Study of the Application of FY-3D/MERSI-II Far-Infrared Data in Wildfire Monitoring

Abstract

In general, the far-infrared channel in the wavelength range of 10.5–12.0 µm plays an auxiliary role in wildfire detection as its sensitivity to high-temperature targets is far lower than the mid-infrared channel in the wavelength range of 3.5–4.0 µm at the same spatial resolution (1 km, which is the spatial resolution of infrared channels in most satellites used for wildfire monitoring in daily operational mode). The Medium-Resolution Spectral Imager II onboard the Fengyun-3D polar orbiting meteorological satellite (FY-3D/MERSI-II) contains far-infrared channels with a spatial resolution of 250 m at the wavelengths of 10.8 μm and 12.0 μm, which promotes the application of far-infrared channels in wildfire monitoring. In this study, the features of FY-3D/MERSI-II far-infrared channels in fire monitoring are discussed. The sensitivity of 10.8 μm (250 m) to fire spots and the influence of solar radiation reflection on the infrared channels are quantitatively analyzed. The method of using 10.8 μm (250 m) as a major data source to detect fire spots is proposed, and several typical wildfire cases are used to verify the proposed method. The results show that the 10.8 μm (250 m) far-infrared channel has the same advantages as the existing method in wildfire monitoring in terms of a more precise positioning of the detected fire pixel, avoiding interference by solar radiation reflections, and reflecting stronger fire regions in large fire fields.

1. Introduction

Wildfire sometimes destroys the resources of forests and grasslands, causing serious losses and heavy air pollution. There are several methods of wildfire monitoring, like ground-based detection, aviation, and remote sensing, and each of them has different application features. Ground-based detection can obtain the accurate position of the fire spot, but the range of detection is limited. Aviation monitoring can obtain a larger range of detection and quite accurately detect the position of the fire; the drone technique developed in recent years can especially detect more detailed information of fire fields [1], but the range of detection is still quite limited. Remote sensing has spatial and temporal features for wildfire monitoring. Meteorological satellites, especially, have the characteristics of a large observation scope, frequent observations, and abundant detection information. Particularly, the mid-infrared channels within the wavelength range of 3.5–4.0 μm in meteorological satellites are sensitive to high-temperature targets, such as wildfire spots. Hence, the use of a meteorological satellite has become an important approach in the application and research of forest and grassland fire monitoring since the early 1980s [2,3,4,5]. For fire spot detection, various automatic methods, including the multi-band threshold method and the contextual method, have been developed [6,7,8,9]. Additionally, deep convolution neural networks have been developed to monitor fire smoke in recent years [10]. Many methods are based on the principle that the brightness temperature of the pixel containing the active wildfire detected by the mid-infrared channel is obviously higher than that of surrounding pixels, and the same can be said for the brightness temperature difference between the mid-infrared channel and the far-infrared channel. A major contribution in identifying fire spots is the contextual method from MODIS (Moderate-resolution Imaging Spectroradiometer) data, which better resolves the influence of the differences in underlying surface types and vegetation coverage between the detected pixels and surrounding pixels on the estimation of background temperature and thus has been widely adopted [11,12,13,14].

Current fire-monitoring technology is mainly based on the mid-infrared channel, while the far-infrared channel with the same spatial resolution is used to calculate the surrounding temperature for reference and as an auxiliary method for fire identification. Although the mid-infrared channel is very sensitive to fire spots, it still has some deficiencies. For example, it easily interferes with by solar radiation reflection, especially in areas with the sun glint phenomenon. Furthermore, the positioning information of flaming fires is not accurate enough. In general, polar-orbiting meteorological satellites have a spatial resolution of 1 km in the infrared channels, which covers a large area. However, more accurate locations of fire spots often need to be obtained in practical applications. As mentioned above, although the far-infrared channel is not influenced by solar radiation, the detection accuracy for fire spots is far lower than that of the mid-infrared channel with the same spatial resolution of 1 km. Therefore, the far-infrared channel is not adopted in the major approach to wildfire detection.

In recent years, the new generation polar-orbiting meteorological satellites of the Fengyun-3 (FY-3) series provides the possibility to improve this problem. The Medium-Resolution Spectral Imager II onboard FY-3D (FY-3D/MERSI-II) is equipped with two far-infrared channels with a spatial resolution of 250 m at 10.8 μm and 12.0 μm, respectively. The far-infrared channel with a spatial resolution of 250 m is far less sensitive to fire spots than the mid-infrared channel in the spectrum aspects, but its spatial resolution is four times finer than that of the mid-infrared channel (1 km). Therefore, the far-infrared channel with a spatial resolution of 250 m can detect slightly larger fire spots, and its detection ability for fire spots is significantly improved compared with that of the far-infrared channel with a spatial resolution of 1 km. Here, we refer to the far-infrared channel with a wavelength of 10.8 μm and a spatial resolution of 250 m as 10.8 μm (250 m).

In this study, the features of 10.8 μm (250 m) in FY-3D/MERSI-II are discussed. The sensitivity of 10.8 μm (250 m) to fire spots and the influence of solar radiation reflection on infrared channels are quantitatively analyzed. The response of 10.8 μm (250 m) to fire spots and the sun glint phenomenon in the real case is presented. The method of using 10.8 μm (250 m) as the major data source to detect fire spots is proposed, and several typical wildfire cases are used to verify the proposed method. The results of fire spot detection by the proposed method are discussed, including the sensitivity and accuracy of the detection, the more precise position, and avoidance of interference by solar radiation, as well as some issues to improve the proposed method. The features of the 10.8 μm (250 m) channel in fire spot detection using the proposed method are concluded in the last section.

2. Data and Methods

2.1. Data

FY-3D/MERSI-II is an improved Medium-Resolution Spectral Imager (MERSI) instrument, carried by a FY-3D polar meteorological satellite, which launched in November 2017. There are 25 channels in FY-3D/MERSI-II, and we mainly use four channels in this study, including channel 3, 4, 20, and 24. Channel 3 and channel 4 are used in determining the “Cloud” and “Water body” pixels. Channel 20 is used to analyze the sensitivity to fire spots and the influence of solar radiation in fire spot detection of the mid-infrared channel. Channel 24 is used to analyze the sensitivity to fire spots and the proposed method of the far-infrared channel in fire spot detection.

Table 1. The parameters of the relative channels of FY-3D/MERSI-II.

Channel	Wavelength (μm)	Band	Spatial Resolution (km)
3	0.650	visible	0.25
4	0.865	near infrared	0.25
20	3.8	mid infrared	1
24	10.8	far infrared	0.25

lists the parameters of these channels. Note that the spatial resolution of the far-infrared channel in the wavelength of 10.8 μm (channel 24) is 250 m, and the spatial resolution of the middle infrared channel in the wavelength of 3.8 μm (channel 20) is 1 km.

Regional images are generated from FY-3D/MERSI-II Level1 data for this study, which contain calibration and earth location information, converted to the brightness temperature for 3.8 μm and 10.8 μm channels and the reflectance for 0.86 μm and 0.65 μm channels, and projected according to the rule of equal latitude and longitude.

To compare the difference in sensitivity to fire spots between 1 km spatial resolution and 250 m spatial resolution of 10.8 µm data, an image of 10.8 µm with 1 km spatial resolution is generated which is called 10.8 µm (1 km). The brightness temperatures derived from the 3.8 µm channel and the 10.8 µm channel of FY-3D/MERSI-II are denoted by T20 and T24, respectively. The reflectances from the 0.865 μm channel and the 0.65 μm channel are denoted by R4 and R3.

2.2. The Principle of Fire Spot Detection by 10.8 µm (250 m)

In this section, we first analyze the sensitivity of 10.8 µm (250 m) to fire spots (2.2.1), and then we present an example of the response of 10.8 µm (250 m) to a fire spot in a real case (2.2.2). We also analyze the influence of solar radiation on the mid- and far-infrared channels for fire spot detection (2.2.3), which is one of the advantages for using the far-infrared (10.8 µm, 250 m) channel.

2.2.1. Analysis of the Sensitivity of 10.8 μm (250 m) to Fire Spots

There are two mid-infrared channels with a spatial resolution of 1 km at 3.8 μm and 4.05 μm and two far-infrared channels with the spatial resolution of 250 m at 10.8 μm and 12 μm in FY-3D/MERSI-II, which can be used for fire spot identification. In this study, the identification method is based on the 3.8 μm mid-infrared channel and 10.8 μm far-infrared channel. The pixels with spatial resolutions of 1 km and 250 m cover an area of 1,000,000 m² and 62,500 m², respectively. Generally, a pixel cannot be fully covered by active fire, and the temperature and radiance of the fire region may reach ten times or even hundred times higher than those of other surfaces without fire in a pixel. Therefore, the radiance of a fire pixel (a pixel with an active fire in it) can be regarded as the linear combination of the radiance in the fire region and non-fire region within this pixel [5], expressed as Equation (1).

$$N_i=P\times N_{ihi}+(1-P)\times N_{ibg}$$

(1)

where $N_i$ is the radiance of the fire pixel at channel $i$, $N_{ihi}$ and $N_{ibg}$ are the radiances of the fire region and non-fire region, respectively, and $P$ is the ratio of the area of fire region in a sub-pixel to the area of this fire pixel.

Based on the Planck function, the temperature difference $\Delta T$ between the fire pixel and the background can be expressed as follows:

$$\Delta T_i=T_i-T_{ibg}=\frac{C_2{V}_i}{Ln(1+\frac{C_1{V}_i{}^3}{N_i})}-\frac{C_2{V}_i}{Ln(1+\frac{C_1{V}_i{}^3}{N_{ibg}})}$$

(2)

where $T_i$ and $T_{ibg}$ are the brightness temperatures of the fire pixel and the background at channel $i$, respectively; $N_i$ and $N_{ibg}$ are the radiance of the fire pixel and the background; $V_i$ is the central wavenumber of channel $i$; C₁ = 1.1910659 × 10⁻⁵ mW (m² sr cm⁻⁴)⁻¹; C₂ = 1.438833 K/cm⁻¹.

Based on the central wavenumbers of the channels at 3.8 μm and 10.8 μm in FY-3D/MERSI-Ⅱ (2631.579 and 925.9259, respectively), the values in $\Delta T$ at mid- and far-infrared channels are obtained under the assumptions of $T_{iB}$ = 290 K, the fire temperature in $T_i$ is 700 K and 1000 K, and P increases from 0.0001 to 0.005 according to Equation (2). As shown in Figure 1, there are obvious differences in the brightness temperature increment between mid- and far-infrared channels in the fire region. The brightness temperature increments of the mid-infrared channel are far higher than those of the far-infrared channel (1 km). In particular, the difference is more remarkable when the area of the fire spot in the sub-pixel is small.

As the spatial resolution of the 10.8 μm channel (250 m) in FY-3D/MERSI-Ⅱ is 4 times finer than that of the 3.8 μm channel (1 km), the $P$ of the 10.8 μm channel is 16 times more than that of the 3.8 μm channel. Therefore, the difference in brightness temperature increment of fire pixel between the 10.8 μm (250 m) channel and the 3.8 μm (1 km) channel is significantly reduced. If $P$ of the far-infrared channel in Equation (2) is multiplied by 16, the ratio of the fire spot area in the same sub-pixel in the mid-infrared channel will be in accordance with that of the far-infrared channel. As shown in the curve of 10.8 μm (250 m) in Figure 1, there is certain difference in $\Delta T$ between the mid-infrared and far-infrared channels when P is small. However, the magnitude of this difference is much smaller than that in the curve for the far-infrared channel 10.8 μm (1 km) in Figure 1, and moreover, the difference decreases gradually with an increasing value of $P$.

We substitute a set of values into the parameters in Equation (2), including $T_f$ (fire temperature), used in $N_{ihi}$ (fire radiance in Equation (1)), $T_{ibg}$ (background temperature of channel $i$), and $P$. From

Table 2. Values of the brightness temperature increment between the fire pixel and surrounding pixels ($\Delta T$).

Channels	P	0.0001		0.0005		0.001		0.005
	${\mathit{T}}_{\mathit{f}}$	700 K	1000 K	700 K	1000 K	700 K	1000 K	700 K	1000 K
3.8 µm (1 km)	$\Delta \mathit{T}$	4.30	17.30	16.60	48.20	27.50	67.70	66.70	141.60
10.8 µm (1 km)		0.10	0.21	0.50	1.06	1.00	2.10	6.46	13.24
10.8 µm (250 m)		1.60	3.50	7.80	15.87	15.10	29.90	78.40	139.25

, we can see that when $T_f$ is 1000 K, $T_{ibg}$ is 290 K and $P$ is 0.0005 (about 500 m² in 1 km spatial resolution), then $\Delta T$ is 48.2 K, 1.06 K, and 15.87 K for mid-infrared, far-infrared (10.8 μm, 1 km), and (10.8 μm, 250 m), respectively, which means that the far-infrared channel (10.8 μm, 250 m) has an evident response to the very high temperature target (1000 K) and can be used to detect slightly larger flaming fires (about 500 m²).

Figure 1 and Table 2 show that FY-3D/MERSI-Ⅱ’s far-infrared channel 10.8 μm is more sensitive to high temperature fires, like flaming fires which have a temperature of around 800 K to 1200 K, and a lower sensitivity to low temperature fires, like smoldering fires which have a temperature of around 600 K to 800 K.

2.2.2. The Response of 10.8 µm (250 m) to Fire Spots in a Real Case

As analyzed in Figure 1 and Table 2, FY-3D/MERSI-II’s far-infrared channel 10.8 µm (250 m) is more sensitive to high temperature fires, and a pixel containing a small fire may cause an increase in the temperature increment. Figure 2 shows the response of 10.8 µm (250 m) to a fire spot in a real case. A forest fire occurred in Southern China on 19 January 2023. Figure 2 shows the different responses of 3.8 µm (1 km), 10.8 µm (1 km), and 10.8 µm (250 m) to the fire spot. Comparing (d), (e), and (f), we can find that the response of 3.8 µm is significant, that of 10.8 µm (250 m) is also obvious, and that of 10.8 µm (1 km) is very small.

Table 3. The brightness temperature increment and the location of the fire pixel.

	3.8 µm (1 km)	10.8 µm (1 km)	10.8 µm (250 m)
Temp. Increment (K)	45	3	15
Location	109.980°E, 22.120°N	109.980°E, 22.120°N	108.975°E, 22.115°N

lists the response of brightness temperature increment to the fire spot of 3.8 µm (1 km), 10.8 µm (1 km), and 10.8 µm (250 m) and the location. We can determine that the $\Delta T$ of 10.8 µm (250 m) is 15 K, which is much more than the surroundings, and that of 10.8 µm (1 km) is 3 K, which is very small and far less than the threshold value usually used to detect fire spots by using infrared channel data.

2.2.3. The Influence of Solar Radiation on Mid- and Far-Infrared Channels

In this section, we analyze the influence of solar radiation reflection on the mid-infrared channel in fire spot detection, which is one of the deficiencies of the mid-infrared channel in fire detection. The irradiance of solar radiation in the mid-infrared channel at 3.8 μm in FY-3D/MERSI-II is higher than the radiance of the ground target. Therefore, when the specular reflection of solar radiation occurs in a pixel in mid-infrared channels, the radiance and brightness temperature of the pixel will increase. Hence, the reflection of solar radiation in mid-infrared channels should be considered to avoid misjudgment of fire spots.

Difference of Radiance Ratio between Mid- and Far-Infrared Channels by Solar Irradiance

Figure 3 shows the emissivity of surface features (regarded as a blackbody) and solar irradiance within the wavelengths of 3.5–4.1 μm under a constant surface temperature of 300 K and under the upper limit of the brightness temperature at the 3.8 μm mid-infrared channel in FY-3D/MERSI-II (366 K). Through the comparison, it can be seen that the solar irradiance at the wavelength of 3.8 μm is obviously higher than the radiance corresponding to the upper limit of the brightness temperature at the mid-infrared channel in FY-3D/MERSI-Ⅱ, as well as the radiance of surface features with a constant surface temperature of 300 K. This indicates that the radiance and brightness temperature in the mid-infrared channel increase remarkably under the effect of the specular reflection of solar radiation.

Figure 4 shows the emissivity of surface features (regarded as a blackbody) and solar irradiance within the wavelengths of 9.5–12.0 μm under a constant surface temperature of 300 K and under the upper limit of the brightness temperature in the 10.8 μm far-infrared channel in FY-3D/MERSI-II (343 K). In contrast, the solar irradiance at the far-infrared wavebands of 9.5–12.0 μm is obviously lower than the radiance of the surface feature with a constant temperature and the radiance corresponding to the upper limit of the brightness temperature in the far-infrared channel in FY-3D/MERSI-Ⅱ. Hence, the reflection of solar radiation has little impact on far-infrared channels, which can be neglected.

Table 4. Solar irradiance, the radiance of a surface feature, and the upper limit of the brightness temperature in the 3.8 μm and 10.8 μm channels.

Wavelength of Channel	Solar Irradiance (W (m2µm)−1)	Radiance to a Temperature of Surface (W (m2 sr µm)−1)	Radiance to the Upper Brightness Temperature (W (m2 sr µm) −1)
3.8 µm	10.57	0.49 (300 K)	4.83 (366 K)
10.8 µm	0.18	9.67 (300 K)	17.03 (343 K)

lists the solar irradiance and the radiance corresponding to a constant temperature surface feature and the upper limit of the brightness temperature in the 3.8 μm and 10.8 μm channels in FY-3D/MERSI-Ⅱ. In the 3.8 µm channel, it is found that the solar irradiance is 20 times higher than the radiance corresponding to a constant temperature surface feature, and 119% higher than the radiance corresponding to the upper limit of the brightness temperature (366 K). In the 10.8 µm channel, the solar irradiance is much lower than the radiance, and is about 2% of the radiance corresponding to a constant temperature surface feature and 1% of the radiance corresponding to the upper limit of the brightness temperature.

The Effect of Sun Glint on Mid- and Far-Infrared Channels

The sun glint phenomenon presented in satellite images is caused by the specular reflection of solar radiation. Figure 5 shows a typical example of a sun glint phenomenon case which includes two days of observation of the Poyang Lake area in Southern China, where each day has a group of observation images from the 3.8 µm mid-infrared channel, 0.65 µm visible channel, and 10.8 µm far-infrared channel from FY-3D/MERSI-II. The observation time is 0540 UTC on 5 June 2021 in the images of (a–d), where no sun glint phenomenon is observed, and the other is at 0520 UTC on 6 June 2021 in the images of (e–g), where there is a strong sun glint phenomenon.

Comparing these two days’ images, we can see that the water body is very clear in the day, with no sun glint phenomenon in the images (a–d) observed on June 5, and that a strong sun glint phenomenon occurred the next day in the images (e–g). Note that the far-infrared channel 10.8 µm image has no sun glint phenomenon, which means that the far-infrared channel can avoid the influence of disturbances by solar radiation reflections.

The occurrence of sun glint is related to the combined effects of the solar azimuth angle, solar zenith angle, satellite azimuth angle, and satellite zenith angle. In this study, the combination of angles mentioned above is called the sun glint angle, which is referred to as the angle between the specular reflection of solar radiation and the satellite viewing line.

To further illustrate the impact of the specular reflection of solar radiation on the mid-infrared channel, the influences of different sun glint angles on mid-infrared, visible, and far-infrared channels were analyzed based on ocean water, as it has uniform underlying surface.

Figure 6 shows the sun glint phenomenon observed by FY-3D/MERSI-II in the ocean area near the east of Australia at 0320 UTC on 16 December 2019. (a) is the composite image from mid-infrared, near infrared, and visible channels in FY-3D/MERSI-II, which displays the phenomenon of sun glint in the ocean area east of Australia at 0320 UTC on 16 December 2019. Specifically, there is a white belt (sun glint) on the ocean east of Australia, where the brightness temperature of the mid-infrared channel and the reflectivities of the visible and near-infrared channels in this area are significantly higher than those in surrounding regions. (b) is the sun glint angle image generated by the solar azimuth angle, solar zenith angle, satellite azimuth angle, and satellite zenith angle related to (a). The grey level is darker, the sun glint angle is smaller, and the pink line within the image is the 8° line of the sun glint angle, which means that the value of the sun glint angle within the 8° line is less than 8°.

Figure 7 shows the profiles of the brightness temperature of mid-infrared, visible, and far-infrared channels and the sun glint angle along the belt of sun glint from point A (149.62°E, 20.94°S) to point B (154.62°E, 20.93°S) in Figure 6. Note that the profiles of brightness temperature fluctuate in some sections due to the blocking of clouds over the ocean. The comparison reveals that when the sun glint angle is below 8°, the brightness temperature of the mid-infrared channel and the reflectivity of the visible channel increase obviously and reach the upper limit. However, the brightness temperature of the far-infrared channel remains unchanged. Thus, the area with a sun glint angle below 8° is called the sun glint area in this study, where specular reflection is prone to occur.

The Influence of the Specular Reflection of Solar Radiation on the Sensitivity of Fire Spot Identification by Mid-Infrared Channels

The solar irradiance in the mid-infrared wave band (wavelengths between 3.5 and 4.0 μm) is much higher than the radiance corresponding to the constant temperature surface feature (300 K), and even higher than the radiance upper limit of the mid-infrared channel. Therefore, although the area proportion of the specular reflecting body in sub-pixels is small, the pixel with a specular reflecting body (such as a water body) may also have a higher radiance and brightness temperature in this pixel in mid-infrared channels. Hence, the difference in the brightness temperature of the mid-infrared channel between the pixel containing the specular reflecting body and the surroundings may reach the threshold of fire spot identification, therefore causing misjudgment.

The difference in the brightness temperature between the pixel containing the specular reflecting bodies in sub-pixels, which we called a mixed pixel, and the surroundings ($\Delta T_{MIR}$) is calculated by Equation (3):

$$\begin{array}{l}\Delta T_{MIR}=T_{MIR\_mix}-T_{MIR\_bg}\\ =\frac{C_2{V}_{MIR}}{Ln(1+\frac{C_1{V}_{MIR}{}^3}{L_{MIR\_mix}})}-\frac{C_2{V}_{MIR}}{Ln(1+\frac{C_1{V}_{MIR}{}^3}{N_{MIR\_bg}})}\end{array}$$

(3)

where $T_{MIR\_mix}$ and $T_{MIR\_bg}$ are the brightness temperatures of a mixed pixel containing the specular reflecting bodies in sub-pixels and surroundings of mid-infrared channels, respectively; $P$ is the area proportion of specular reflecting body in sub-pixel to the pixel; and $L_{MIR\_mix}$ and $L_{MIR\_bg}$ are the radiances of the mixed pixel and surroundings of mid-infrared channels, respectively.

$L_{MIR\_mix}$ is the radiance of a mixed pixel with a specular reflecting body in sub-pixels, which is calculated by Equation (4):

$$\begin{array}{l}{L}_{\mathrm{MIR}\text{\_}\mathrm{mix}}=P\times L_{\mathrm{MIR}\text{\_}\mathrm{sung}}+(1-P)\times L_{\begin{array}{l}\mathrm{MIR}\text{\_}\mathrm{bg}\\ \end{array}}\\ =P\times (L_{\mathrm{MIR}\text{\_}\mathrm{sun}}\times R_{\mathrm{sung}}+L_{\mathrm{MIR}\text{\_}\mathrm{water}})+(1-P)\times L_{}{}_{\begin{array}{l}\mathrm{MIR}\text{\_}\mathrm{bg}\\ \end{array}}\\ =P\times (L_{\mathrm{MIR}\text{\_}\mathrm{sun}}\times R_{\mathrm{sung}}+L_{\mathrm{MIR}\text{\_}\mathrm{water}}-L_{\mathrm{MIR}\text{\_}\mathrm{bg}})+L_{\mathrm{MIR}\text{\_}\mathrm{bg}}\\ =P\times (L_{\mathrm{MIR}\text{\_}\mathrm{sung}}\times R_{\mathrm{sung}})+L_{\mathrm{MIR}\text{\_}\mathrm{bg}}\end{array}$$

(4)

where $L_{MIR\_sung}$ is the radiance of the specular reflecting water body in sub-pixels, which includes the reflected solar radiance and the radiance emitted by the water body; $L_{MIR\_sun}$ is the solar irradiance in the mid-infrared channel, which adopts a value based on the meteorological industry standard, Solar Constant, and Zero Air Mass Solar Spectral Irradiance; $L_{MIR\_water}$ is the emissivity of the specular reflecting body, that is, the radiance emitted by the water body; and $R_{sung}$ is the ratio of solar radiance reflected by specular reflecting body in sub-pixel to the solar irradiance.

In Equation (4), it is assumed that the emitted radiance of a water body is equal to that of the surroundings. Specially, the reflected solar radiance varies with the reflectance $R_{sung}$ due to the specular reflectance angle and atmospheric attenuation. The wavelength of the mid-infrared channel is the same as that of FY-3D/MERSI-II (3.8 μm). In this study, $L_{MIR\_water}$ takes the value of $L_{MIR\_bg}$.

$L_{MIR\_mix}$ will be changed along with $P$ and $R_{sung}$, which subsequently leads to the variation in the brightness temperature increment in the pixel containing the specular reflecting water body. Figure 8 shows the variations in $\Delta T_{MIR}$ with different $P$ and $R_{sung}$ under the background temperature of 290 K. It can be seen that when $R_{sung}$ is high (100%) or relatively high (70%), the small specular reflecting body may also result in large $\Delta T_{MIR}$ in mid-infrared channels. In addition, when $R_{sung}$ is relatively low (50% or 30%), $\Delta T_{MIR}$ can be also increased to an equivalent level with an increase in $P$.

In order to obtain a higher sensitivity for fire spot identification, the threshold of $\Delta T_{MIR}$ (the temperature difference between the detected pixel and the background) is usually set to 8 K.

Table 5. Values of $\Delta T_{MIR}$ corresponding to different combinations of $R_{sung}$ and $P$. The values in bold represent the values exceeding the threshold of fire spot identification.

	Rsung (%)	100	70	50	30
P (%)
1		4.98	3.52	2.51	1.45
2		9.17	6.63	4.80	2.82
3		12.81	9.42	6.90	4.13
4		16.04	11.95	8.86	5.37
5		18.95	14.28	10.69	6.56
6		21.59	16.43	12.40	7.70
7		24.02	18.43	14.02	8.80

lists the values of $\Delta T_{MIR}$ corresponding to different combinations of $R_{sung}$ and $P$. When $R_{sung}$ and $P$ adopt the following sets of values, 100% and 2%, 70% and 3%, 50% and 4%, and 30% and 7%, the corresponding values of $\Delta T_{MIR}$ will exceed the threshold for fire spot identification (values in bold in Table 5, which are bigger than 8 K).

The above analysis shows that when the area of the specular reflecting body in sub-pixels reaches a certain level (the size and the sun glint angle), the radiance of the mixed pixel containing a specular reflecting body in the mid-infrared channel will increase. This will lead to an increase in the brightness temperature in the mid-infrared channel, which may reach the threshold for fire spot identification or even the upper limit of the brightness temperature at the 3.8 µm channel. It is hard to distinguish the difference between the sun glint phenomenon and the fire spot in the sun glint area; thus, in the existing method, the sun glint area will be ignored for fire spot detection to avoid misjudgment.

2.3. Method

From the analysis of the sensitivity of 10.8 µm (250 m) to fire spots in Section 2.2.1, we know that the major principle of fire detection by using infrared channels is the temperature difference between the pixel containing a fire and the background. As the temperature of the background cannot be taken from the fire pixel, we need to use the average of the temperature from the surrounding pixels around the detected pixel to evaluate the background temperature. In the evaluation, we need to select those pixels which have the possibility of the occurrence of wildfire to calculate the average temperature of the surrounding pixels and those pixels with no possibility of the occurrence of wildfire should not be involved in the background temperature evaluation, like clouds, water bodies, deserts, and very low temperature areas. High temperature pixels should not be involved in the background temperature evaluation either, as they may raise the average of the background temperature. We also need to consider the difference in the temperature of surrounding pixels, which may influence the accuracy of the evaluated temperature, so we calculated the standard deviation of the background temperature. After the background temperature is calculated, the determination on the fire spot will be conducted. The major steps in the algorithm include four parts. Firstly, each pixel is checked to see if it is a potential fire pixel or non-potential fire pixel and then given a mask for the pixel; secondly, the potential fire pixels will be selected to evaluate the background temperature and the suspected fire pixels will be not involved in background temperature evaluation; thirdly, the average and standard deviation of the background temperature will be calculated; and fourthly, the determination of the fire spot will be conducted by comparing the difference in the brightness temperature between the detected pixel and the background.

2.3.1. Masking the Potential and Non-Potential Fire Pixels

A potential fire pixel means that the covered range of that pixel has the possibility of wildfire occurrence. A non-potential fire pixel means that there is no possibility of wildfire occurrence in the coverage of that pixel, which includes types of pixels like clouds, water bodies, very low temperature areas, and deserts. We will check each pixel in the image to determine if it is a potential or non-potential fire pixel and assign a relevant mask to each pixel. The conditions of determination of non-potential fire pixels used are as below:

Cloud Determination

If the pixel satisfies the conditions in Equation (5), it will be considered as a cloud pixel and will be masked as a non-potential fire pixel.

R₃ > R_{3_cth} and T₂₄ < T_{24_cth.}

(5)

where R₃ is the reflectance of channel 3 (0.65 μm) and R_{3_cth} is the threshold of the cloud determination of channel 3. T₂₄ is the brightness temperature of channel 24 (10.8 μm) and T_{24_cth} is the threshold of the cloud determination of channel 24.

The reference value R_{3_cth} = 0.2; T_{24_cth} = 270 K.

Water Body Determination

If the pixel satisfies the conditions in Equation (6), it will be considered as a water body pixel and will be masked as a non-potential fire pixel.

R₄ < R_{4_wth} and (R₄. − R₃) < 0

(6)

where R₄ is the reflectance of channel 4 (0.86 μm) and R_{4_wth} is the threshold of the water body determination of channel 4.

The reference value R_{4_wth} = 0.1.

Low Temperature Determination

If the pixel satisfies the conditions in Equation (7), it will be considered as a low temperature pixel and will be masked as a non-potential fire pixel.

T₂₄ < T_{24_lwth}

(7)

where T_{24_wth} is the threshold of the low temperature determination of channel 24.

The reference value T_{24_lwth} = 265 K.

Desert Determination

Desert determination refers to the land-surface-type data set. If the land surface type of a pixel is desert, it will be considered as a desert pixel and will be masked as a non-potential fire pixel.

All pixels marked as non-potential fire pixels will not be considered in the subsequent processing, and the rest of the pixels will be marked as potential fire pixels.

2.3.2. Selecting the Background Pixels

The fire pixel detection will be handled pixel by pixel from the potential fire pixels. First, the pixels around the detected pixel for calculating the background temperature will be selected. Suspected fire pixels will be determined as they may raise the background temperature and cause missed fire spot detection. The determination of a suspected fire pixel is as below:

If a potential fire pixel satisfies the condition in Equation (8), it will be considered as a suspected fire pixel:

T₂₄ > (T_{24_avg} +ΔT_{24_sp}) or T₂₄ > T_{24_wmth}

(8)

where T_{24_avg} is the average temperature of 7 × 7 surrounding potential fire pixels around the detected pixel. ΔT_{24_sp} is the threshold of the brightness temperature increment. T_{24_wmth} is the threshold of the warm source pixel which may raise the background temperature.

The reference value ΔT_{24_sp} = 12 K; T_{24_wmth} = 330 K.

The number of potential fire pixels around the detected pixel will be selected for background temperature calculation and the suspected fire pixel will be excluded. In the beginning, the 7 × 7 pixels around the detected pixel are checked, excluding the detected pixel itself, and the potential fire pixels are selected from among them. If the number of selected potential pixels is not equal or bigger than 20% of the number of checked pixels or less than 8, then the range of surrounding pixels will be increased to 9 × 9, 11 × 11,…, 19 × 19; if it still does not satisfies the conditions, this pixel will be ignored in further processing.

2.3.3. Calculating the Background Temperature

After enough number of potential fire pixels are selected, the average and standard deviation of the background temperature will be calculated by using Equations (9) and (10):

$$T_{24\_\mathrm{bg}}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{T}_{24,}{}_i}$$

(9)

$$\delta T_{24\_\mathrm{bg}}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n(T_{24,}{}_i-T_{24\_\mathrm{bg}}}{)}^2}$$

(10)

where $T_{24\_\mathrm{bg}}$ and $\delta T_{24\_\mathrm{bg}}$ are the average and standard deviation of the background temperature of channel 24 (10.8 μm), respectively. n is the number of potential fire pixels selected to calculate $T_{24\_\mathrm{bg}}$ and $\delta T_{24\_\mathrm{bg}}$.

After the standard deviation is calculated, it will be checked to see if it is too small to avoid misjudgment.

If $\delta T_{24\_\mathrm{bg}}$ is less than 2 K, it will be adjusted to 2 K.

2.3.4. Fire Pixel Detection

After the background temperature is calculated, fire spot detection will be conducted as below:

If the conditions in (11) are satisfied, the potential fire pixel will be considered as a fire pixel.

$$T_{24}\ge T_{24\_\mathrm{bg}}+k\delta T_{24\_\mathrm{bg}}\mathrm{or}{T}_{24}>T_{24\_\mathrm{th}}$$

(11)

where k is the coefficient of the standard deviation of the background temperature. T_{24_fth} is the threshold of an absolute high temperature fire pixel.

The reference values are k = 4; T_{24_fth} = 340 K

2.3.5. Flow Chart of the Processing

The steps of processing for the proposed method using the FY-3D/MERSI-II far-infrared channel (10.8 μm, 250 m) are shown in the flowchart in Figure 9.

3. Results

This section introduces the results of fire detection in four typical wildfire events by using the proposed method. The first one demonstrates the increase in accuracy of the positioning of the fire spot; the second one is presenting stronger fire areas (like flaming fires) in a large range of wildfires; the third one shows that the proposed method more accurately predicts the real condition of the grass land fire line compared to the existing method; and the fourth one demonstrates that the proposed method can detect fire spots in sun glint areas where the existing method ignores to avoid misjudgment.

The accuracy of the method will be evaluated by the fire information derived from the mid-infrared channel data of FY-3D/MERSI-II with the method used in the global fire products of FY-3D/MERSI-II, which we call the existing method. The accuracy of the existing method is quite close to MODIS fire products [15]. Since the sensitivity of the mid-infrared channel is much higher than that of the far-infrared channel, we consider the result of the fire pixel detected by the existing method as the truth. In the evaluation of the accuracy, will use the distance of the longitude and latitude between the fire pixels detected by these two methods. The distance is calculated using Equation (12):

$$D_{M-F}=\sqrt{{({\lambda }_{far}-{\lambda }_{mid})}^2+{({\phi }_{far}-{\phi }_{mid})}^2}$$

(12)

where $D_{M-F}$ is the distance of longitude and latitude between the fire pixels detected by 10.8 µm (250 m) and 3.8 µm (1 km). ${\lambda }_{far},{\phi }_{far}$ is the position of the fire pixel derived from 10.8 µm (250 m) using the proposed method; and ${\lambda }_{mid},{\phi }_{mid}$ is the position of the fire pixel derived from 3.8 µm (1 km) using the existing method. Since the spatial resolution of the mid-infrared channel of FY-3D/MERSI-II is 1 km, which is close to 0.01° longitude/latitude, we consider the accuracy of fire detection by using the proposed method as passing the truth test for that fire pixel when $D_{M-F}$ is less than 0.02°.

A set of the results processed for each event is presented, including a group of images and the statistics about the fire pixels detected. The group of images includes two multiple channel composition images and two fire spot thematic images. The multiple-channel composition images are composed of the mid-infrared channel (3.8 µm), the near infrared channel (0.8 6 µm), and the visible channel (0.65 µm), or by the far-infrared channel (10.8 µm), the near infrared channel (0.86 µm), and the visible channel (0.65 µm), and can present intuitive information about the environment of fire fields, like burning areas, burned areas, non-burned areas, smoke, clouds, and water bodies. Note that the spatial resolution of the images containing 10.8 µm (250 m) is 250 m and containing 3.8 µm (1 km) is 1 km. The fire pixel thematic images present the spatial distribution of fire pixels detected by the proposed method using 10.8 µm (250 m) and the existing method using 3.8 µm (1 km) data. The statistics including the accuracy evaluation, the number of pixels, and the total size of fire pixels detected by both 10.8 µm (250 m) and 3.8 µm (1 km) will be used to analyze the features of fire detection by using 10.8 µm (250 m) data.

3.1. Forest Fire Monitoring in Southern China on 19 January 2023

A forest fire occurred in Southern China on 19 January 2023, and FY-3D/MERSI-II data are used to detect the fire spot. Figure 10a,b shows the multiple channel composition images composed by 3.8 µm, 0.86 µm, 0.65 µm channels, and by 10.8 µm, 0.86 µm, and 0.65 µm channels. Comparing Figure 10a,b, we can find that the high temperature area reflected by 3.8 µm (Figure 10a) is much bigger than 10.8 µm (Figure 10b). Figure 10c,d shows the fire thematic images generated by using the 3.8 µm channel with the existing method (Figure 10c) and by using the 10.8 µm (250 m) channel with the proposed method (Figure 10d). Comparing Figure 10c,d, we can see that the size of the fire spot detected by the existing method is much larger than that using the proposed method.

The distances of each fire pixel detected by using the proposed method to the fire pixel detected by using the existing method were calculated and the values are all 0, which means all the fire pixels detected by using the proposed method passed the accuracy test.

The results of fire spot detection using both the existing method (3.8 µm, 1 km) and the proposed method (10.8 µm, 250 m) are listed in

Table 6. The number of pixels and the total size of fire pixels detected by the existing method and the proposed method.

3.8 µm (1 km) Existing Method		10.8 µm (250 m) Proposed Method
Number of Fire Pixels	Size Coverd	Number of Fire Pixels	Size Coverd
8	9.16 km2	5	0.36 km2

, which includes the number of fire pixels detected and the size of the fire pixels covered by both methods. In Table 6, we can see the size (9.16 km²) of the fire spot detected by using the existing method is much bigger than the size (0.36 km²) detected by using the proposed method, which means the far-infrared channel (10.8 μm, 250 m) can give the more stronger fire pixels in the fire spot as it is more sensitive to very high temperature targets, like flaming fires (around 1000 K).

3.2. Large Wild Fire Monitoring in Australia on 4 January 2020

From the summer of 2019 to the beginning of 2020, a massive wildfire occurred in eastern Australia, and a wide range of forest was burned. FY-3D/MERSI-II was used to monitor this fire in the local area of eastern Australia at 0355 UTC on 4 January 2020. Figure 11a,b shows the multiple-channel composition images composed by 3.8 µm, 0.86 µm, and 0.65 µm channels, and by 10.8 µm (250 m), 0.86 µm, and 0.65 µm channels. Comparing Figure 11a,b, we find that the high temperature area reflected by 3.8 µm (Figure 11a) is significantly larger than that of 10.8 µm (Figure 11b), and many high temperature regions in Figure 11a are shaped like flakes or blocks and in Figure 11b like dots or lines. Figure 11c,d shows the fire thematic images generated by using the 3.8 µm channel with the existing method (Figure 11c) and by using the 10.8 µm (250 m) channel with the proposed method (Figure 11d). Comparing Figure 11c,d, we can see that the area of fire spot detected by using the existing method is significantly larger than that using the proposed method, and most of the bigger fire spot detected by the existing method is also detected by the proposed method, even though its size is much smaller than the existing method.

The results of fire spot detection by using both the existing method (3.8 µm, 1 km) and the proposed method (10.8 µm, 250 m) are listed in

Table 7. The statistics of the number of fire pixels and total covered size of the fire spots detected by 10.8 µm (250 m) and 3.8 µm (1 km) of FY-3D/MERSI-II data.

3.8 µm (1 km) Existing Method		10.8 µm (250 m) Proposed Method
Number of Fire Pixels	Size Coverd	Number of Fire Pixels	Size Coverd
5670	5622 km2	11,267	700.25 km2

, which include the number of fire pixels detected and the total size of the fire pixels covered by both methods. In Table 7, we can see the size (5622 km²) of the fire spot detected by using the existing method is much bigger than the size (700.25 km²) detected by using the proposed method, which means that only a small part of the fire pixels detected by the proposed method are much stronger than the others detected by the existing method, because the far-infrared channel (10.8 μm, 250 m) is more sensitive to flaming fires.

Table 8. The statistics of the distances between the fire pixels detected by 10.8 µm (250 m) and 3.8 µm (1 km) of FY-3D/MERSI-II.

Distance of Longitude/Latitude	Number of Fire Pixels Detected	Ratio
0	10,763	0.955
<0.01	10,965	0.973
<0.02	11,002	0.976
≥0.02	264	0.023

lists the statistics of the distances between the fire pixels detected by 10.8 µm (250 m) and 3.8 µm (1 km) of FY-3D/MERSI-II, which shows that more than 97% of fire pixels are detected by using the 10.8 µm (250 m) channel with the proposed method, which passed the accuracy test.

3.3. Grass Land Fire Monitoring in Northeast Asia on 19 September 2022

A grass land fire occurred in northeast Asia on 19 September 2022, and FY-3D/MERSI-II data were used to detect the fire spot. Figure 12a,b shows the multiple channel composition images composed by 3.8 µm, 0.86 µm, and 0.65 µm channels, and by 10.8 µm, 0.86 µm, and 0.65 µm channels. Comparing Figure 12a,b, we find that the high temperature area reflected by 3.8 µm (Figure 12a) is much bigger than 10.8 µm (Figure 12b). Figure 12c,d are the fire spot thematic images generated by using the 3.8 µm channel with the existing method (Figure 12c) and by using the 10.8 µm (250 m) channel with the proposed method (Figure 12d). Comparing Figure 12c,d, we can see that the size of the fire spot detected by using the existing method is much larger than that using the proposed method, and the fire spot in Figure 12c is shaped like a block and that in Figure 12d is shaped like a line.

The distances of each fire pixel detected by using the proposed method to the fire pixel detected by the existing method were calculated and the values are all 0, which means all the fire pixels detected by using the proposed method passed the accuracy test.

The results of fire spot detection by using both the existing method (3.8 µm, 1 km) and the proposed method (10.8 µm, 250 m) are listed in

Table 9. The statistics of the number of fire pixels and the total size of the fire spots detected by 10.8 µm (250 m) and 3.8 µm (1 km) of FY-3D/MERSI-II data.

3.8 µm (1 km) Existing Method		10.8 µm (250 m) Proposed Method
Number of Fire Pixels	Size Coverd	Number of Fire Pixels	Size Coverd
45	36.5 km2	57	2.88 km2

, which includes the number of fire pixels detected and the size of the fire pixels covered by both methods. In Table 9, we can see the size (36.5 km²) of the fire spot detected using the existing method is much bigger than the size (2.88 km²) detected by using the proposed method, which means the far-infrared channel (10.8 μm, 250 m) is sensitive to very high temperature targets (around 1000 K) and can provide information regarding grass land fire lines, which are flaming fires at a temperature around 1000 K.

3.4. Fire Spot Identification in Sun Glint Areas

There is a sun glint area in the observed FY-3D/MERSI-II data in eastern Australia at 0320 UTC on 16 December 2019. The FY-3D/MERSI-II far-infrared 10.8 µm (250 m) channel was used to detect a fire spot within the sun glint area. Figure 13a is the multiple channel composition image composed by 10.8 µm, 0.86 µm, and 0.65 µm channels, and there is a sun glint area in the ocean near eastern Australia which presents as a large white band due to the influence of the specular reflection of solar radiation in the 0.86 µm and 0.65 µm channels. There is also a sun glint 8° line in the sun glint area, which means the sun glint angle within the two sun glint 8° lines is less than 8°. Usually, the existing method is not used to detect fire spots in sun glint areas, as the mid-infrared channel is easy interfered with by solar radiation reflections. Figure 13b is the portion of Figure 13a in the yellow square, where two fire spots are indicated by yellow arrows. Figure 13c is the 10.8 µm (250 m) channel of Figure 13b. Figure 13d is the fire spot thematic image which has two fire spots, indicated by black arrows, detected by using the 10.8 µm (250 m) channel with the proposed method.

There are 12 fire spots containing 109 fire pixels in Figure 13d, and some fire spots are very small, at only one or two pixels; thus, such fire spots cannot be seen in the thematic image.

4. Discussion

4.1. Sensitivity

The result shows that the areas detected by using the FY-3D/MERSI-II far-infrared channel (10.8 μm, 250 m) with the proposed method are much smaller than the areas detected by using the mid-infrared channel (3.8 μm, 1 km) with the existing method. This is in accordance with the hypothesis analyzed in the principle section which presents that 10.8 μm (250 m) is more sensitive to flaming fires with temperatures around 800 K to 1200 K than smoldering fires with temperatures around 600 K to 800 K. Especially in grass land fire detection, the fire area detected by the 10.8 μm (250 m) channel is a linear-like shape, which is closer to the real case. Fuel loading in grass land is limited, and the smoldering area (charcoal fire area) during a grassland fire is generally small and lasts for a short time. Therefore, the width of the grass land fire perimeter is only several to dozens of meters, and is generally affected by the wind speed. The information of grassland fire detected by the 10.8 μm (250 m) channel can more accurately capture the spatial distribution characteristics of grassland fires.

4.2. More Precise Location

As the spatial resolution of the FY-3D/MERSI-II far-infrared channel (10.8 μm) is four times higher than the mid-infrared channel, the accuracy of the location of the fire spot detected by 10.8 μm (250 m) is much more precise, like in the example of forest fire detection in Southern China, which is very useful for forest fire departments, as in areas with complex terrains, a distance of 1 km may be blocked from sight which affects the search for fire fields.

4.3. Avoiding the Disturbance of Solar Radiation Reflection

Fire spots within a sun glint area can be detected by using the FY-3D/MERSI-II far-infrared channel (10.8 μm) with the proposed method, as the far-infrared channel is not affected by solar radiation reflection, thus avoiding missing larger fire spots in fire detection with the existing method.

4.4. Accuracy

The statistics regarding the accuracy of detection of the proposed method in the results show that more than 97% of fire pixels detected by the 10.8 μm (250 m) channel passed the accuracy test during large wildfire monitoring, in which the proposed method detected 11,267 fire pixels using the 10.8 μm (250 m) channel. Additionally, in other examples, all fire pixels detected by the proposed method using the 10.8 μm (250 m) channel passed the accuracy test.

4.5. Some Issues for Using 10.8 μm (250 m) in Fire Detection

It is worth noting that due to the warming of land by solar radiation in the daytime and the higher temperature of water bodies than land in the nighttime, there may be a misjudgment in fire spots identified only by the far-infrared channel. By referring to the brightness temperature increment in the mid-infrared channel, the misjudgment of taking high-temperature area as fire spots can be effectively avoided.

In addition, although the far-infrared channel (250 m spatial resolution) of FY-3D/MERSI-II can be used to detect small fire spots due to the improved spatial resolution, the fire areas in the pixels of the mid-infrared channel may cross several pixels in the far-infrared channel, resulting in a decrease in the brightness temperature difference between the fire spot pixel and the surrounding pixels in the far-infrared channel and failing to reach the fire spot identification threshold. Therefore, small fire spots (just reaching the fire spot identification threshold) detected by the mid-infrared channel may be not obviously reflected in the images of the far-infrared channel. A deficiency in the proposed method is that the reference value of coefficient k in Equation (11) may only be suitable for grasslands or large ranges of forests, and not suitable for areas with a mixture of different coverage of vegetation, as the temperature difference in adjacent pixels at 250 m spatial resolution is usually larger than that at 1 km spatial resolution. To further improve this issue, a group of reference values for k should be established which are suitable for various areas with different vegetation coverage, like dense forests, grasslands, a mixture of forests, crop fields, villages, and towns, which will then improve the accuracy of fire spot detection.

5. Conclusions

Compared to the existing method for fire monitoring by using meteorological satellites and the mid-infrared channel as the major data source, the FY-3D/MERSI-II far-infrared channel (10.8 μm, 250 m) used in the proposed method has the following advantages in wildfire monitoring:

(1)

Provides effective support for fire spot identification in sun glint areas, since the far-infrared channel is not disturbed by solar radiation reflection.

(2)

Raises the accuracy of the location of detected fire spots, as the spatial resolution of the FY-3D/MERSI-II far-infrared channel (10.8 μm) is 250 m, which is four times better than 1 km spatial resolution; thus, the problem that the relatively coarse resolution of 1 km cannot determine the precise position of fire fields in the fire pixel can effectively be solved.

(3)

Identifies the spatial distribution of regions with strong fires in large-scale fire fields, especially distinguishing flaming fire lines from smoldering fire in grass land fires.

Hence, the FY-3D/MERSI-II far-infrared channel (10.8 μm, 250 m) can be supplementary to the existing method of wildfire monitoring and it has a practical application value in wildfire monitoring.