1. Introduction
Vegetation fire, as an important component of biogeochemical cycles, combusts fuels on the land and releases water vapor, trace gases, and aerosol emissions into the atmosphere [
1]. Global biomass burning contributes a substantial amount of emissions that significantly affect the atmospheric carbon budget, weather conditions, air quality, and climate [
2,
3]. For example, global fires annually burn an area of ~350–422 Mha [
4,
5] and release ~2.2 Pg carbon [
6], among which 67% (burned area) and 52% (carbon) are, respectively, contributed by biomass burning in Africa alone [
4,
6]. Fire emissions are likely underestimated from satellite observations due to a substantial number of undetected small fires [
6,
7,
8]. One of the most important fire parameters observed from satellites is fire radiative power (FRP)—the instantaneous fire radiative energy [
9]. FRP has been increasingly used to understand fire characteristics [
10,
11,
12], analyze impacts of fire weather conditions on fire activity [
13], and predict smoke injection height [
14,
15]. More importantly, FRP is a proxy of the rates of biomass combustion and emissions [
9,
16,
17], which has been confirmed in lab-based combustion experiments [
18], prescribed fires [
19], and landscape wildfires [
20,
21,
22,
23]. Thus, satellite-based FRP provides an effective way to quantify biomass-burning emissions.
Among all satellite-based FRP products, the most commonly used and scientifically reliable one is from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the NASA Earth Observing System (EOS) Aqua and Terra satellites. The MODIS active fire products have provided FRP data since 2000, which has been widely applied to estimate regional-to-global trace gases and aerosol emissions [
24,
25,
26,
27,
28,
29]. However, global applications of MODIS FRP are affected by the MODIS sampling limitations and the equatorial swath gaps. First, the MODIS sensor is able to detect a fire pixel of above 10 MW confidently at nadir but its fire detection capability decreases quickly at off-nadir angles due to degrading pixel resolution [
30,
31,
32]. Thus, MODIS could miss observations of fires (at off-nadir angles) that are detectable at nadir, which results in the underestimation of MODIS FRP and emissions estimates. To mitigate FRP underestimation caused by the sampling limitation-, a quadratic model has been proposed separately for Aqua MODIS and Terra MODIS [
30], and a mapping method has been developed based on cumulative distribution function (CDF) [
33]. Both methods assume that MODIS FRP or FRP CDF in all scan angles should be the same as that at nadir if MODIS pixel size is the same across all angles. This assumption would be valid only if fires burn consistently in the same locations with constant FRP values during a long period when the sensor senses the fires in all angles with the same observation frequency. It could be the case for gas flares but not landscape wildfires because wildfires evolve fast with environmental conditions and its intensity is very dynamic [
34]. Wang et al. [
35] mitigate the underestimation of emissions at MODIS large view angles in the Northern Hemisphere of Africa, by using emissions estimated with the smallest view angle within ±2 days, assuming persistent burning patterns during the same period. This may not work well because burning sites could vary largely due to human activity [
36].
Second, due to MODIS swath gaps between adjacent orbits at low latitudes, MODIS needs two days to provide a full coverage of the equatorial regions between ~30°S and ~30°N [
37]. As a result, MODIS misses observations of fires daily inside its swath gaps, which has been known in fire detection [
38,
39] and emissions estimation [
40]. Due to the lack of comparable reference data, the effects of swath-gap-caused missing fire observations on FRP and emissions estimates are poorly understood quantitatively. To deal with the swath gap issue, studies often make assumptions that fire activity within the swath gaps is consistent during a day or two days [
24,
35].
The imagery bands (I-band) of Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the Suomi National Polar-orbiting Partnership (S-NPP) satellite for the first time provides daily global fire observations at a resolution of 375 m without any gaps [
41]. As a complement of the VIIRS 750 m active fire product [
42], the VIIRS I-band 375 m active fire data theoretically detect many more small and cool fires than the 1 km MODIS active fire data, which have been demonstrated in agriculture burnings in China [
43] and India [
44]. Therefore, the VIIRS I-band 375 m active fire data has a large potential to improve biomass-burning emissions estimation. Both the S-NPP VIIRS and Aqua MODIS cross the equator approximately at 01:30 and 13:30 local time and observe the same area with an overpassing time difference of fewer than 50 min [
37,
45].
In this study, we use the 375 m VIIRS fire data as a reference to investigate the missing MODIS fire observations due to the sampling limitations and the equatorial swath gaps. This study is conducted in the following steps. First, we introduce the sensing geometry of MODIS and VIIRS I-band and compare their fire detection capability across swath. Second, MODIS and VIIRS active fire detections are preprocessed to correct inter-scanline repeat fire detections and extract fire detections sensed contemporaneously by the two sensors. Third, we quantify VIIRS fire observations missed by MODIS and compare contemporaneous FRP estimates from the two sensors at a continental scale and in grids of seven different resolutions to establish empirical MODIS FRP adjustment models. Finally, we investigate the swath-gap-caused missing MODIS FRP observations.
5. Discussion
The MODIS sensor has a relatively low fire detection capability and misses a large portion of FRP observations relative to the VIIRS I-band sensor. The minimal FRP (per fire pixel) across swath suggests that the MODIS sensor misses fires with intensity less than ~4.3 MW at nadir and ~31.7 MW at the scan edge (
Figure 4a), whereas VIIRS I band (375m) can detect approximate 3-11 times smaller fires (
Figure 4b). Across the African continent, fires missed by MODIS result in the underestimation of MODIS FRP by 42.8% (26% + bias interpreted as 16.8% of daily mean MODIS FRP in
Figure 5b) because fires with FRP<31.7 MW account for a significant portion of all fires as indicated by VIIRS FRP frequency density distributions (
Figure 7b). Our previous comparison between MODIS and VIIRS M-band FRP suggests that the daily MODIS sum FRP is only underestimated by approximately 10.5% across Africa [
32]. Thus, much more underestimation of MODIS FRP is revealed by comparing with VIIRS I-band FRP. Similar findings on underestimation of MODIS FRP have also been demonstrated in cropland fires [
43,
44]. For example, compared with the Aqua MODIS monthly summed FRP in agriculture burnings, VIIRS I-band measures 120% more FRP in Punjab, India [
44], and synthesized VIIRS I-band and M-band radiances measure FRP about six times larger in eastern China [
43]. Additionally, it is worth noting that, for relatively small fires detected by the MODIS sensor, high albedo across the fire-prone Africa savannas also causes uncertainty in MODIS FRP retrievals [
54,
55].
Furthermore, missing MODIS FRP due to relatively lower fire detection capability is also evident at the grid level. For a given grid size from the finest 0.05° to the coarsest 0.5°, MODIS senses no fires in a large number of grids where only VIIRS has fire detections (
Figure 6). It indicates that VIIRS senses many more fires with FRP beyond MODIS fire detection capability. These MODIS missing fires occur in many areas across Africa, as suggested by the large percentages of grids where only VIIRS senses fires (
Figure 6a). For example, the percentage is still 34% even at a grid size of 0.5° (
Figure 6a), although fires in these grids contribute to a very small portion of daily VIIRS sum FRP (
Figure 6b,c). In addition, the increase in the ratio of VIIRS to MODIS grid FRP with view angles suggests that MODIS misses more FRP at large view angles than in nadir as its detection capability reduces at off nadir. For example, as grid size grows larger than 1.0° (
Figure 6), the underestimation of MODIS FRP could increase by up to ~60% at the scan edge relative to in nadir (
Figure 8e,f).
MODIS FRP underestimation due to relatively low detection capability can be partly mitigated. At a continent scale, daily MODIS FRP underestimation can be directly adjusted based on the derived relationship between MODIS and VIIRS FRP illustrated in
Figure 5b. This adjustment undoubtedly improves the estimation of total biomass-burning emissions from MODIS FRP across Africa. On a grid level, the fitted models at seven different grid resolutions can be applied to adjust MODIS FRP in grids where MODIS senses fires. Indeed, these fitted models can help to adjust grid-based MODIS FRP that is used in most of biomass-burning emissions inventories at a grid resolution from 0.1°–1.0°, including GBBEPx—Global Biomass Burning Emissions Products [
28]—GFAS—the Global Fire Assimilation System [
24]—and FEER—Fire Energetics and Emissions Research [
26]. For example, at the grid size of 0.5° the model can improve MODIS FRP by 40–50%. It is worthwhile to note that there is a trade-off between grid size and degree to which the MODIS FRP underestimation can be adjusted. At a fine grid resolution (e.g., 0.25°), there is a considerable number of grids with only VIIRS fire detections but without MODIS fire detections. In these grids, the missing MODIS FRP cannot be adjusted using the corresponding models that require FRP values available from both MODIS and VIIRS fire detections. For the grid size being larger than 2.5°, almost all grids contain both VIIRS and MODIS fire detections. In this case, the fitted models can be applied to improve MODIS FRP underestimation at almost all coarse grids, but spatial details are missed relative to fine-resolution ones. Therefore, choosing an appropriate grid resolution to adjust the MODIS FRP underestimation depends on the application purposes.
The FRP comparisons in different grid resolutions provide direct evidence to evaluate the published methods for adjusting the bow-tie effect caused MODIS FRP underestimation. First, our results show that MODIS FRP in 0.5° grids is relatively underestimated by 40–50% compared to VIIRS (
Figure 8d and
Table 1). This is very similar to the previous finding that MODIS FRP at a 0.5° grid could be improved by 44% across Africa by correcting the MODIS bow-tie effect [
30]. However, this study indicates that the MODIS FRP underestimation should be attributed to two factors: (1) the MODIS bow-tie effect and (2) the MODIS coarser-resolution effect relative to VIIRS I-band in nadir. In other words, VIIRS I-band, theoretically, is able to sense smaller and/or cooler fires than MODIS even at nadir, although both sensors observe fires in the same grid. Thus, it is likely the algorithm of bow-tie correction [
30] could overestimate MODIS FRP.
Second, Kaur et al. [
33] correct the MODIS bow-tie effect on FRP by mapping the CDF of grid FRP at large VZA to nadir and find that MODIS FRP could be improved by up to 44% for 1° grids with a VZA > 55°. Our results show that MODIS FRP is underestimated by 47–85% relative to VIIRS in 1° grids as VZA varies from nadir to the scan edge. Because VIIRS fire detection capability is almost consistent across VZA (
Figure 1 and
Figure 4b), it can be derived that the underestimation of MODIS FRP attributed to VZA effect could be as large as 38% (38% = 85% − 47%), which is slightly smaller than the 44% found by Kaur et al. [
33]. However, MODIS FRP in nadir is still underestimated by 47% compared to VIIRS FRP.
Third, Wang et al. [
35] corrects MODIS FRP at large view angles (VZA>35°) in 27 km × 27 km grids by using FRP at smallest VZA observed within two days and finds that emissions estimates could be increased by a factor of two across Northern Hemisphere Africa, including the FRP improvements by correction of clouds and swath gaps. However, our results show that MODIS grid FRP on average is smaller than VIIRS grid FRP by up to 31% at a grid resolution of 0.25° (~25 km at the equator,
Figure 7c). Even when all VIIRS I-band FRP inside and outside MODIS swath gaps across Africa are considered, the continental-scale MODIS FRP is relatively underestimated by ~62.5% (49% + bias interpreted as 13.5% of daily mean MODIS FRP in
Figure 5a), much smaller than the difference of a factor of two. It is likely that the assumption proposed by Wang et al. [
35] for correcting MODIS FRP at large VZA and swath gaps may not work well and could overestimate MODIS FRP. This is demonstrated in
Figure 11, which shows the variations of mean VZA and daily FRP of MODIS and VIIRS I-band in a 1° grid centering at 30.232°E and 4.139°N in January 2017. The grid FRP estimates from the two sensors show similar temporal patterns, although MODIS swath gap occurred on January 4 and 20. According to the assumption in Wang et al. [
35], MODIS FRP requires correction for a total of 23 days in January 2017. The correction of VZA and swath gaps obviously results in the overestimation of MODIS grid FRP and alters a temporal pattern compared to the observed VIIRS grid FRP, especially during days from Jan 5-10 and Jan 15-20 (
Figure 11).
The MODIS equatorial swath gaps also significantly affect its observation of actively burning fires at low latitudes. By referencing to VIIRS FRP, MODIS swath gaps result in a mean monthly omission error of ~12.5% in MODIS FRP, which varies largely in different seasons. It is similar to a previous finding that swath gaps explain 14% of omission error by comparing MODIS fire detections with field observations in the Yucatán forest, Mexico [
38]. At low latitudes, it is well known that cloud obscuration is one of key factors that limit fire observations [
56]. For instance, cloud contamination could lead to an omission error of 11% in MODIS fire detections across the Amazon region [
56], which is very close to the omission error caused by swath gaps. Thus, MODIS swath gaps play an equally important role in affecting MODIS active fire data as clouds. Therefore, the prediction of MODIS missing FRP inside swath gaps for historical MODIS fire data during the past two decades could significantly improve the estimation of biomass-burning emissions and benefit related applications.
Mitigation of the swath-gap-caused underestimation of MODIS FRP is very challenging. Although MODIS swath gap has been a known problem in fire detection since the first release of the MODIS active fire product [
57], very limited studies have investigated potential approaches to mitigate the swath gap effect on MODIS FRP. For any day without valid observations due to swath gaps, a commonly used strategy is to use the FRP observations of the previous day (or previous days with the minimal VZA) directly [
24,
35]. This approach could largely underestimate or overestimate the missing FRP inside swath gaps, as illustrated in
Figure 11, because burn sites most likely do not persist from day to day due to human activity and changes in fire weather [
36]. For the time period with the VIIRS fire observations available, a potential mitigation approach is to examine the daily FRP relationship between the MODIS and VIIRS sensors over the areas outside swath gaps and apply it to predict FRP missed by MODIS inside swath gaps.