IEEE Geoscience and Remote Sensing Magazine - March 2021 - 267

data may negatively affect the methods' accuracy, both in
terms of undersampling, due to the temporal resolution of
Landsat, and false positives, because of the solar-irradiation
effects in daytime images.
BISPECTRAL METHODS
Hot sources such as gas flares, wildfires, or steel mills produce a contrast when observed at different IR wavelengths
(Figure 5). The larger the temperature, the sharper the contrast between the radiances at two given wavelengths. The
methods that exploit such contrasts are called bispectral.
The difference between the brightness temperatures measured in the MWIR and TIR channels by AVHRR [108] and
MODIS [87], [109], [110] has been widely used to recognize
subpixel hot sources. Matson et al. [108] demonstrated the
feasibility of the distinction of GF among other subpixel hot
sources (e.g., steel mills) in terms of temperature and area
by means of the Dozier differential method using AVHRR
bands. Using this algorithm, the radiance measured in a
spectral channel is modeled as the linear combination of the
radiances emitted from two temperature components: one
representing the hot target (dominant in the MWIR but negligible in the TIR bands) and the other the background (the
signal present in the MWIR and TIR bands). GF is characterized by a contrast larger than that of other hot sources. For
MODIS, Elvidge et al. [109] applied a threshold to the MWIRTIR difference to isolate gas flares, while the authors in [87]
and [110] processed this signal using a standardized variable
before applying a time-persistence criterion for discriminating gas flares from other hot sources. These MODIS-based
findings resulted in good agreement with other satellite
products (OLS in [109] and VIIRS in [29]).
Considering subpixel hot sources, the share of the background signal is larger in the TIR, less in the SWIR, and
intermediate in the MWIR. In the particular case of GF, in
which combustion temperatures are much higher than,
e.g., wildfires, the share of the hot source component is
larger in the SWIR than in the MWIR and is practically
nonexistent in the TIR. Therefore, it makes sense to exploit
the contrast between the SWIR and the MWIR. The AVHRR
instruments were not able to measure in the MWIR and the
SWIR at the same time, while MODIS registered only the
250 and 500 m bands (among the SWIR bands) at daytime.
The AATSR series and later moderate resolution imagers
(i.e., VIIRS and SLSTR) have the ability to register nighttime SWIR and MWIR, allowing for a bispectral approach
with improved accuracy and reliability levels. Zhang et al.
[89] proposed using the spectral difference between the VIIRS SWIR and MWIR bands (i.e., M10 and M12) to identify
gas flares (where the M10 radiance is expected to be larger
than at M12) together with the shape of the spectral radiance curve at eight M bands (from NIR to TIR), showing a
peak at M10 for flare-containing pixels. A fixed thresholding on SWIR/MWIR radiances, based on both VIIRS [111]
and ATSR [96] channels converted in SWIR/SWIR ratios
for SLSTR radiances, also proved to be a valid way to filter
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

the persistent thermal anomalies associated with industrial gas flares.
In summary, the AVHRR saturation limit (321.5 K, [112])
restricted the applicability range of the Dozier method [108]
while MODIS appeared to be more suitable than OLS for
gas flares identification thanks to its higher observations
times, onboard calibration, rare saturation of gas flares in the
thermal bands, and data-collection continuity with VIIRS
[109]. The global spatial distribution of flaring and nonflaring sites produced by the VIIRS and ATSR data sets
showed an overall similarity, with some discrepancies due
mainly to the different imaging geometry, detection algorithm procedure, and thermal anomaly sensitivity of the two
sensors (e.g., due to signal-tonoise ratio, VIIRS's smaller
pixel area is better maintained
TO DATE, USING SATELLITEacross the swath).
BASED APPROACHES

ALLOWS FOR LOCATING
CONTEXTUAL METHODS
Contextual approaches comFLARING SITES AND
pare the hot spot spectral beCOMPUTING THEIR
havior with the surrounding
TEMPERATURE, AREA, AND
background. When applied,
RADIANT POWER.
the threshold value is derived
from the background or the
scene signal in the MWIR and
SWIR channels. For the former (the MWIR based), the authors in [95] combined radiometric (an apply threshold)
and spatial filtering (a high-pass filter) in parallel, both
based on fixed thresholds, on the MODIS MWIR radiances
to detect gas flares in the Niger Delta, recognizing a possible underdetection of low-intensity flares due to the coarse
spatial resolution of MODIS data. For the latter (the SWIR
based), a SWIR channel from VIIRS [82], SLSTR [86], and
TM/ETM+/OLI [88], [94] has been reported as the primary
band used for GF identification.
For Elvidge et al. [82], who developed the VNF algorithm,
a SWIR (1.6-μm) hot spot is confirmed as gas flare if the corresponding location also exhibits a radiance value above
the threshold in one of the remaining NIR or MWIR bands.
For the SWIR and NIR bands, the threshold is defined as
the mean plus four standard deviations of the surrounding
background pixels. For the MWIR bands, the hot pixel-detection threshold is the background mean plus three standard deviations, where the background corresponds to a 10
× 10 pixel window around the original SWIR hot pixel.
The authors in [86] used an approach similar to that of
[82], exploiting the SWIR and MWIR bands acquired by
the SLSTR onboard the Sentinel-3A satellite in the detection
scheme. The main differences with VNFs are that the SLSTR-based method 1) analyzes the integrated radiances of
contiguous clusters of hot pixels instead of the local maximal SWIR radiances, 2) analyzes two SWIR channels, and
3) uses the TIR channels when fitting the sum of the two
Planck curves. The methodology appeared able to detect
less-intense gas flares (less active or colder) thanks to the

267

IEEE Geoscience and Remote Sensing Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2021