Indonesian fire activity and smoke pollution in 2015 show persistent nonlinear sensitivity to El Niño-induced drought

Significance

The 2015 Indonesian fire season, in terms of fire activity and pollution, was the most severe since the NASA Earth Observing satellite system began observations in the early 2000s. Our estimates show that the 2015 CO₂-equivalent biomass burning emissions for all of Indonesia were between the 2013 annual fossil fuel CO₂ emissions of Japan and India. Longer-term records of airport visibility in Sumatra and Kalimantan show that 2015 ranked among the worst episodes on record. Analysis of dry season rainfall shows that, due to the continued use of fire to clear and prepare land on degraded peat, the Indonesian fire environment continues to have nonlinear sensitivity to dry conditions, and this sensitivity appears to have increased over Kalimantan.

Abstract

The 2015 fire season and related smoke pollution in Indonesia was more severe than the major 2006 episode, making it the most severe season observed by the NASA Earth Observing System satellites that go back to the early 2000s, namely active fire detections from the Terra and Aqua Moderate Resolution Imaging Spectroradiometers (MODIS), MODIS aerosol optical depth, Terra Measurement of Pollution in the Troposphere (MOPITT) carbon monoxide (CO), Aqua Atmospheric Infrared Sounder (AIRS) CO, Aura Ozone Monitoring Instrument (OMI) aerosol index, and Aura Microwave Limb Sounder (MLS) CO. The MLS CO in the upper troposphere showed a plume of pollution stretching from East Africa to the western Pacific Ocean that persisted for 2 mo. Longer-term records of airport visibility in Sumatra and Kalimantan show that 2015 ranked after 1997 and alongside 1991 and 1994 as among the worst episodes on record. Analysis of yearly dry season rainfall from the Tropical Rainfall Measurement Mission (TRMM) and rain gauges shows that, due to the continued use of fire to clear and prepare land on degraded peat, the Indonesian fire environment continues to have nonlinear sensitivity to dry conditions during prolonged periods with less than 4 mm/d of precipitation, and this sensitivity appears to have increased over Kalimantan. Without significant reforms in land use and the adoption of early warning triggers tied to precipitation forecasts, these intense fire episodes will reoccur during future droughts, usually associated with El Niño events.

The 2015 fire season in Indonesia began in July in Sumatra and a month later in Kalimantan and was mostly confined to the part of the country in the Southern hemisphere. By September, much of Sumatra and Kalimantan were blanketed in thick smoke that lasted through October, with the haze extending to Singapore, Malaysia, and Thailand. Millions of people were exposed to hazardously poor air quality for 2 mo (1).

Fig. 1 shows the monthly Moderate Resolution Imaging Spectroradiometer (MODIS) active fire detections (described in the next section) between August and November 2015. This period comprised the bulk of the fire season, with 85% of total annual fire detections. September and October were the months with the highest number of active fire detections (68% of total). Most fires burned in the lowlands of southern Sumatra and Kalimantan, often in areas underlain by peat deposits. The locations of the fires and the progression of the fire season resembled 2006, but there were more fires in 2015 in the main fire-affected provinces except for western Kalimantan. The key difference with other years is in the amount of fire activity.

Fig. 1.

High confidence Aqua and Terra MODIS active fire density in 2015 (0.25° × 0.25°) over Sumatra and Borneo for (A) August, (B) September, (C) October, and (D) November.

The fire and haze in 2015 was a repeat of events that have occurred periodically in Kalimantan since the 1980s (2–6) and in Sumatra since at least the 1960s (7). From those studies, 1982/83, 1987, 1991, 1994, 1997, and 2006 can be considered severe fire years over Sumatra and Kalimantan, relative to years where little or moderate fire occurs because it is too wet during the dry season for sustained burning. Fires are set to clear logging waste, agricultural waste, and, to maintain or secure land tenure, regrowth (8, 9). The fires often occur on drained and degraded peat lands (10). During abnormally dry years typically associated with El Niño conditions, the peat becomes dry enough to burn (11). Fires on the surface can escape underground, where, because they are so difficult to extinguish and have a large source of fuel, they burn continuously until the return of the monsoon rains (12).

It is widely accepted that the worst event on record was in 1997, with the total CO₂ emissions equivalent to between 13% and 40% of mean annual global fossil fuel emissions at the time (11). The last major event occurred in 2006 over southern Sumatra and south-central Kalimantan, under a combination of moderate El Niño and positive Indian Ocean Dipole conditions (12). The 2006 burning episode registered uniquely in satellite measurements sensitive to pollution in the midtroposphere (13–15). In terms of extent and duration, retrieved CO in the upper troposphere in 2006 was the highest during the 2004–2011 observation record across the whole of the tropics (16). There have been brief episodes under isolated dry conditions that produced locally high pollution levels, for example, in the central Sumatran province of Riau in 2013 (10), but in general, fire activity and pollution levels have not approached those of 2006 in the intervening years.

To quantify the magnitude of 2015 compared with past events and understand the drought conditions under which they occurred, we analyzed data from the NASA Earth Observing System (EOS) period, namely MODIS active fire detections, five different satellite measurements of tropospheric pollution, airport visibility records as a longer-term proxy, and precipitation estimates from satellites and rain gauges. We make the case that the 2015 Indonesian fire season was the most severe season since the NASA’s Earth Observing satellite system began observations in the early 2000s, and, by examining visibility data before the EOS period, that 2015 ranked after 1997 and alongside 1991 and 1994 as among the worst Indonesian fire events on record.

Data

During the EOS period, we used midtropospheric CO data from the Terra Measurement of Pollution in the Troposphere (MOPITT) (17) and Aqua Atmospheric Infrared Sounder (AIRS) instruments (18) and upper-tropospheric CO data from the Aura Microwave Limb Sounder (MLS) (19). Aerosols were characterized using MODIS aerosol optical depth (AOD) over ocean (20) and land (21) and the Aura Ozone Monitoring Instrument (OMI) aerosol index (AI) (22). Fire activity was characterized by Terra and Aqua MODIS active fire detections (23). Details of these data are provided as SI Text.

At the surface, airport visibility is a useful indicator of severe fire emissions in Indonesia (7, 24–26) because of high emissions per unit area burned and poor ventilation due to typically gentle surface winds. Visibility records were obtained for World Meteorological Organization (WMO)-level surface stations located at three airports in each of southern Sumatra (Rengat, Jambi, and Palembang) and south-central Kalimantan (Pangkalan Bun, Palangkaraya, Muaratewe) from the NOAA Integrated Surface Database for 1990–2015. We computed the total extinction coefficient (B_ext) from the visibility using the empirical Koschmieder relationship, B_ext = 1.9/v, where v is the visibility in kilometers (7). For the sake of computation, reports of zero visibility during the worst of the haze were replaced with 0.05 km, the next lowest reported value.

Precipitation estimates for 2000–2015 were obtained from the Tropical Rainfall Measurement Mission (TRMM) (27) 3B42RT product, which is produced using a consistent retrieval, but lacks radar data assimilation after mid-2015. Strictly gauge-based precipitation estimates were obtained for 1990–2015 from the NOAA Climate Prediction Center’s global daily precipitation dataset (28).

The MODIS active fire detections, precipitation, and extinction coefficient from surface visibility were analyzed over the primary burning regions in southern Sumatra (6°S–0°, 99°E–106°E) and south-central Kalimantan (4°S–0°, 110°E–117°E). The OMI AI, MODIS AOD, AIRS CO, MOPITT CO, and MLS CO were analyzed separately over Sumatra (10°S–10°N, 90°E–105°E) and Kalimantan (10°S–10°N, 105°E–120°E) to include the larger regions affected by the smoke. A broader pollution signature is seen in the AIRS, MOPITT, and MLS CO, but this will require a more thorough examination of transport mechanisms that was beyond the scope of this study.

Results

Fig. 2 shows the time evolution of the 2015 event (black) across the satellite data for Sumatra and Kalimantan compared with the last major event in 2006 (red). All data have been averaged over the previous 7 d.

Fig. 2.

Time series of various species for 2015 and 2006 over Sumatra and Kalimantan. OMI AI, MODIS AOD, MOPITT CO, AIRS CO, and MLS CO are averaged for Sumatra over (10°S–10°N, 90°E–105°E) and for Kalimantan over (10°S–10°N, 105°E–120°E). MODIS active fires and TRMM precipitation are averaged over the smaller primary burning regions, for Sumatra (6°S–0°, 99°E–106°E) and for Kalimantan (4°S–0°, 110°E–117°E). Subscripts for the MLS, AIRS, and MOPITT CO are the altitude in hPa over which the data were analyzed. All data have been averaged over the previous 7 d.

In southern Sumatra, the 2015 dry season captured by TRMM precipitation began in mid-June, with limited fire activity (<100 detections/d) appearing late in the month and interrupted by brief periods of rain in mid-July and early August. Fire activity increased in late August and by early September, 7-d average fire detections varied around 600/d. Brief rain during the third week of September caused a temporary decrease in fire activity, which was followed by persistently high fire detections until the return of the monsoon in early November.

The increases in MODIS AOD over the larger region including Sumatra lag behind the increase in fire activity by roughly 10 d but varied around 0.6 for September, increasing slowly through the first 3 wk of October. At the end of October, average AOD increased sharply to ∼1.4. The rapid AOD decrease in early November with the arrival of the monsoon lagged by roughly a week behind the drop in fire activity. Increases in OMI AI followed those in the MODIS AOD, but with the timing of peak values (∼0.9) more closely following the peaks in MODIS fire activity in early and late October. A close examination of the OMI AOD retrieval showed that, although pixels with low to moderate loading of aerosols were reported as the best quality data, retrievals with higher reflectivity (>0.3) at 388 nm are flagged as less reliable in the AOD inversion. OMI pixels having reflectivity larger than 0.3 directly over the biomass burning are excluded from the retrieval process due to higher reflectivity that is often associated with the clouds. However, the UV-AI is derived and reported for all-sky conditions regardless of reflectivity of the scene.

The signature of the event can be seen in the MOPITT CO retrieval from the surface to 200 hPa but is particularly distinct at 500 hPa. CO in the midtroposphere lags behind the increases in fire activity and aerosols through August and September. MOPITT requires cloud-free observations for CO retrievals so many scenes over Indonesia are excluded. Of the remaining high-quality retrievals, CO approached 300 ppbv, slightly higher than in 2006, but the amount of missing data in 2015 makes a comparison between the two years difficult.

The AIRS CO at 500 hPa had less missing data due to greater data coverage and the use of extrapolation to cloud-free radiances before the retrieval. By October, CO increased to concentrations as high as 300 ppbv, dropping sharply in early November with the return of the monsoon. The MLS CO at 215 hPa (∼12 km in altitude over Sumatra) increased steadily through September and varied slightly above 200 ppbv during the first 3 wk of October. A rapid increase to CO exceeding 400 ppbv at the end of October corresponded to the sharp increases in the MODIS AOD. In the uppermost troposphere at 100 hPa (∼16 km), the increase in CO only began in mid-October, but rapidly approached 175 ppbv before the end of the burning season.

Fire activity and pollution for 2006 over Sumatra showed similar precipitation-driven timing to 2015, but was overall lower in magnitude and shorter in duration. MLS CO at 215 hPa briefly exceeded 200 ppbv and at 100 hPa mostly remained below 100 ppbv. The less severe conditions in 2006 were due to more precipitation from June through mid-September.

Over Kalimantan, the timing of drying, fire activity, and tropospheric pollution during 2015 was very similar to that over Sumatra. Fire activity increased in August, varying about 800 detections/d through September. Precipitation in early October caused a temporary decrease in fire activity, MODIS AOD, and OMI AI, but was followed in late October by sharp increases similar to Sumatra, particularly in the aerosol-related retrievals. AIRS and MOPITT CO at 500 hPa varied around 225 ppbv for October, with fewer excluded MOPITT profiles than for Sumatra. MLS CO showed an increase in late October, as Sumatra, but to lower CO concentrations, briefly exceeding 300 ppbv at 215 hPa and 150 ppbv at 100 hPa. Compared with 2006, the earlier start of fire activity over Kalimantan in 2015 was offset by an earlier onset of the monsoon. Other than higher CO at 100 hPa, 2015 and 2006 were of comparable magnitude over Kalimantan. Even after the fires stopped with the return of the monsoon, both regions continued to have 100 hPa CO well above “background” (60 ppbv) through November 2015.

This event represents the largest enhancement in the MLS record of CO at 215 hPa (i.e., since August 2004; Fig. 3). The CO peaks approaching 300 ppbv over western Indonesia form part of a broad signature stretching from the western Indian Ocean to the southwest Pacific Ocean, exceeding the extent, magnitude, and duration of the 2006 event. High (200 ppbv) is also measured regularly over eastern South America (boxes Sa11 and Sb11) due to burning in the Arc of Deforestation around the Southeastern Amazon and the Cerrado (savanna) further south, but the upper tropospheric CO signature has a much smaller extent than over Indonesia.

Fig. 3.

Biweekly averaged MLS CO at 215 hPa (∼12 km) for 15° × 30° cells during 2004–2015. Line colors denote year. Numbers and letters in black boxes along the top and right edges identify regions for discussion in the text.

Fig. 4 shows the extinction coefficient (B_ext) for 2015 along with 1991 and 1997. Burning in 1994 was also severe and is discussed in the next section. B_ext is computed using visibility from three weather stations in each of southern Sumatra and south-central Kalimantan’s main burning regions over which fire activity and precipitation was averaged in Fig. 2. Over Sumatra in 2015, the B_ext peaks in mid-September and early and late October correspond closely to those seen in fire activity in Fig. 2, reinforcing the usefulness of airport visibility as a severe haze indicator in Indonesia. The 2015 B_ext increase is more severe and longer in duration than 2006 but is much lower than the 1991 and 1997 episodes. The magnitude of the 1997 event reflects much lower antecedent rainfall beginning in June. The later 1991 peak in early October compared with 1997 was due to significant rainfall in early September. The late October interruption in haze in 1991 also followed significant precipitation. Overall, B_ext data indicate that the 2015 haze in southern Sumatra was less severe than 1991 or 1997, which is easy to explain for 1997 given the decreased precipitation that year during the exceptionally strong El Niño. However, it is more difficult to explain for 1991, which was only slightly drier than 2015.

Fig. 4.

NOAA Climate Prediction Center (CPC) gauged-based precipitation and extinction coefficient (B_ext) calculated from horizontal visibility. CPC precipitation averaged over the smaller primary burning regions, for Sumatra (6°S–0°, 99°E–106°E) and for Kalimantan (4°S–0°, 110°E–117°E) for 3 high fire years. Extinction coefficient calculated over Sumatra from the airports in Rengat, Jambi, and Palembang, and over Kalimantan from Pangkalan Bun, Palangakaraya, and Muaratewe. All data have been averaged over the previous 7 d.

In Kalimantan, 2015 B_ext has two peaks in late September and late October that correspond to those in fire activity. The 2015 B_ext was also lower than in 1997 due to the near absence of precipitation that year between mid-July and early October. The early September onset of 2015 haze was comparable to that in 1991 and its termination earlier, but with weaker isolated precipitation events than in 1991, making it slightly more severe.

Across the satellite observations, we can conclude with a fair amount of certainty that 2015 was a worse fire year than 2006, because of its earlier start in Sumatra, higher fire activity in September over Kalimantan, and despite an earlier end in Kalimantan. There is greater uncertainty associated with the visibility-based B_ext record due to possible changes in observing procedures and more missing records in the early 1990s, but the available data suggest more severe 1990s burning in Sumatra compared with 2015, and in Kalimantan less severe burning in 2015 than in 1997 but more than 1991.

To give a more complete picture of the relationship between annual fire or pollution magnitude and the underlying dry conditions, Fig. 5 shows the annual mean dry season (August–November) precipitation plotted against the different fire and haze indicators, for all years over which each source of data are available. Mean precipitation is averaged over the previous 12 wk to include the effects of antecedent drying for each month during the dry season. In each case, we estimated the strength of the nonlinear relationship using piecewise linear regression, which includes an estimated change-point parameter α. We interpret α as the precipitation threshold below which fire and pollution magnitude increase rapidly and above which conditions are too wet for high fire activity and pollution. The estimates of α also provide an empirical means of separating severe from nonsevere fire years.

Fig. 5.

Annual August–November relationships between CPC precipitation over the source regions and fire/pollution metrics. Subscripts for the MLS, AIRS, and MOPITT CO are the altitude in hPa over which the data were analyzed. The numbers above each point are the last two digits of each year. Dashed lines show the piecewise-linear regression fit for each region, the coefficient of determination (R²), and change-point estimate of precipitation (α, mm/d), which are provided in the figure legends.

There is a consistently nonlinear relationship between dry season precipitation and the different indicators of fire and pollution. Across all indicators of fire and haze during the EOS period, the estimates of α ranged from 3.9 to 5.2 mm/d. That is, for average dry-season precipitation greater than 6 mm/d, there is little fire activity or pollution. Between 4 and 6 mm/d, there is some increase in fire and pollution, and below 4 mm/d, fire and pollution increase rapidly. This nonlinearity has been seen before over broadly the same regions for MODIS fire detections (29) and the MODIS-based Global Fire Emissions Database (12). The nonlinearity was also seen between seasonal precipitation in B_ext, depending on the period considered for Sumatra and Kalimantan (7).

During the EOS period, the nonlinear relationship with precipitation is strongest (R² between 0.85 and 0.98 depending on the region) for MODIS AOD, AIRS, and MOPITT CO at 500 hPa, and MODIS fire detections. It is still present, but weaker (R² between 0.69 and 0.90) for the OMI AI, and MLS CO at both levels, presumably due to their higher-altitude retrieval sensitivity, and therefore additional dependence on a vertical transport mechanism, which has been examined for Indonesian biomass burning using different transport models (30–32). The greatest separation between Sumatra and Kalimantan is for the MLS 215 hPa CO and precipitation relationships. We speculate that at this altitude, pollutant concentrations are strongly dependent on nearby deep convection, but that higher at 100 hPa, there is a greater influence of horizontal advection and the subsequent mixing of pollutants between the two regions. This height-dependent mixing is supported by a parameterized case study (32) for the 2006 event, in which the convective supply of CO from the surface peaked at 200 hPa, and was very limited at 100 hPa.

For the longer-term B_ext haze proxy, there is a strong nonlinear relationship with precipitation over Sumatra (R² = 0.90), which weakens over Kalimantan (R² = 0.77). This difference is due to a weaker linear relationship over Kalimantan for years with seasonal precipitation below the estimated 3.7-mm/d threshold, which is discussed further below.

Discussion

Fire activity is known to increase in the tropics during droughts, as long as fuels are abundant (33). However, the consistency and nonlinearity of this relationship in Indonesia across such a diverse set of satellite-based measurements during the EOS era is remarkable. We are unaware of any other large region where interannual variation in fire activity and pollution through the depth of the troposphere is so strongly, and nonlinearly, related to the dry conditions on the ground. The uniqueness of the relationship for Indonesia is due to the ubiquitous use of fire that grows out of control during droughts, its large area of degraded peatlands, and, presumably, the strong control that precipitation has over whether the peat becomes dry enough to burn (34).

The B_ext plot in Fig. 5 suggests an increase in fire sensitivity over Kalimantan since the 1990s. Despite occurring under comparably dry, or even wetter conditions, the burning in 2006, 2015, and also 2002, was more severe than in 1991 and 1994, which is what weakens the nonlinear relationship with precipitation compared with Sumatra. This increase in fire sensitivity over Kalimantan seems to be a continuation of a longer-term trend (7). Severe fire was absent in the 1960s and 1970s despite regularly occurring drought years. Severe fire appeared only in the 1980s, and strengthened in the 1990s, which was attributed to intensifying land use change (7). In southern Sumatra, 2015 and 2006 were somewhat less severe than 1994 despite similar seasonal rainfall, perhaps suggesting a decrease in fire sensitivity. This decrease in fire sensitivity could correspond to an increase in fire prevention and suppression on larger industrial plantations in the provinces of South Sumatra and Jambi, or to a northward shift in the intensiveness of fire, as noted by recent case studies during the secondary dry season in Riau province (10, 12) and satellite records of tree-cover loss (35). Possible changes in fire sensitivity inferred from B_ext will need to be studied further, taking into account changes in data completeness, the effect of different year-to-year transport patterns relative to the airport locations, and most importantly, corroboration with estimates of changing land use.

Since 1997, annual emissions from fires in all of Indonesia have been between 6 (in 2010) and 1,046 (in 1997) Tg C according to the Global Fire Emissions Database version 4s (GFED4s), updated from previous versions (36). Total emissions for 2015 were estimated to be 380 Tg C, which translates to 1.5 billion metric tons CO₂ equivalent when also including emissions of methane and nitrous oxide. This emissions estimate is in between the 2013 annual fossil fuel CO₂ emissions of Japan and India (37). Known sources of uncertainty in the emissions estimate are a possible underestimation of burned area due to cloud and smoke cover (38) and a possible overestimation relating to recent work (39, 40) showing that the depth of peat burning decreases for successive fires, which is not yet taken account for repeated fires in the same area in GFED4s. Viewed historically, these events are nevertheless a large part of what makes Indonesia’s land use change-related greenhouse gas emissions much larger than its fossil fuel emissions compared with other countries (41).

Eliminating fire from degraded peatlands is a long-term goal and will require major reforms in land use and land tenure in the context of Indonesia’s need for economic development. In the short term, fire prevention, suppression, and mitigation measures must be tied to early warning triggers. Our analysis over five different indicators of fire activity and atmospheric pollution from NASA EOS data suggests that doing so is a matter of being able to anticipate extended periods of less than 4 mm/d of rain. Given the skill with which strong El Niño impacts can increasingly be predicted (42, 43), tying these predictions to early warning triggers based on these types of precipitation thresholds should be a priority.

SI Text

Satellite Data.

The following are brief descriptions of the satellite-based fire and pollution indicators.

MODIS has been aboard polar-orbiting Terra and Aqua satellites since 1999 and 2002, respectively. Measuring radiance in 36 spectral bands ranging from 0.41 to 14.2 μm, MODIS observes a swath of 2,300 km (2-d global coverage), with nominal (at nadir) pixel spatial resolutions of 1 km or better.

MODIS active fire detections (often referred to as fire counts) are based on the radiance in the 3.9- and 11-µm channels with a 1-km spatial resolution (23). We used the high-confidence observations from the MCD14DL product, which contain the locations of all active fire detections of MODIS on board of both Aqua and Terra providing almost four overpasses each day. Active fires can be seen through optically thin clouds or smoke but may be obscured during less favorable atmospheric conditions.

From selected bands in visible, near-IR, and infrared regions, the collection 5 version of the MODIS dark-target algorithm retrieves aerosol properties over ocean (20) and land (21) at 10-km spatial resolution. Although the land and ocean algorithms are separate, both operate by matching the observed spectral reflectance to lookup tables that contain simulated TOA reflectance. Assumptions are made about the surface characteristics and likely aerosol optical properties. Results include spectral AOD, estimates of relative aerosol size, and diagnostics including quality assurance and confidence (QAC) criteria. These level 2 (L2) retrievals are performed under cloud-free conditions and where surfaces are sufficiently dark. By comparing to global, ground-based sunphotometer data, expected error envelopes for the L2 AOD at (550 nm) product have been reported as ±(0.05 + 15%) over land (44) and ±(0.04 + 10%) over the ocean (45). Daily level 3 data used here are created by aggregating the L2 data and accounting for QAC values (46).

Since its launch in July 2004, the OMI on board NASA’s Aura satellite has measured reflected radiation from Earth in the 270- to 500-nm wavelength range of the spectrum globally on a daily basis. OMI is a hyperspectral instrument viewing in 740 wavelength bands. With a cross-track swath of about 2,600 km at ground level, OMI scans the entire Earth in 14–15 orbits, with a nadir ground pixel spatial resolution of 13 × 24 km². Satellite observations at 354 and 388 nm made by OMI are used to derive UV aerosol index (UV-AI), as well as the AOD using a near-UV algorithm (OMAERUV) that takes the advantage of the well-known sensitivity to the aerosol absorption in the UV spectral region (22, 47, 48). After 2008 operation, the OMI sensor has encountered an instrument problem, formally referred to as “row anomaly,” in which a portion of the instrument’s field-of-view (FOV) has been likely obstructed by an external object affecting the quality of radiance at all wavelengths for a particular viewing direction or row number. This row anomaly problem has significantly affected the sampling during post-2008 OMI measurements, where about half of the OMI swath is blanketed by row anomaly flags. Despite the instrumental issue, OMI has maintained the quality of its measurements for the unaffected rows and has been validated and used in the scientific publications (49, 50). In this paper, we primarily use OMI/OMAERUV (v1.4.2)-measured UV-AI, briefly discussing retrieved AOD at 388 nm to characterize the aerosol loading over the Indonesia fire region.

The MOPITT instrument on Terra is in a sun-synchronous orbit with a mean local overpass time of 10:30 AM. It has a 22- × 22-km horizontal footprint with global coverage in ∼3 d. MOPITT uses gas filter correlation radiometry at 4.6 and 2.3 μm to detect atmospheric CO and has a data record that spans nearly 16 y. MOPITT CO data have been validated using aircraft in situ measurements (17). Here we use day and night MOPITT V5J data that have been filtered to require cloud-free scenes from both MOPITT and MODIS cloud detection, scenes with corresponding valid National Centers for Environmental Prediction reanalysis water vapor fields, and degrees of freedom for signal >1.0.

The AIRS instrument is a 2,300-channel infrared grating spectrometer on the Aqua spacecraft, in a sun-synchronous orbit with northward equator crossing time of 1:30 PM. AIRS CO is retrieved with horizontal resolution of 45 km at nadir, in a swath of width 30 FOVs or about 1,600 km. This orbit gives global coverage in the tropics every 2 d. The retrieval uses a cloud-clearing methodology providing CO with sensitivity that peaks around 500 hPa, with ∼0.8–1.2 degrees of freedom of signal for 50–70% of scenes. More sampling and higher information content are obtained in clear scenes (18).

The MLS on the Aura spacecraft measures CO (and a range of other trace species) profiles from the upper troposphere to the mesosphere, with CO information derived from observations of the 230.5-GHz spectral line. We use the version 4.2x MLS CO product at 215 and 100 hPa (19), screening them according to the rules given therein. The vertical resolution is ∼5 km, with an along track resolution of ∼700 km. Systematic errors are estimated to be ±30 ppbv and ±30%. MLS observations are unaffected by aerosols and all but the thickest clouds. CO from AIRS, MOPITT, and MLS are here reported individually, but not with the intent of comparing across instruments, which requires accounting for temporal and spatial sampling differences and the application of instrument operators that take account for a priori profiles and different vertical sensitivities in the retrievals.

Piecewise Regression Analysis.

The nonlinear relationships between the different fire and haze indicators were estimated using piecewise regression analysis (7, 12). This model consists of two linear segments constrained to meet at an unknown change point. In our case, this approach has the advantage over fitting a higher-order curve in providing an objective estimate of the precipitation level below which fire and pollution increase rapidly with decreasing precipitation.

Specifically, the model is given by

$$y={\{}_{{\beta }_0+{\beta }_{1}x+{\beta }_2\left(x-\alpha \right),\hspace{0.17em}\hspace{0.17em}x>\alpha }^{{\beta }_0+{\beta }_{1}x,\hspace{0.17em}\hspace{0.17em}x\le \alpha },$$

where x is the explanatory variable (precipitation), y is the fire or pollution data, α is the change point to be estimated, β₁ is the slope of the line for precipitation below α, and β₁ + β₂ is the slope of the line for precipitation above α.