries, much of dryland Africa, the Middle East, and South and Southeast Asia. Particularly high average exposure is found in China (56 μg m⁻³, more than five times the WHO concentration guideline). One hundred out of 156 countries have concentrations above WHO guidelines for PM_2.5. At the recommendation of the AG, average exposure was also calculated at province/state levels to highlight concentration differentials particularly within large, densely populated countries (SOM, figure 1).

Zoom In Zoom Out Reset image size

The limitations of air quality satellite measurements include deriving surface-level concentrations from column measurements, understanding spatial patterns at scales finer than the native satellite spatial resolution (e.g., 10 km² for daily MODIS AOD although 3 km² products are becoming available), dealing with missing data, and maintaining continuity of measurements as satellites degrade and are replaced. While the team developed approaches to address most of these limitations, the AG felt that the satellite-based air quality indicators are more properly understood as a communication tool, and would not be useful per se in identifying mitigation strategies. This indicator does not have sufficient detail to support development of pollution control strategies since it does not address critical issues such as pollutant transport and chemistry. However, it should be noted that the use of the indicator for communication is consistent with our definition of a policy-relevant indicator.

AG engagement provided a number of additional insights on the suitability of satellite data for indicator development. The requirements for indicators to be scientifically rigorous, policy relevant, transparent, and sustainable underlie the suitability of satellite data (and particularly, air quality data) for environmental indicators. The two methods tested, the second of which involved using time series satellite AOD data with simplified assumptions, illustrate trade-offs among these criteria for alternate methods of applying satellite-based air quality data. Method 1 employs higher temporal and spatial resolution data than Method 2, more accurately capturing spatial and temporal variations in pollutant levels. Method 1 also relies on computing resources and knowledge of chemical transport models. Method 2 can be calculated by a geographic information system expert with internet access and a detailed methodology including a file of AOD/PM_2.5 ratios and filters, and can be developed as a time series and routinely updated⁵ . Method 2 thus has higher sustainability in terms of lending itself to self-calculation by non-experts.

AG members cited the complex relationship between satellite-derived AOD and surface PM_2.5 exposure—especially as it relates to disease risk—as a potential factor limiting indicator readiness. Also, one AG member noted that if the observation time of the satellite does not correspond to peak PM_2.5 concentrations locally, then the corresponding satellite data may not provide the best representation of population exposure to air pollution. Although this problem may not be significant on a global scale, it should be addressed in future analyses to prevent potential misinterpretation or 'over-trusting' of the indicator results by decision makers who are not familiar with the details of the satellite dataset.

There are trade-offs between the use of satellite data and other forms of air quality monitoring and modeling outputs to support environmental indicators. Given that air quality monitors are relatively sparse in many parts of the developing world, satellite data helps fill a significant in situ data gap that hinders inclusion of air quality in global environmental indicators. The amount of labor required to operate a few monitors in a city is virtually the same as the labor required to calculate a satellite-based indicator for the entire world, on the order of a few hundred hours per year. Ground-based air quality monitors can cost tens of thousands of US dollars for initial purchase and several thousand US dollars for operation and maintenance. They also require well-functioning institutions and trained technicians to produce consistent, high quality data. Development of satellite derived estimates requires considerably less investment and expertise. Possible next steps are described in the SOM.

Biomass burning has a number of environmental impacts, including greenhouse gas emissions, health impacts from fire-related pollution (Johnston et al 2012), and ecosystem effects (Nepstad et al 1999). Some 20% of global emissions of greenhouse gases are due to fire activity (Bowman et al 2009), and the resulting emissions of black carbon are a major contributor to climate change and human health issues (Shindell et al 2012). Research suggests that climate change is contributing to an increase in fire activity in some regions (Westerling et al 2009, Flannigan et al 2009), resulting in a positive feedback loop of forest drying contributing to fires, which contributes to greenhouse gas emissions and ultimately more climate change and fires.

For the purposes of this indicator, it was assumed that most fire activity is either directly or indirectly due to human activity. An example of the former would be fires intentionally set to clear land for agriculture or to prepare land for cultivation, or accidentally set owing to failure to contain a fire in combination with drought conditions. An example of the latter would be forest fires produced by lightning strikes in areas where land clearing processes or climate change has resulted in a substantial alteration to the moisture content of the vegetative cover. Although the evidence for the human influence on fire activity is mixed (see for example Krawchuk et al 2009 and Le Page et al 2010), in the context of changing land use and climate conditions there is reason to believe that trends can be ascribed to human influences. For example, in a global assessment of fire activity, Chuvieco et al (2008) found a high association between population distribution and fire persistence. Therefore, our indicator focused on patterns and trends (spatial and temporal variation) in biomass burning with an emphasis on emissions and ecosystem impacts, with a goal of identifying countries with either increases or decreases in biomass-burning related emissions. Decreases in emissions would presumably signal countries that are improving their fire activity management over time, although the potential for major influence from short term climate variability or longer term climate change cannot be ignored as a contributing factor to fire activity. We address this in section 4.4.

For this indicator the AG recommended that we leverage prior work in assembling the Global Fire Emissions Database 3 (GFED3), a 13-year validated time series gridded emissions database (1997–2009) (van der Werf et al 2010). GFED3 is summarized on a half degree grid, though the emissions estimates come from higher resolution underlying burned area and active fire data. For the time period since 2000, more than 90% of the global burned area was mapped using MODIS 500 m resolution data. GFED3 emissions estimates (annual estimated tons of carbon per pixel) are further subdivided by fire type within every half degree cell: deforestation and degradation fires, savanna fires, woodland fires, forest fires, agricultural waste burning, and tropical peatland fires.

The following steps were followed to create the indicators. We added total emissions from forest, deforestation and degradation, woodland, savanna, and tropical peatland fire emissions; we omitted agriculture waste burning because this reflects seasonal fires that tend to have low net carbon emissions (crops quickly grow back resulting in re-absorption of carbon dioxide from the atmosphere). In ArcGIS 10 we created annual grids of emissions for these biomass burning types. Using R, we produced a cluster map on a gridcell basis depicting areas where there is both high frequency and intensity (total) of emissions (results presented in SOM). We regridded the data set at 2.5 min (∼4 km at equator), so that it would better correspond to the CIESIN country boundary data, and then we produced total emissions by country and year.

The country-based indicator of average annual emissions over the last five years of the time period shows very high emissions in the United States, northern South America, Central Africa, Australia, India and China (table 2). Figure 3 depicts emissions divided by land area. Among countries with temperate climates the US, Canada, Russia, and Australia are major emitters per land area, as are most countries with dense tropical forests with the exception of Liberia, Gabon and Congo-Brazzaville.

Zoom In Zoom Out Reset image size

We sought to determine if there were significant trends in country level emissions over the ten year period. We could not identify any significant trends in the slope of ton of carbon emissions by year given the small sample size (n = 10), but there are a number of countries that saw large percentage increases and decreases in emissions when comparing the first half of the period to the second half (figure 4). Brazil, Peru and Bolivia saw large increases of 37%, 20%, and 73%, respectively, presumably owing to recent droughts and increases in forest fire activity in the Amazon Basin over this period. Algeria saw a near doubling in emissions. Much of South Asia is also a hotspot of increased burning between the two time periods.

Zoom In Zoom Out Reset image size

An original goal of this indicator was to determine which countries are improving their fire management over time, as signaled by declining emissions. This goal turns out to be difficult to achieve. Although human land management activities are indeed important and can affect overall emissions, year-on-year variations are more likely to be driven by climatic factors. In work not presented here, we sought to identify statistically significant trends in emissions, but given the short time series (n = 10) and high inter-annual variability, very few countries had significant trends. The percent change map (figure 4) could be used to conduct further analysis regarding the possible impact of management, positive or negative. Normalizing emissions by rainfall could potentially result in an indicator that more effectively identifies areas where land management activities are contributing to increasing trends. One additional issue with the trend approach is that forest burning depletes forest biomass, so there may be an upper asymptote beyond which biomass burning tends to level off. In other words, declining emissions could mean a country has largely been deforested.

A second goal of pinpointing areas of high biomass burning for policy attention was partially accomplished, though it would be good to use the cluster analysis results (presented in the SOM) in combination with information on the type of fire activity (forest, agriculture, peatland, etc) to identify fire 'hotspots' that might link to potential policy responses. This relates to a critique by the AG, that the indicator conflates a number of separate policy issues, including deforestation and land degradation (in some cases not sustainable); sustainable agriculture and land use processes; and peatland burning. It also does not adequately address the nature of burning activity in geographically disparate regions such as the boreal and tropical forests. The types of burning and locations where burning is taking place will require different policy responses, and the degree to which the indicator is relevant to a policy process depends on separating these out.

Overall, the AG identified this as an indicator with high salience (owing to concern over GHG emissions) and high robustness that was ready, once some of the above issues are addressed, to be applied in policy contexts.

The flow of nutrients into coastal waters from land-based sources has seen a worldwide increase over the last decades (Boesch et al 2009, Glibert et al 2008). A 20 year analysis (1980–2000) of Coastal Zone Color Scanner (CZCS) and Sea-viewing Wide Field-of-view Sensor (SeaWiFS) data found dramatic increases in global average chlorophyll concentrations of close to 22%, particularly in the southern hemisphere and intertropical regions (Antoine 2005, Antoine et al 2005). The resulting change in water quality has many potential impacts on coastal and marine ecosystems. Phosphorus and nitrogen contribute to enhanced algae growth, and subsequent decomposition reduces oxygen availability to benthic sea creatures like fish, shell fish, and crustaceans. Changes to nutrient loadings can also change the phytoplankton species composition and diversity. In extreme cases, eutrophication can lead to hypoxia or oxygen-depleted 'dead zones' (Falkowski et al 2011) and harmful algal blooms, which have been spreading (Diaz and Rosenberg 2008, Kahru and Mitchell 2008, Glibert et al 2008). Yet in situ monitoring systems are even more sparse than for atmospheric pollution concentrations (EPA (US Environmental Protection Agency) 2008).

Given high variability in chlorophyll concentrations in the coastal zone depending on the type of coastal waters (e.g., estuaries or upwelling areas), and even within short distances in the same waters, it is not easy to determine a suitable target or threshold for 'harmful' chlorophyll concentrations similar to those that can be established for air pollutant concentrations. Furthermore, bottom reflectance can affect acquisition of ocean color parameters, rendering a snap shot in time relatively meaningless. Thus, in consultation with the AG, our approach to measuring coastal water quality was to measure trends in chlorophyll-a (chl-a) concentrations from 1997–2007, with the assumption that any increase in concentrations is likely to signal a negative change, which is consistent with the approach taken by others (Antoine 2005, Antoine et al 2005). While this assumption may not hold in every coastal area, on a global scale there can be little doubt that widespread increases would signal important changes in coastal water quality, which in turn may be related to patterns of eutrophication or harmful algae blooms. We further assume that changes in chl-a concentrations are due largely to land-based practices such as over-fertilization, inadequate sewage treatment, or other nutrient sources, all of which are amenable to policy responses. The degree to which this presumption is justified is further addressed below.

Based on recommendations from the AG, we focused on SeaWiFS Level 3 monthly data acquired from September 1997 to December 2007. The date ranges represent the most reliable period of SeaWiFS acquisitions; the sensor was launched in 1997 and after December 2007 the instrument's reliability began to decline. This work built on an earlier pilot efforts in 2007, which examined trends in coastal chlorophyll concentrations using annual SeaWiFS composites from 1998–2007 (Center for International Earth Science Information Network (CIESIN)/Columbia University 2009)⁶ . The annual composites were deemed to be unsuitable because of the oversampling during periods with low cloud cover and undersampling during periods of high cloud cover. It is known that periods of coastal upwelling in the Pacific Northwest, for example, are correlated with periods of higher cloud cover, and hence the sample from which annual composites are drawn is biased.

The work required a number of data transformations. We limited the spatial extent of the SeaWiFS data based on country's Exclusive Economic Zones (EEZs) out to 100 km based on EEZ boundary data from the Flanders Marine Institute (VLIZ) Maritime Boundaries Geodatabase. We eliminated grid cells with less than 45 observations over the 124 months, which effectively excluded many areas of constant cloud cover, especially in the northern latitudes. We also eliminated monthly averages based on <2 observations. We exported the data to IBM SPSS predictive analytics software and ran a regression of chlorophyll-a concentrations against year with dummy variables for each month to account for seasonal variability. The end results was a slope coefficient which measures change in chlorophyll-a by year, as well as a significance level for the slope.

Figure 5 presents trend results for the United States and Mexico, while additional figures for China and the Mediterranean are presented in the SOM. All maps depict change in coastal chl-a in the form of slopes with chl-a as the dependent variable and year as the independent variable (controlling for seasonal variation with the month dummies). Declines in chl-a over the time period 1997–2007 are depicted in shades of green and increases in chl-a (potential problem areas) are depicted in shades from yellow to red⁷ . Although we were unable to perform a validation against external data sources, the results are nevertheless interesting. Several areas have seen significantly increasing trends in chl-a concentrations, including Long Island Sound, the Chesapeake Bay, coastal Texas, and much of the Pacific Northwest.

Zoom In Zoom Out Reset image size

Although pilot results were promising, the AG members representing the EPA and MCC found that this would not be easily used by their agencies because of the still advancing science on attribution and interpretation of results, and the high uncertainty in satellite-derived chl-a concentrations in near coastal regions. Ocean color sensors also detect colored dissolved organic matter (CDOM), and hence in areas experiencing heavy sediment loading increases, the signal needs to be interpreted differently. While AG members acknowledged that increases in chl-a and CDOM both signal declining water quality, addressing suspended sediments requires different policy responses than does reducing nutrient loads. Thus, more research would be required to understand patterns and specific causes (cf Beman et al 2005). For example, there were increases in the Chesapeake Bay, which is known to be influenced by nutrient loading from the Susquehanna and Potomac Rivers and urbanization (Kemp et al 2005), but large areas around the Mississippi Delta appear to have negative slope lines, which is somewhat counterintuitive.

The AG found that this indicator probably has limited relevance as a country-level comparative performance indicator. This is owing to the difficulties in (1) interpreting trends and attributing country responsibility for changes in coastal water quality, and (2) attributing changes to land-based activities. There is also no straightforward way to create a country level indicator representing the balance of positive and negative trends in the country's coastal area. Given a sufficiently long time series, one could conceivably measure the area of the coastal zone affected by 'negative' trends (positive slopes), but uncertainties are still high.

On balance, however, there was agreement among AG members that the map outputs can be useful for identifying changes that warrant further investigation, and may in some cases point to the need for policy responses.