DOI QR코드

DOI QR Code

An Extraction of Solar-contaminated Energy Part from MODIS Middle Infrared Channel Measurement to Detect Forest Fires

  • Park, Wook (Department of Earth System Sciences, Yonsei University) ;
  • Park, Sung-Hwan (Department of Geoinformatics, University of Seoul) ;
  • Jung, Hyung-Sup (Department of Geoinformatics, University of Seoul) ;
  • Won, Joong-Sun (Department of Earth System Sciences, Yonsei University)
  • 투고 : 2019.01.29
  • 심사 : 2019.02.05
  • 발행 : 2019.02.28

초록

In this study, we have proposed an improved method to detect forest fires by correcting the reflected signals of day images using the middle-wavelength infrared (MWIR) channel. The proposed method is allowed to remove the reflected signals only using the image itself without an existing data source such as a land-cover map or atmospheric data. It includes the processing steps for calculating a solar-reflected signal such as 1) a simple correction model of the atmospheric transmittance for the MWIR channel and 2) calculating the image-based reflectance. We tested the performance of the method using the MODIS product. When compared to the conventional MODIS fire detection algorithm (MOD14 collection 6), the total number of detected fires was improved by approximately 17%. Most of all, the detection of fires improved by approximately 30% in the high reflection areas of the images. Moreover, the false alarm caused by artificial objects was clearly reduced and a confidence level analysis of the undetected fires showed that the proposed method had much better performance. The proposed method would be applicable to most satellite sensors with MWIR and thermal infrared channels. Especially for geostationary satellites such as GOES-R, HIMAWARI-8/9 and GeoKompsat-2A, the short acquisition time would greatly improve the performance of the proposed fire detection algorithm because reflected signals in the geostationary satellite images frequently vary according to solar zenith angle.

키워드

1. Introduction

Most of fire detectable sensors (e.g., AVHRR, GOES, ATSR, MODIS, VIIRS) have used multispectral infrared channels including the middlewavelength infrared (3~5 μm; MWIR) channel. When compared to the thermal infrared channel (8~13 μm; TIR), MWIR radiation is much stronger at high temperatures. Thus, the radiation difference in hightemperature targets between the MWIR and TIR channels has been widely used to detect fires (Dozier, 1981; Matson and Dozier, 1981; Prins and Menzel, 1992, 1994; Kaufman et al., 1998;Justice et al., 2002; Giglio et al., 2003; Zhukov et al., 2006; Schroeder et al., 2014). However, the fire detection algorithms do not easily detect fire occurences during the day compared to night due to solar reflection and heating of the surface (Giglio et al., 2003; Zhukov et al., 2006; Wooster et al., 2012). The difficulty arises because the solar reflection in the MWIR channel almost equally contributes to surface radiation due to natural characteristics (e.g., emissivity and thermal inertia) during the day (Li andBecker, 1993; Goiita andRoyer, 1997; Petitcoloin and Vermote, 2002; Boyd and Petitcolin, 2004; He and Je, 2013). In particular, the solar effect in urban, bare ground, and sparsely vegetated areas increases the brightness value of the MWIRimages.Therefore, the removal ofthe reflected signal in an MWIRimage resultsin a reduction in false alarms of fire detection.

Recent geostationary satellites images including those of GOES-R, HIMAWARI-8/9 and GeoKompsat2A can provide a higher spatial resolution of approximately 2 km and hence they can be used to monitor fire evolution with a longer revisit time. Because these geostationary satellite images are affected by differentsolarradiation at each observation time, the reflected signals can vary with the acquisition time. Thus, fire detection using these geostationary satellite images would be more challenging due to the variation in reflected signals as opposed to detection using polar-orbiting satellite images, which visit at a relatively constant time. This is very important for the fire-detection algorithm because most fire-detection algorithms are based on brightness temperature. However, removing reflected signals of an MWIR image is not easy because the reflectivity or emissivity of the MWIR channel is estimated.

In previous studies, methods using the temperature at ground level(He and Li, 2011; 2012) and calculating the reflection signal using the TIR channel (NOAA NESDIS; Schmidt et al., 2010) to correct solar reflection in day MWIR images have been proposed. A method of using the temperature at ground level is theoretically betterto use for a fire detection algorithm; however, it requires an atmospheric correction algorithm and emissivity for all used channels. He and Li (2011) proposed a method to detect fires using the thermal radiation at a ground level from MODIS data having the MOD11C3 emissivity product and creating a look-up table (LUT) for the atmospheric corrections. They reported that the performance of the method improved by approximately 20% when compared to the MODIS fire algorithm. However, there are some drawbacks in applying it to fire detection. The classification-based method does not reflect actual surface changes well including surface wetness and senescent vegetation (Snyder et al., 1998). This may lead to an error in estimating the emissivity in dry season environments. The LUT for atmospheric corrections needs to consider abundant auxiliary data depending on the set up of the LUT such as total column water vapor contents (TWV), atmospheric temperature, digital elevationmodels(DEM), geometry parameters, and so on.

Alternatively, the GOES-R fire detection algorithm (Schmidt et al., 2010) uses the brightness temperature of the TIR channel to estimate the reflected signal of the MWIR channel, but this method does not consider the difference in the atmospheric effects of the two channels.In addition, thisis not independent oftheTIR channel. Thus, it cannot be used to calculate the difference of brightness temperature for both two channels.Therefore, it can only be used to evaluate the amount of reflected signal.

Taking these points into consideration, we try to mitigate the solar reflection effect from the brightness temperature at the top-of-atmosphere (TOA) level by using the calculated surface reflectance. We propose i) a simple method to estimate the MWIR emissivity using the relation between the MWIRand visible (VIS) channels as well as ii) an improved method to detect fires by mitigating the reflected signals from the estimated MWIR emissivity.

The main idea of this study noted the red channel. The red channel has been used as an evaluation of the solar effect in existing fire detection algorithms(Giglio et al., 2003, 2016; Schmidt et al., 2010; Wooster et al., 2012). In other fields, the relationship between the MWIR and red channels has been investigated in previous forest studies (Kaufman and Remer, 1994; Boyd et al., 1999). Kaufman and Remer (1994) replaced the MWIR reflectance with the red channel reflectance when calculating the vegetation index 3 (VI3). The results from the studies showed that the reflectance variation in the MWIR channel is very similar to the red channel, although the process of reflection is different for both two channels. The idea has not been used to estimate the emissivity of the MWIR channel from the red channel but to calculate the vegetation index from the MWIR channel instead of the red channel (Kaufman and Remer, 1994; Boyd et al., 1999). We exploit that the converse relationship between the red and MWIR reflectances is applicable to improving the fire detection algorithm. In addition, if the solar irradiance and the two-way atmospheric transmittance are known, the reflected signal can be easily removed.

MODIS images and the fire detection algorithm (Gigloio, 2015) were used to test for the removal of solar contamination. Thus, the performance validation of the proposed method was performed by comparing it to the MODISCollection 6 active fire product(MOD 14 collection 6; Giglio, 2015). Because the proposed fire detectionmethod does not largely depend on sensor types, it would seem that the proposed method can be applied to most satellite images.

2. Method to mitigate solar contamination

The fire detection algorithms are divided into two different types: 1) the absolute fire detection algorithm and 2) the contextual fire detection algorithm (Justice et al., 2002; Giglio et al., 2003, 2016; Zhukov et al., 2006; Wooster, 2012; Schroeder et al., 2014). The algorithms are performed by thresholding a brightness temperature value according to the sensorspecifications and day/night conditions. However, in the MWIR channel, day images contain surface radiation, atmospheric effect and reflected solar radiation (Li and Becker, 1993; Goiita and Royer, 1997; Petitcoloin and Vermote, 2002; Boyd and Petitcolin, 2004). In particular, reflected solar radiation is much more significant in high-reflection targets including urban, bare ground, and sparsely vegetated areas. There is a greater than approximately 10 K brightness temperature rise in 20% surface reflectivity at a 300 K surface temperature (He and Li, 2011). When the reflected solar radiation is removed, both the day and night images have the same processto transfersurface emission (except for the surface heating effect). As shown in Fig. 1, the MWIRobservations during the day and night are approximately performed using equations (1) to (4) (refer to Chandrasekhar, 1960 within Tang and Li, 2014) as follows:

\(L_{s a t, \lambda}=\left[\varepsilon_{\lambda} B_{\lambda}\left(T_{s}\right)+\left(1-\varepsilon_{\lambda}\right) L_{j}^{\downarrow}\right] \cdot \tau\left(\theta_{s}\right)_{i}^{\uparrow}+L_{\lambda}^{\uparrow}+E_{s u n, \lambda}\)       (1)

\(L_{s a t, \lambda}=\left[\varepsilon_{i} B_{i}\left(T_{s}\right)+\left(1-\varepsilon_{i}\right) L_{\lambda}^{\downarrow}\right] \cdot \tau\left(\theta_{s}\right)_{i}^{\uparrow}+L_{i}^{\uparrow}\)       (2)

where Lsat, λ is the radiance observed by the satellite sensor, ελ and ρλ are the surface emissivity and surface reflectivity at a given wavelength, respectively, Bλ (Ts) is Planck’s function at the surface temperature, Lλ and Lλ are the upwelling and downwelling atmospheric radiances,τ(θs)λ isthe upward atmospheric transmittance at given zenith angle (θs), and Esun, λ isthe reflected solar radiance from the surface as given by the following:

\(E_{s u n, \lambda}=\frac{\rho_{\lambda} \cdot I_{g, \lambda} \cdot \tau\left(\theta_{s}\right)_{\lambda}^{\uparrow}}{\pi}\)       (3)

\(I_{g, \lambda}=I_{s u n, \lambda} \cdot \cos \left(\theta_{z}\right) \cdot \tau\left(\theta_{z}\right)_{i}^{\downarrow}+I_{\text {scat}, \lambda}\)       (4)

where Ig, λ is the total (directed and scattered) solar irradiance at the ground level at the given wavelength, Isun, λ isthe solarirradiance in the exoatmosphere (TOA) at the given wavelength, Iscat, λ is the solar irradiance scattered by the atmosphere, \(\tau\left(\theta_{z}\right)_{\lambda}^{\downarrow}\) is the downward atmospheric transmittance at the solar zenith angle (θz), and \(\tau\left(\theta_{s}\right)_{i}^{\uparrow}\) is the upward atmospheric transmittance at the sensor zenith angle (θs).

OGCSBN_2019_v35n1_39_f0001.png 이미지

Fig. 1. Illustration of the radiative transfer model in the MWIR spectral region. Solar effects appear during the day

In this study, we used the low-gain MWIR channel given astheMODISband 22.However,in orderto avoid saturation, the image pixels with a high temperature greater than 330 K were changed into image pixels from MODIS band 21. Fig. 2 shows the relative spectral response function of MODIS bands 21 and 22 (https://mcst.gsfc.nasa.gov/calibration/parameters) and the solar irradiance at ground level (Ig) with multiple scattering due to the solar zenith angle (SZA) calculated using the MODTRAN 5.0 radiative transfer code (MODO v5.0@ ReSe Applications Schläpfer, Wil, Switzerland) using the U.S. standard atmosphere model.The estimatedTOAsolarirradiance of MODIS band 22 was approximately 9.17 W/m2 /sr/μm, and the maximum variation due to the Sun-Earth distance was approximately 3.9%. In order to remove the reflected solar radiance defined by Equation (3) from the at-sensor radiance of Equation (1), the two-way atmospheric transmittance and surface reflectivity need to be determined.

OGCSBN_2019_v35n1_39_f0002.png 이미지

Fig. 2. Relative sensor response functions of the MODIS band 21 and 22 MWIR channels and atmospheric transmissivity spectrum of the U.S standard atmosphere model calculated using the MODTRAN 5.0 radiative transfer code (MODO v5.0@ReSe).

1) Transmittance

To remove the reflected solar radiance from an atsensor signal, it should be calculated for the upward and downward atmospheric transmittance in the optical path of equation (3). MODIS band 22 was designed to avoid the main effects ofCO2 and O3 absorption as well as water vapor contents. Thus, MODIS band 22 is slightly affected by atmospheric effects in the 3-5 μm MWIR spectral region.

The MODTRAN 5.0 radiative transfer code was used to obtain the atmospheric transmittance. Six standard atmospheric profiles such as mid-latitude summer and winter, sub-arctic summer and winter, and tropical and US standard in MODTRAN were examined with scaled water vapors of 0.2 to 4.0 g/cm2 with an interval of 0.2 g/cm2. The uniformly mixed gasses such as CO2 and O3 were used at default values for a standard atmosphere, which are 370 ppm for CO2 and 330 DU for O3. Finally, the atmospheric transmittance was convolved with the MODIS band 22 spectral response function. The atmosphere transmissivity is affected by the geometry as well asthe effect of water vapor contents. Because the MODIS sensor has a large field of view (FOV) of 55 degrees, an optical path of from 0 to 60 degrees with an interval of 10 degrees was used in the test

Fig. 3 shows atmospheric transmittance variations according to optical path and the atmospheric total column water vapor contents (g/cm2) in the MODIS band 22 MWIR channel. The total atmosphere transmissivity in the water contentfrom0.0 to 4.0 g/cm2 was estimated using the MODTRAN simulation. The variation in the estimated atmosphere transmittance was as small as 0.025 in a nadir angle, as shown in Fig. 3. This meansthat the transmissivity due to water content in the atmosphere would cause a small error. Fig. 4 shows the atmospheric transmittance variation according to air mass, which is the direct optical path length through the atmosphere, by a solar zenith angle. The error bar of Fig. 4 representsthe standard deviation of the atmospheric transmittance estimated from the water vapor content variation from 0 to 4 g/cm2 . The result suggests that the optical path is much more important than water vapor content in the MODIS band 22. Thus, in order to remove the solar reflected signal effectively, the calculation of the two-way atmosphere transmittance can be done by considering the solar zenith angle and sensor zenith angle as given by the following simple empirical equation:

OGCSBN_2019_v35n1_39_f0003.png 이미지

Fig. 3. Variation in the atmospheric transmittance according to the optical path and the atmospheric total column water vapor contents (g/cm2) in the MODIS band 22 MWIR channel. The graphs were obtained using the MODTRAN simulation. The optical paths from 0, 10, 20, 30, 40, 50 and 60 degrees are represented by the solid and open circles, solid and open triangles, solid and open squares, and cross, respectively.

OGCSBN_2019_v35n1_39_f0004.png 이미지

Fig. 4. Relationship between the optical path length and the atmospheric transmittance. The error bar shows the standard deviation of the atmospheric transmittances estimated from the water vapor content variations from 0 to 4 g/cm2.

\(\begin{array}{r} \tau(\theta)=-0.143 \mathrm{m}^{2}+0.193 \mathrm{m}+0.823 \\ \quad(\therefore \mathrm{m}=\mathrm{sec}(\text { Zenith angle })) \end{array}\)       (5)

2) Determination of surface reflectance

The estimation of surface reflectance or emissivity in the MWIR channel is usually categorized into 1) a method based on thermal infrared spectral indices (Beker and Li, 1990, 1992; Li andBecker, 1990; Nelly et al., 1998; Petitcolin and Vermote, 2002) and 2) empirical methods. The TISI-based method has a significant constraint in that it needs cloud-free images during both the day and night. Thus, the method is unsuitable for detecting fires because fire detection requires near-real time processing. The empirical methods can easily calculate surface emissivity values, and include amethod based on the normalized difference vegetation index (NDVI) and the classification-based method.They have an advantage in that they can easily calculate emissivity.The NDVI-basedmethod has been popularly used to estimate emissivity in the TIR channel (Van de griend and Owe, 1993; Valor and Caselles, 1996; Sobrino andRaissouni, 2000;JiménezMuñoz and Sobrino, 2003; Sobrino et al., 2008). However, the relationship between MWIR emissivity and vegetation index has not been well studied. In addition, this method has difficulties in application to non-vegetated areas. This is an important issue in the MWIR channel because the variations in the MWIR emissivity are generally larger than those of TIR emissivity as shown in various spectral libraries (Elvidge, 1988; Salisbury and D’Aria, 1994; Snyder et al., 1997, 1998). Alternatively, the MWIR emissivity is estimated by using a fixed value according to the classified land cover with the spectral library. The classification-based method has an advantage in that it can easily calculate emissivity values (Siroski et al., 1997). The classification-based method had also been proposed for the fire detection algorithm (He and Li, 2011; 2012). However, the method does not readily reflect actualsurface changesincluding surface wetness and senescent vegetation (Snyder et al., 1998). This may lead to an error in estimating emissivity in a dry season environment. Thus, it cannot reflect the actual surface emissivity.

To overcome the drawbacks of these conventional methods, we exploit that the reflectivity of the MWIR and red channels has a similar pattern (Kaufman and Remer, 1994). In previousstudy, thisidea has not been used to estimate the emissivity of the MWIR channel from the red channel but to calculate the vegetation index from the MWIR channel instead of the red channel (Kaufman and Remer, 1994; Boyd et al., 1999). In thisstudy, the converse relationship between red and MWIR reflectance is applicable to estimate surface emissivity in the MWIR channel.

The MODIS MCD43 16-day product for one year from Jan. 1, 2014, to Dec. 31, 2014, was used to analyze the relationship between MWIR and red reflectivity. In this test, we excluded the polar, ocean, and coastal areas, because we could not find any correlations between MWIR and red channelsin areas covered by water,snow, and ice. These coversshowed a constant MWIR reflectivity without regard to the variations in red channel reflectivity. To consider seasonal variation, we sampled the MODIS data every 8 days and applied the linear regression method to the sampled data. Consequently, a total of 14,578,097 samples were used to establish the relationship between red reflectance (Band 1) and MWIR emissivity (Band 22). Fig. 5 showsthe estimated linearregression model between red reflectivity and MWIR emissivity. From the regression analysis, we simply derived the MWIR emissivity (εMWIR) and reflectivity (ρMWIR) as follows:

\(\varepsilon_{M V T R}=-0.288 \rho_{R e d}+0.972\)       (6)

\(\rho_{M W R}=1-\varepsilon_{M W I R} (by Kirchoff's law)\)       (7)

OGCSBN_2019_v35n1_39_f0005.png 이미지

Fig. 5. Estimation of linear regression function between red reflectivity and MWIR emissivity. The estimation was performed by using the MODIS MCD43 16-day products acquired in 2014.

3) Elimination of solar effects

Fig. 6 shows an example of solar effect mitigation using the proposed method. The test image was the MOD02 L1Bobtained on 2:55 UTConApril 23, 2014. The two-way atmospheric transmittance was calculated from Equation (5) using the solar and sensor zenith angle from the MOD03 geolocation product as an auxiliary data. And the surface emissivity was estimated from Equation (6) using the MODIS Band 1 red channel reflectance image (Fig. 6 (a)). Finally, the reflected solar radiance was calculated by using Equations (3) and (4). Fig. 6 (b) and (c) show the MWIR brightness temperature and the solar-effectmitigated MWIR brightness temperature in the red box of Fig. 6 (a), respectively. Fig. 6 (d) shows the temperature difference before and after.The temperature showed a difference of approximately 2 to 7 K.

OGCSBN_2019_v35n1_39_f0006.png 이미지

Fig. 6. An example of solar effect mitigation from the MOD021KM and MOD03 geolocation product at 2:55 UTC on April 23, 2014: (a) The MWIR (Band 22) land surface emissivity map obtained from Equations (6) and (7) using the day red channel (Band 1) image, (b) brightness temperature from the MWIR image, (c) brightness temperature from the solar effect-mitigated MWIR image and (d) difference map between the brightness temperatures. (b), (c) and (d) show the area enlarged by the red box shown in (a).

3. Threshold optimization

If the solar reflection effect in the day MWIR image is well corrected, the corrected MWIR image would not depend on surface reflectivity. In addition, because the brightness temperature is reduced overall in the corrected MWIR image, the threshold value should be adjusted as much as the influence of the reflection signal mitigation for detecting fire. Therefore, in this chapter we want to test howmuch adjustment is needed to optimize the threshold values needed to detect a forest fire when the solar reflection signal is mitigated.

1) Study area and data

The test site is in Northeast Asia centered on the Korean peninsula.The study area largely included bare ground and sparsely vegetated areas in the Inner Mongolia, Manchuria, and North Korea region. Moreover, these regions have a burning period in the spring season because of slash-and-burn farming. Among the MOD021KM and MOD14 (collection 6) images obtained in 2014, a total of 113 images were selected in the study area because they include relatively low cloud cover. In order to reduce the false alarms of MOD14, a total of 8,254 fires with a confidence level of more than 30% were finally selected and used for the threshold optimization.

2) Determining Thresholds

In order to determine the optimal threshold value, the brightness temperature was calculated from the MOD021KM Band 22 MWIR image. As previously mentioned in chapter 2, the brightness temperature pixels higherthan 330K were substituted intoBand 21. After the image was corrected using the solar effect mitigation approach proposed by thisstudy, we created a solar effect-mitigated brightness temperature map. The threshold values for the potential and exclusion fires, which are used for the contextual algorithm (Giglio et al., 2003, 2016), were extracted from the original and solar-effect-mitigated brightnesstemperature maps in fire pixels that are provided by the MOD14 dataset. Potential fires are provided by the MOD14 collection 6 product using the variable thresholding approach, and hence the threshold values used for each fire pixel could not be known. Consequently, the potentialfire pixels were identified when the brightness temperatures in the MWIR channel were from 300 K to 325 K and the exclusion fire pixels were classified when the brightness temperatures were higher than 325 K. Fig. 7 shows the histograms of potential fire temperature (300 K < T4 < 325 K), potential fire temperature differences (\(\overline{\Delta T}\)> 10 K), exclusion fire temperature (T4 ≥ 325K), and exclusion fire temperature differences (\(\overline{\Delta T}\)≥ 20 K), which were created by using the potential and exclusion fire pixels identified from MOD14 fire pixels in the original and solar-effectmitigated images. Supposing that all the identified fire pixels are true, the minimum values of the histograms can be considered as threshold values.

OGCSBN_2019_v35n1_39_f0007.png 이미지

Fig. 7. Comparison of the brightness temperature on MOD14 collection 6 fire pixels between the original and solar-effect-mitigated images: (a) potential fire temperature (300 K < T4 < 325 K), (b) potential fire temperature differences (\(\overline{\Delta T}\) > 10 K), (c) exclusion fire temperature (T4 ≥ 325 K), and (d) exclusion fire temperature differences (\(\overline{\Delta T}\) ≥ 20 K).

The minimum temperatures of potential fires were approximately 300.15 K and 295.82 K in the original and solar-effect-mitigated images, respectively. Thus, the minimum variable threshold of the potential fire temperature T4 was determined as 295 K using the proposedmethod. The brightnesstemperature differences of potential fires were, respectively, 10028 K and 6.37 K in the original and solar-effect-mitigated images, and hence the minimum variable threshold of the potential fire temperature difference ΔT was determined as 6 K using the proposed method.And the minimum variable thresholds of the exclusion fire temperature T4 and temperature difference ΔT were also changed to 321 K and 17 K from 325 K and 20.00 K,respectively. Table 1 summarizes and compares the minimum variable thresholds of the MODIS fire algorithm and the proposed method

Table 1. Threshold values used for the MODIS fire algorithm and the proposed method (unit: K)

OGCSBN_2019_v35n1_39_t0001.png 이미지

3) Relationship between fire pixels and background pixels

The solar effect mitigation enables us to reduce the commission and omission errors caused by the fire detection contextual algorithm, because the solarreflected signals can be mitigated from the MWIR channel (He and Li, 2013). Thus, we analyzed the processing parameters used for the MODIS fire algorithm and proposed method in the extracted fire pixels. Table 2 compares the processing parameters estimated from the MODIS fire algorithm to those of the proposed method. The processing parameters are as follows: 1) mean fire pixel temperature, Mean (T4), 2) mean background temperature, Mean (\(\overline{T_4}\)), 3) mean difference deviation (MAD) background temperature, Mean (MAD(\(\overline{T_4}\))), and 4) mean MAD background temperature difference, Mean (MAD(\(\overline{\Delta T}\)). The differences between the solar-effect-mitigated and original images were approximately -2.01K, -3.40K, 0.23K and 0.17K in Mean (T4), Mean (\(\overline{T_4}\)), Mean (MAD(\(\overline{T_4}\))), and Mean (MAD(\(\overline{\Delta T}\))), respectively. When the differences were applied to the fire detection equations (Giglio et al., 2003, 2016), the differences were approximately -1.49 K and -2.76 K in Equations (8) and (9),respectively.Thismeansthat the differences are able to improve the number of fire detections.

\(T_{4}>\left(\bar{T}_{4}+3 \mathrm{MAD}\left(\bar{T}_{4}\right)\right)\)       (8)

\(\Delta T>(\overline{\Delta T}+3.5 \mathrm{MAD}(\overline{\Delta T}))\)       (9)

Table 2. Comparison of the processing parameters between the MODIS fire algorithm and the proposed method (unit: K)

OGCSBN_2019_v35n1_39_t0002.png 이미지

4. Processing results and Evaluation

A total of 52 MOD021KM images acquired inApril 2014 was used for performance validation of the proposedmethod.TheMOD14 productswere compared to the fire detection result from the proposed method. Fig. 8 shows the detailed processing workflow of the proposed method for detecting fires from MODIS images. The algorithm used for the MOD14 collection 6 (Giglio et al., 2016) was modified to improve the fire detection accuracy. The reflected solar radiance mitigation and threshold values were modified in the proposed method, as previously mentioned.

OGCSBN_2019_v35n1_39_f0008.png 이미지

Fig. 8. Detailed processing flow of the proposed fire detection method.

Fig. 9 shows the fires detected from the 52 MOD021KM images acquired in April 2014. The black solid circles show the fires detected by both the MODIS algorithm and the proposed method. The red solid circles represent the fires detected only by the proposed algorithm while the blue solid circles present the fires detected only by the MODIS algorithm. Table 3 summarizesthe number of commonly detected fires, and the fires detected by the proposed and MODIS algorithms. The total number of detected fires was 10,638 and 12,433 from the MODIS and proposed algorithms,respectively.The number oftotalfires using the proposed method increased 16.87% compared to that ofthe MOD14 product.The number of commonly detected fires by the two algorithms was 9178 and the number of differently detected fires was 3255 and 1460 for the proposed and MODIS algorithms, respectively. In order to compare the performance of the proposed and MODIS algorithms, we carried out visual analysis and a confidence level test of the MOD14 product.

OGCSBN_2019_v35n1_39_f0009.png 이미지

Fig. 9. Fire detection results in the study area in April 2014. The black solid circles show the fires detected by both the MODIS algorithm and proposed method. The red solid circles represent the fires detected only by the proposed algorithm while the blue solid circles present the fires detected only by the MODIS algorithm.

Table 3. Total number of fires detected by the MODIS fire algorithm and the proposed method​​​​​​​

OGCSBN_2019_v35n1_39_t0003.png 이미지

To evaluate the performance of the proposed method, we carried out visual analysis using a subset image from a part of Inner Mongolia and Manchuria. The performance of the proposed method could be effectively evaluated in the areas because the areas had a relatively higher reflectance in the MWIR channel. All fire pixels were found visually by searching for smoke plums and burn scars. Thus, we could classify most fire pixels into smoke and non-smoke fires. The fires with smoke plumes can be assumed to be true fires, while a non-smoke fire does not necessarily mean it is a false fire. That is, non-smoke fires can be true or false fires.

Table 4 comparesresultsfromtheMOD14 algorithm and the proposed method.The two algorithms detected a total of 1006 and 1339 fires,respectively. Smoke fires were detected a total of 781 and 1009 times as shown in Table 4. The additionally detected smoke fires comprised 281 fires. The increased detection rate of smoke fires was 29.19%. In total, 77.63% and 75.35% ofsmoke fires were detected by the MOD14 algorithm and the proposed method respectively, in the validation dataset. The results of the MOD14 fire algorithm are shown to be of slightly higher accuracy, 2.28%, compared to that of the proposed method. On the other hand, the proposed method detected more than 30% more fires.Consequently, both algorithmsshow similar levels of accuracy, but the proposed algorithm detects approximately 30% more fires.

Table 4. Comparison between the MODIS fire algorithm and the proposed method in part of Inner Mongolia and Manchuria​​​​​​​

OGCSBN_2019_v35n1_39_t0005.png 이미지

Moreover, we found that the proposed method reduced the false alarms of artificial structures or urban areas.An example of reducing the false alarm from an artificial structure effect is shown in Fig. 10. The blue circle of Fig. 10 indicates an industrial area. The MODIS algorithm detected some bright buildings as fires while the proposed algorithm did not detect them. Thisis because the solar effect was mitigated from the MWIRchannels.The red circle of Fig. 10 presentsfires with relatively low temperature.We could confirm that the fires were undetectable using the MODIS fire algorithm, but detectable using the proposed method. This also indicatesthat the solar effect was successfully mitigated from the MWIR channels. The result shows as one of many examplesthat the proposed method has a better detection performance in avoiding false alarms of artificial structures or urban areas than the MODIS fire algorithm

OGCSBN_2019_v35n1_39_f0010.png 이미지

Fig. 10. An example of exclusion false alarm from the proposed method. Blue circle shows an industrial area. Red circle shows fires with relatively low temperature. It is noted that the events did not occur on the same day (The figure on the right refers to a Google Earth image).algorithm.​​​​​​​

We analyzed the confidence level of unrecognized fire using the proposed algorithm. Fig. 11 shows cumulative distribution functions (CDFs) of the commonly detected fires and the undetected fires with respect to detection confidence. The confidence level of less than 30% in the commonly detected fires included 346 fires (approximately 3.77%) among a total of 9,178 fires while that of the undetected fires was 344 (23.56%) among a total of 1,460 fires. This means that the number of fires with low confidence in the unrecognized fire wassix times greater than that of commonly detected fire.

OGCSBN_2019_v35n1_39_f0011.png 이미지

Fig. 11. CDFs of commonly detected fires and undetected fires according to detection confidence. The black solid line shows commonly detected fires, and the black dashed line represents fires undetected by the proposed algorithm.estimation was performed by using the MODIS MCD43 16-day products acquired in 2014.​​​​​​​

5. Conclusions

Thisstudy proposed an efficient method to mitigate the solar effect in improving a fire detection algorithm. Although solar effect mitigation in the MWIR channel has been applied to fire detection in a previous study, the proposed method has some advantages in that it i) adoptstheRed channelforretrieving MWIRemissivity and ii) can correct atmospheric effects by calculating atmospheric transmissivity. To calculate the MWIR emissivity usingRed channelreflectance, we analyzed the relationship between red reflectance and MWIR emissivity using fourteen million paired samples from a one-year MODIS global dataset. Moreover, a simple formequation for atmospheric transmittance calculation of MODIS band 22 was proposed in thisstudy through MODTRAN simulations. Consequently, the MODIS fire algorithm was modified by adding solar effect mitigation of MWIR images and was used to change threshold values.

Through this study, the Red-channel-based emissivity has a better practical advantage when compared to classification-based emissivity because i) land classification-based emissivity has difficulty in reflecting actualsurface emissivity, while red channelbased emissivity can readily reflect it, and ii) red channel-based emissivity can be easily estimated from the red channel obtained at the same acquisition time. However, red channel-based MWIR emissivity has a drawback in that the Red reflectance channel is vulnerable to the aerosol effect. This contrasts with the advantages of using MWIRchannelreflectance instead of red channel reflectance as proposed by Kaufman and Remer (1994). Thus, red channel-based MWIR emissivity may be overestimated due to the aerosol effect.

The detection performance of the proposed method was evaluated using 52 MODISTerra L1Bscenes over Northeast Asia acquired in April 2014. We found that the proposed method had an improved performance of 16.87% when compared to the MODIS fire algorithm. In addition, a visual analysis and confidence level test were performed to supplement the comparison between algorithms. The visual analysis showed that the detection rate of the proposed method increased approximately 30% in the high solar reflection areas. And the false alarms from some of the artificial structures were excluded suing the proposed fire detection algorithm. The confidence level test showed that the unrecognized fires in the proposed method reject low confidence fires. Although this test is not perfect, we can suggest that the proposed algorithm has higher accuracy compared to the MODIS fire algorithm.

In conclusion, our results showed that the proposed method is very efficient in mitigating solar reflection with an improved detection rate and a higher confidence. The method does not require additional external data such as land-cover maps and auxiliary atmospheric data. Although this study was carried out using MODIS data, the proposed method could be applied to other satellite images acquiring both the red and MWIR channels. Especially for geostationary satellites such as GOES-R, HIMAWARI-8/9 and GeoKompsat-2A, the short acquisition time would greatly improve the performance of the proposed fire detection algorithm because the reflected signalsin the geostationary satellite imagesfrequently vary according to sun zenith angle.

Acknowledgements

This work wassupported by “Development of Scene Analysis & Surface Algorithms” project, funded by ETRI, which is a subproject of “Development of GeostationaryMeteorologicalSatelliteGroundSegment (NMSC-2016-01)” programfunded byNMSC(National Meteorological Satellite Center) of KMA (Korea Meteorological Administration).​​​​​​​

참고문헌

  1. Becker, F. and Z.-L. Li, 1990. Temperature-independent spectral indices in thermal infrared bands, Remote Sensing of Environment, 32(1): 17-33. https://doi.org/10.1016/0034-4257(90)90095-4
  2. Boyd, D. S., G. M. Foody, and P. J. Curran, 1999. The relationship between the biomass of Cameroonian tropical forests and radiation reflected in middle infrared wavelengths (3.0-5.0 ${\mu}m$), International Journal of Remote Sensing, 20(5): 1017-1023. https://doi.org/10.1080/014311699213055
  3. Boyd, D.S. and F. Petitcolin, 2004. Remote sensing of the terrestrial environment using middle infrared radiation (3.0-5.0 ${\mu}m$), International Journal of Remote Sensing, 25(17): 3343-3368. https://doi.org/10.1080/01431160310001654356
  4. Carlson, T. N. and D. A. Ripley, 1997. On the relation between NDVI, fractional vegetation cover, and leaf area index, Remote Sensing of Environment, 62(3): 241-252. https://doi.org/10.1016/S0034-4257(97)00104-1
  5. Chandrasekhar, S., 1960. Radiative transfer, Dover Publications Inc., New York, NY, USA.
  6. Dozier, J., 1981. A method for satellite identification of surface temperature fields in subpixel resolution, Remote Sensing of Environment, 11: 221-229. https://doi.org/10.1016/0034-4257(81)90021-3
  7. Giglio, L., J. Descloitres, C. O. Justice, and Y. J. Kaufman, 2003. An enhanced contextual fire detection algorithm for MODIS, Remote Sensing of Environment, 87(2-3): 273-282. https://doi.org/10.1016/S0034-4257(03)00184-6
  8. Giglio, L., W. Schroeder, and C. O. Justice, 2016. The collection 6 MODIS active fire detection algorithm and fire products, Remote Sensing of Environment, 178: 31-41. https://doi.org/10.1016/j.rse.2016.02.054
  9. Goita, K. and A. Royer, 1997. Surface temperature and emissivity separability over land surface from combined TIR and SWIR AVHRR data, IEEE Transactions on Geoscience and Remote Sensing, 35(3): 178-733.
  10. He, L. and Z. L. Li, 2011. Enhancement of a fire-detection algorithm by eliminating solar contamination effects and atmospheric path radiance: application to MODIS data, International Journal of Remote Sensing, 32(21): 6273-6293. https://doi.org/10.1080/01431161.2010.508057
  11. He, L. and Z. L. Li, 2012. Enhancement of a fire detection algorithm by eliminating solar reflection in the mid-IR band: application to AVHRR data, International Journal of Remote Sensing, 33(22): 7047-7059. https://doi.org/10.1080/2150704X.2012.699202
  12. Jimenez-Munoz, J.C. and J.A. Sobrino, 2003. A generalized single-channel method for retrieving land surface temperature from remote sensing data, Journal of Geophysical Research, 108(D22).
  13. Justice, C. O., L. Giglio, S. Korontzi, J. Owens, J. T. Morisette, D. Roy, J. Descloitres, S. Alleaume, F. Petitcolin, and Y. Kaufman, 2002. The MODIS fire products, Remote Sensing of Environment, 83(1-2): 244-262. https://doi.org/10.1016/S0034-4257(02)00076-7
  14. Kaufman, Y. J. and L. A. Remer, 1994. Detection of forests using Mid-IR reflectance: An application for aerosol studies, IEEE Transactions on Geoscience and Remote Sensing, 32(3): 672-683. https://doi.org/10.1109/36.297984
  15. Kaufman, Y. J., C. O. Justice, L. P. Flynn, J. D. Kendall, E. M. Prins, L. Giglio, D. E. Ward, W. P. Menzel, and A. W. Setzer, 1998. Potential global fire monitoring from EOS-MODIS, Journal of Geophysical Research: Atmospheres, 103(D24).
  16. Li, Z. L. and F. Becker, 1993. Feasibility of land surface temperature and emissivity determination from AVHRR data, Remote Sensing of Environment, 43(1): 67-85. https://doi.org/10.1016/0034-4257(93)90065-6
  17. Matson, M. and J. Dozier, 1981. Identification of subresolution high temperature sources using the thermal IR, Photogrammetric Engineering and Remote Sensing, 47(9): 1311-1318.
  18. Nerry, F., F. Petitcolin, and M. P. Stoll, 1998. Bidirectional reflectivity in AVHRR channel 3: application to a region in northern Africa, Remote Sensing of Environment, 66(3): 298-316. https://doi.org/10.1016/S0034-4257(98)00066-2
  19. Petitcolin, F. and E. Vermote, 2002. Land surface reflectance, emissivity and temperature from MODIS middle and thermal infrared data, Remote Sensing of Environment, 83(1-2): 112-134. https://doi.org/10.1016/S0034-4257(02)00094-9
  20. Prins, E. M. and W. P. Menzel, 1992. Geostationary satellite detection of biomass burning in South America, International Journal of Remote Sensing, 13(15): 2783-2799. https://doi.org/10.1080/01431169208904081
  21. Schroeder, W., P. Oliva, L. Giglio, and I. A. Csiszar, 2014. The new VIIRS 375m active fire detection data product: Algorithm description and initial assessment, Remote Sensing of Environment, 143: 85-96. https://doi.org/10.1016/j.rse.2013.12.008
  22. Schmidt, C. C., J. Hoffman, E. Prins, and S. Lindstrom, 2010. GOES-R advanced baseline imager (ABI) algorithm theoretical basis document for fire/hot spot characterization, version 2.0, NOAA NESDIS Center for Satellite Applications and Research, NOAA, Silver Spring, MD, USA.
  23. Siegert, F., B. Zhukov, D. Oertel, S. Limin, S. E. Page, and J. O. Rieley, 2004. Peat fires detected by the BIRD satellite, International Journal of Remote Sensing, 25(16): 3221-3230. https://doi.org/10.1080/01431160310001642377
  24. Snyder, W. C., Z. Wan, Y. Zhang, and Y.-Z. Feng, 1998. Classification-based emissivity for land surface temperature measurement from space, International Journal of Remote Sensing, 19(14): 2753-2774. https://doi.org/10.1080/014311698214497
  25. Sobrino, J.A., J.C. Jimenez-Munoz, G. Soria, M. Romaguera, L. Guanter, J. Moreno, A. Plaza, and P. Martinez, 2008. Land surface emissivity retrieval from different VNIR and TIR sensors, IEEE Transactions on Geoscience and Remote Sensing, 46(2): 316-327. https://doi.org/10.1109/TGRS.2007.904834
  26. Tang, H. and Z. L. Li, 2013. Quantitative remote sensing in thermal infrared: theory and applications, Springer Science & Business Media, Berlin, Germany.
  27. Valor, E. and V. Caselles, 1996. Mapping land surface emissivity from NDVI: Application to European, African, and South American areas, Remote Sensing of Environment, 57(3): 167-184. https://doi.org/10.1016/0034-4257(96)00039-9
  28. Van de Griend, A. and M. Owe, 1993. On the relationship between thermal emissivity and the normalized difference vegetation index for natural surfaces, International Journal of Remote Sensing, 14(6): 1119-1131. https://doi.org/10.1080/01431169308904400
  29. Vermote, E., D. Tanre, J. L. Deuze, M. Herman, J. J. Morcrette, and S. Y. Kotchenova, 2006. Second Simulation of a Satellite Signal in the Solar Spectrum - Vector (6SV), 6S User Guide Version 3, pp. 1-55.
  30. Wan, Z. and Z. L. Li, 1997. A physics-based algorithm for retrieving land-surface emissivity and temperature from EOS/MODIS data, IEEE Transactions on Geoscience and Remote Sensing, 35(4): 980-996. https://doi.org/10.1109/36.602541
  31. Wooster, M. J., W. Xu, and T. Nightingale, 2012. Sentinel-3 SLSTR active fire detection and FRP product: Pre-launch algorithm development and performance evaluation using MODIS and ASTER datasets, Remote Sensing of Environment, 120: 236-254. https://doi.org/10.1016/j.rse.2011.09.033
  32. Zhukov, B., E. Lorenz, D. Oertel, M. Wooster, and G. Roberts, 2006. Spaceborne detection and characterization of fires during the bi-spectral infrared detection (BIRD) experimental small satellite mission (2001-2004), Remote Sensing of Environment, 100(1): 29-51. https://doi.org/10.1016/j.rse.2005.09.019