Integrating Machine Learning and a Spatial Contextual Algorithm to Detect Wildfire from Himawari-8 Data in Southwest China

Abstract

Timely wildfire detection is helpful for fire monitoring and fighting. However, the available wildfire products with high temporal resolutions face problems, such as high omission error and commission error (false alarm) rates. This study proposed a wildfire detection algorithm combined with an improved spatial contextual algorithm and machine learning method in southwest China. First, a dataset consisting of a formation of high-confidence fire pixels combining the WLF (Himawari Wild Fire product) and VIIRS wildfire products was constructed. Then, a model to extract potential fire pixels was built using the random forest method. Additionally, an improved spatial contextual algorithm was used to identify actual fire pixels from potential fire pixels. Finally, strategies such as sun glint rejection were used to remove false alarms. As a result, the proposed algorithm performed better, with both a lower omission error rate and a lower commission error rate than the WLF product. It had a higher F1 score (0.47) than WLF (0.43) with VIIRS for reference, which means it is more suitable for wildfire detection.

1. Introduction

With the acceleration of industrialization, the amounts of wildfires worldwide have been increasing since the 1950s [1]. China is also a country that suffers from wildfire. Wildfire, which causes significant damage to the ecological environment and socio-economic development, threatens the safety of people’s lives and property and destroys surface vegetation and topsoil [2,3,4,5]. Therefore, timely and accurate wildfire detection is necessary for protecting the ecological environment and stabilizing the social economy.

Commonly used methods of wildfire detection include manual patrol monitoring [6], video monitoring [7,8,9], UAV (Unmanned Aerial Vehicle) monitoring [10,11], and remote sensing monitoring [12,13]. Manual patrol monitoring requires many human resources. Meanwhile, video and UAV monitoring have problems, such as their small monitoring scope and high cost. These problems make them difficult to apply to large-scale wildfire detection. On the other hand, remote sensing technology has an extensive monitoring range, high monitoring timeliness, and low monitoring cost, making it available for large-scale wildfire detection. Today, high spatial resolution sensors borne by polar-orbiting satellites such as MODIS, AVHRR, and VIIRS are widely used to detect wildfire. However, these satellite data have problems, such as low temporal resolution, making them unable to meet the requirements of near-real-time wildfire detection. Compared to the above data, the geostationary satellite Himawari-8 has a high temporal resolution (full-disk observation for every 10 min), stable data quality, and a large amount of data accumulation. These advantages make it more suitable for near-real-time wildfire detection.

Wildfire detection methods based on remote sensing mainly include the fixed threshold method [14,15], spatial contextual method [16], multi-temporal method [12,17,18] and machine learning method [18,19]. The main principle of the detection methods is to identify anomalies in the surface temperature [20,21].

Wildfire detection based on a fixed threshold value aims to find pixels with abnormal brightness temperatures at infrared bands [22]. Though this method can detect some wildfire pixels, its poor robustness may result in a high omission error rate [23,24]. The multi-temporal method extracts fire pixels by detecting the abnormal changes in the brightness temperature in terms of time dimension. However, the previous data required are easily contaminated by clouds, making it difficult to obtain an accurate estimation of the brightness temperature [12]. The spatial contextual wildfire detection algorithm, which is widely used in wildfire detection of satellite data, has advantages such as fewer omission errors [25,26,27]. This method needs to set a loose threshold to extract potential fire pixels in the image. Then, by comparing the brightness temperature difference between each potential fire spot and its surrounding background pixels, the non-fire spots are removed and accurate fire spot pixels are obtained [17,23,28]. However, this algorithm also has a high commission error rate. Moreover, using a fixed threshold based on experience to screen potential fire spots can easily miss out on smaller, lower-temperature fire pixels. Machine learning methods are also widely used for wildfire detection [29,30,31]. They have an efficient information-mining ability, strong learning ability, and good adaptability [18,32]. The method first analyzes and extracts the hidden features in the dataset for training. Then, a model that best fits all the samples in the training dataset is obtained through iteration [33]. However, the accuracy of machine learning models largely depends on the quality of the dataset.

Taking Yunnan, Sichuan, and Guizhou provinces as the research areas and using Himawair-8 satellite data, a high-precision, near-real-time wildfire detection algorithm combining machine learning and spatial contextual methods was proposed in this study to solve the above problems. This algorithm alleviates the problem of missing potential fire pixels extracted by fixed thresholds in traditional spatial contextual algorithms by adding a random forest model to the screening stage of potential fire pixels. On this basis, the spectral and spatial contextual information of the pixels is integrated to detect the real fire pixels in the potential fire pixels and remove false alarms. False alarm rejection methods, including sun glint rejection, desert boundary rejection, and forest-clearing rejection, are added to further remove false alarms due to the study area’s complex physical and geographical environment. Finally, the accuracy of the proposed algorithm is evaluated. The results show that the proposed algorithm has advantages in terms of the overall detection accuracy (F1 score = 0.47). Furthermore, its omission error rate (0.66) and commission error rate (0.23) are lower than the available near-real-time wildfire detection product WLF. Therefore, the results indicate that this algorithm is more suitable for near-real-time wildfire detection.

2. Materials and Methods

2.1. Study Area

The study area of this study is southwest China, including Sichuan Province, Yunnan Province, and Guizhou Province, which includes the Sichuan Basin, its surrounding mountains, and the high mountains and hills in the Yunnan–Guizhou Plateau. The terrain there is complex and varied, with an average altitude of more than 500 m (as shown in Figure 1b). The study area is a subtropical monsoon climate zone, with overall low precipitation and a relatively dry spring climate, making these areas prone to wildfire. Due to these reasons, southern Sichuan, northern and eastern Yunnan, and western Guizhou are high-risk areas for fire incidents (as shown in Figure 1a). Regarding forest resources, the depth and breadth of vegetation cover in southwest China are relatively high, and the types of forest resources are diverse [34]. The rich forest resources, complex terrain, and relatively dry climate in the study area mean that destructive fires can easily occur in these areas. Therefore, there are high requirements for wildfire monitoring and control.

2.2. Materials

2.2.1. Himawari-8 Image Data

Himawari-8 is a geostationary satellite operated by Japan Meteorological Agency (JMA). It was launched from Tanegashima Space Center using an H-IIA rocket on October 2014 and has been generating data since July 2015. The AHI (Advanced Himawari Imagers) imager borne by Himawari-8, which is greatly improved compared with those of the MTSAT series, contains 16 bands of data, from visible to infrared [35].

Table 1. Himawari-8 AHI Imager specifications.

Type	Band	$\mathbf{Wavelength} \left(\mathsf{\mu }\mathbf{m}\right)$	$\mathbf{Spatial} \mathbf{Resolution} \left(\mathbf{k}\mathbf{m}\right)$
Visible	1	0.47	1
	2	0.57
	3	0.61	0.5
Near-infrared	4	0.86	1
	5	1.6	2
	6	2.3
Infrared	7	3.9
	8	6.2
	9	6.9
	10	7.3
	11	8.6
	12	9.6
	13	10.4
	14	11.2
	15	12.4
	16	13.3

shows the specifications of the AHI imager, whose spatial resolution varies from 500 m to 2 km. The temporal resolution of AHI full-disk data is 10 min, and its scope includes regions such as East Asia and Australia. Additionally, the image data provided by Himawari-8 are in NetCDF (network Common Data Form) format. The data used in this study include albedo in visible and near-infrared bands (bands 1–6), brightness temperature in infrared bands (bands 7–16), and geometric observation information of the sun and satellite, such as solar zenith angle.

2.2.2. Wildfire Data

This study used WLF and VIIRS wildfire products as the wildfire dataset for model training and testing. See Section 2.3.1 for relevant content on dataset construction.

WLF, a wildfire position and fire radiation power product produced based on the data of the Himawari-8 satellite, is published by Japan Aerospace Exploration Agency (JAXA). It provides data between 60° N to 60° S and 80° E to 160° W, with a spatial resolution of 2 km and a temporal resolution of 10 min. Compared to WLF, the VIIRS wildfire product published by NASA, with a spatial resolution of 375 m, can provide a more accurate location of wildfire pixels and cause fewer omission errors and false alarms due to its higher spatial resolution. However, VIIRS has a low temporal resolution of four times a day. Taking both of them into account, WLF and VIIRS data at the corresponding fire time in the study area are collected to evaluate the detection effect of the algorithm quantitatively. Before calculating accuracy, samples marked as low confidence in the products are eliminated.

Some examples of fires that occurred in the study area from 2020 to 2022 are collected. On the one hand, they serve as time reference for image data collection. On the other hand, they help verify whether the algorithm can accurately detect the occurrence of wildfires in practical applications.

2.3. Methods

This study proposes a near-real-time wildfire detection algorithm based on Himawari-8 satellite data, combining random forest model with spatial contextual algorithm. The process flowchart is shown in Figure 2. For each scene of image to be detected, cloud masking and water masking are first performed to reduce the interference to subsequent process. After that, a random forest model is used to detect potential fire pixels with relatively strict fixed thresholds to reduce omission errors. Then, an improved spatial contextual algorithm is used to filter these potential fire pixels, and various false alarm rejection methods are applied to minimize commission error. Finally, two kinds of methods are used to evaluate detection accuracy.

2.3.1. Cloud Masking and Water Masking

Clouds and water can interfere with detecting wildfire in the mid-infrared band, so conducting cloud masking and water masking before wildfire detection is required. The algorithm for cloud masking varies between day and night. Pixels satisfying condition (1) are marked as night.

$$soz>85°$$

(1)

where $soz$ is the solar zenith angle. Cloud has high albedo in the visible and near-infrared channels, while its brightness temperature in the far infrared band is low. A pixel is considered to be cloud-obscured if the following conditions are satisfied:

$${bt}_{15}<265K$$

(2)

$$a_3+a_4>1.2$$

(3)

$$a_3+a_4>0.7 and {bt}_{15}<285K$$

(4)

where $a_3$ and $a_4$ are the albedo of band 3 and band 4, respectively, ${bt}_{15}$ is the brightness temperature of band 15. During the daytime, pixels that satisfy any of the conditions in 2, 3, and 4 are marked as clouds. In addition, at night, all pixels that satisfy condition 2 will be marked as clouds.

As most of the study area is inland, the majority of the water in the study area is from inland rivers and lakes, which may vary with seasons. Because of the possible differences in water distribution between seasons, using existing water products for water masking is not feasible. The proposed algorithm uses the strong absorption of water at band 6 of Himawari-8 data to extract water.

$$NDVI=\frac{a_4-a_3}{a_4+a_3}$$

(5)

$$a_6<0.05 and NDVI<0$$

(6)

where $a_6$ is the albedo of band 6, $NDVI$ is the normalized difference vegetation index calculated based on Himawari-8 image data. Pixels that satisfy condition 6 will be marked as water.

2.3.2. Construction of Wildfire Dataset

This study selected remote sensing image data and existing wildfire products at the corresponding times of fire occurrence in the study area from 2017 to 2018 to construct a wildfire dataset. The wildfire data in the dataset comes from WLF and VIIRS wildfire products. When a pixel in a Himawari-8 image is marked as having medium or high confidence of fire by both wildfire products, it is added to the dataset as a fire pixel sample. A total of 576 fire point pixels were selected as positive samples in the dataset. After that, 10 times the number of non-cloud, non-water, and non-fire pixels are randomly chosen as negative samples. Meanwhile, to minimize the impact of erroneous samples on the detection results, all fire point data in WLF and VIIRS are not involved in the construction process of negative samples. After that, all the samples are divided into test sets and training sets in a 1:4 ratio. The extracted features for each sample pixel include albedo information, brightness temperature information, angle information, and spatial contextual information. The detailed information of feature is shown in

Table 2. Features included in the dataset.

Information Category	Content and Abbreviation
Albedo	Albedo of bands 1 (a1), 2 (a2), 3 (a3), 4 (a4), and 6 (a6)
Brightness temperature	Brightness temperature of bands 7 (bt7), 11 (bt11), 14 (bt14), and 15 (bt15) and the difference between the brightness temperature of band 7 and band 14 (dt)
Angle	Solar zenith angle (soz) and satellite zenith angle (saz)
Spatial contextual information	The mean, variance, and mean deviation of the brightness temperature of the background pixels at bands 7 (meanw7, devw7, and varw7) and 14 (meanw14, and devw14, varw14) and the difference between them in the window (dmeanw, ddev, and dvarw)

, where dev means the mean deviation and w means the value calculated in the window. The determination of window information is detailed in Section 2.3.4.

To make the model detect more fires with lower temperatures, an enhanced dataset is constructed as follows: 314 selected pixels in this dataset are those marked as having medium or high confidence by VIIRS products with relatively high brightness temperatures but not recognized as fires by WLF products. These pixels may be missed due to the low spatial resolution of Himawari-8. This portion will not be divided into test sets and will all be input to the model as the training set. The distribution of all samples in the dataset is shown in Figure 3.

2.3.3. Potential Fire Pixel Detection

Detecting potential fire pixels based on fixed thresholds will miss some smaller and lower temperature fire spots. Therefore, random forest model is required to screen the potential fire spots further. Random forest model was proposed by Breiman in 2001 [36], and has been applied to remote sensing wildfire detection [37].

The essence of random forest is a combination classifier based on decision trees. Each decision tree is composed of a root node and multiple internal leaf nodes [38]. The final output of the model is obtained by voting based on the results of each independent tree. Compared with other machine learning models, random forest model has better robustness to overfitting. Moreover, random forest model can derive the importance of each feature during the training process.

Firstly, the constructed fire point dataset is divided into test set and training set at a ratio of 1:4, and the enhanced dataset and training set are merged. Then, the integrated dataset is input into the model for training. Due to the relatively small number of positive samples in the dataset, attention mechanism is introduced in the training process to give higher weights to fire samples. In the algorithm process, the random forest model will test each non-water pixel that is not cloud-obscured and output its probability of belonging to a fire spot. Pixels classified as fire by random forest model are marked as potential fire pixels.

In addition, to avoid possible omission errors in random forests, this algorithm also uses relatively strict fixed thresholds to screen potential fire spots. Pixels that satisfy conditions 7 or 8 will also be marked as potential fire spots.

$${bt}_7>307K and \left({bt}_7-{bt}_{14}\right)>7K and a_4<0.4$$

(7)

$${bt}_7>305K and \left({bt}_7-{bt}_{14}\right)>7K and night$$

(8)

where ${bt}_7$ is the brightness temperature of Himawari-8 at band 7 and ${bt}_{14}$ is the brightness temperature of Himawari-8 at band 14, and $night$ is judged by condition (1). The output of the fixed threshold algorithm is combined with the potential fire pixels obtained from the random forest model to obtain the final set of potential fire pixels.

2.3.4. Fire Detection Based on Spatial Contextual Algorithm

The aim of spatial contextual algorithm is to identify the real fire pixels in potential fire pixels by selecting appropriate background pixels around the potential fire pixels and determining whether the brightness temperature of the critical bands of the pixel is significantly different from the background ones [24,39]. It mainly includes global contextual method [40] and window contextual method. Global contextual method regards all the valid non-fire pixels in the study area as background pixels, which may lead to an inaccurate estimation of the background features of fire pixels. Window contextual method solves this problem by choosing background pixels from the ambient pixels around the potential fire pixel.

In this study, a window contextual algorithm is used to filter all the extracted potential fire pixels, which can not only reduce the false alarms generated by the random forest algorithm but also extract real fire pixels. Firstly, all pixels that satisfied condition 9 are directly marked as fire.

$${bt}_7>345K or ({bt}_7>320K and night)$$

(9)

Condition 9 indicates that the fire has reached a certain scale, resulting in a sharp increase in the brightness temperature of pixels compared to normal conditions. Therefore, it is needless to use background pixels for screening. Pixels that do not satisfy condition 9 need to be delimited around a window to select background pixels. At least eight eligible background pixels are required in a window, and at least 25% of the pixels in the window are required to be eligible background pixels. Eligible background pixels refer to pixels with non-cloud, non-water, and non-fire properties, where the background fire pixel is determined by condition 10. The study of Giglio et al. [41] shows that selecting a window within a range of 20 km around the pixel to be detected can accurately represent the background characteristics of it. However, Himawari-8 has a low spatial resolution in the infrared band, which means that only 10 pixels correspond to 20 km. So, the problem of insufficient background pixels in the window is likely to occur. To solve this problem, this study sets the maximum window size to 15 × 15, which is the range of 30 km around the potential fire pixel.

$${bt}_7>304K and \left({bt}_7-{bt}_{14}\right)>7K$$

(10)

The initial size of the window is 5 × 5. If the number of background pixels does not meet the requirements, the scope will be expanded, and the window will be retested. When the window size reaches the preset upper limit, the potential fire pixel in the center will be discarded. After determining the window size, the spatial information of background pixels is calculated.

$$dt>{dt}_a+3{dt}_d$$

(11)

$$dt>{dt}_a+4.5K$$

(12)

$${bt}_7>{bt7}_a+3{bt7}_d$$

(13)

$${bt}_{14}>{bt14}_a+{bt14}_d-4.5K$$

(14)

$${fpd}_7^{\prime }>3K$$

(15)

where dt is the brightness temperature difference between band 7 and band 14 of Himawari-8 satellite data. Subscripts a and d are the mean and mean deviation of corresponding spectral band in the background pixels within the window, respectively. ${fpd}_7^{\prime }$ is the mean deviation of brightness temperature of band 7 for all background fire pixels in the window. Pixels that satisfy conditions 16 or 17 are marked as fire pixels.

$$\left(11\right) and \left(12\right) and \left(13\right) and (\left(14\right) or (15\left)\right)$$

(16)

$$\left(11\right) and \left(12\right) and \left(13\right) and night$$

(17)

The purpose of the above criteria is to screen abnormal pixels that have significant differences in brightness temperature from surrounding background pixels. It is particularly important to point out that the aim of condition 15 is to compensate for possible misjudgment caused by condition 14 when a large fire is included in the background pixel.

2.3.5. False Alarm Rejection

This step aims to remove as many non-fire pixels as possible that have spectral similarities to fire pixels. Considering the complex natural geographical environment and climatic conditions of the study area, methods including sun glint rejection, desert boundary rejection, and forest-clearing rejection are taken to remove false alarms and reduce the commission error rate.

1.

Sun glint rejection

In some areas, such as water bodies and humid soil, the light spots reflected by the sun can be captured by sensors, resulting in the saturation of specific channel values. When the infrared band is saturated, the algorithm may recognize it as a brightness temperature anomaly, which results in a false alarm. Such false alarms can be removed effectively by calculating the angle between vectors pointing in the surface-to-satellite and specular reflection directions. The angle ${\theta }_g$ is calculated by Equation (18).

$$cos{\theta }_g=cos \left(soz\right)cos \left(saz\right)-sin \left(soz\right)sin \left(saz\right)cos\phi$$

(18)

where $soz$ and $saz$ are solar zenith angle and satellite zenith angle, respectively, and φ is the relative azimuth between satellite and sun. The criteria for sun glint rejection are as follows:

$${\theta }_g<2°$$

(19)

$${\theta }_g<10° and a_3>0.1 and a_4>0.2 and a_6>0.12$$

(20)

$${\theta }_g<15° and N_W>0$$

(21)

where $N_W$ is the total number of pixels recognized as water in the window. The addition of this indicator can reduce false alarms caused by pixels with high reflectivity around water. Pixels that satisfy any one of the conditions 19, 20, and 21 are marked as false alarms.

2.

Desert boundary rejection

When filtering fire pixels through spatial contextual algorithm, some ground pixels with higher temperatures may be incorrectly identified as background fire pixels. This may lead to a generally low brightness temperature of the background pixel in the window and result in false alarms in the detection results. By analyzing the number and spectral characteristics of fire pixels in the window, it is found that they have a higher mean and mean deviation in band 7 compared to ordinary, high-temperature background pixels. This feature can be used to remove such false alarms. The criteria are as below, where $N_{bk}$ and $N_f$ are the number of background pixels and fire pixels in the window, respectively, and ${fpa}_7^{\prime }$ is the mean of band 7 brightness temperature of all fire pixels in the window. Pixels that satisfy any one of the conditions 22, 23, and 24 are marked as false alarms.

$$N_f>0.1N_{bk} and N_f>3$$

(22)

$$a_4>0.18 and {fpa}_7^{\prime }<320K and {fpd}_7^{\prime }<2.25K$$

(23)

$${bt}_7<{fpa}_7^{\prime }+6{fpd}_7^{\prime }$$

(24)

3.

Forest-clearing rejection

The study area has a high vegetation coverage and relatively dense forests, but there is some vacant space between the trees. Due to significant spectral differences between these pixels and their surrounding ground objects, they may be mistakenly identified as fire pixels sometimes. Compared to remote sensing image data with high spatial resolution, the lower spatial resolution of Himawari-8 somewhat reduces this false alarm, but it still needs to be removed. Pixels satisfying both conditions 25 and 26 are marked as false alarms, where ${a4}_a$ is the mean of albedo of band 4 in the window. These conditions mainly aim to identify forest-clearing pixels with significantly higher brightness temperatures than surrounding vegetation pixels.

$${bt}_{14}>{bt14}_a+3.7{bt14}_d$$

(25)

$${a4}_a>0.28 and {bt}_7<325K$$

(26)

After removing false alarms from the wildfire pixels obtained by the spatial contextual algorithm in three ways, the resulting fire pixels are the output of the algorithm.

2.4. Accuracy Evaluation Methods and Indicators

This study will evaluate the accuracy of the algorithm from two aspects. On the one hand, the detection effect of the algorithm is evaluated with examples by testing images corresponding to the time of previous fires. On the other hand, the omission error rate and commission error rate of the algorithm are quantitatively calculated with WLF and VIIRS wildfire products for reference to evaluate the detection accuracy of the algorithm.

The omission error rate and commission error rate represent the degree of omission and false alarm of the wildfire pixels detected by the algorithm compared to the reference wildfire product, respectively. The omission error rate is defined by Equation (27).

$$omi=1-\frac{rightfp\left(res\right)}{totalfp\left(ref\right)}$$

(27)

where $rightfp\left(res\right)$ is the number of pixels in the reference data that coincides with the pixel position of the algorithm detection result, and $totalfp\left(ref\right)$ is the total number of pixels in the reference fire point data. The commission error rate is defined by Equation (28).

$$com=1-\frac{rightfp\left(ref\right)}{totalfp\left(res\right)}$$

(28)

where $rightfp\left(ref\right)$ is the number of pixels in the algorithm detection result that coincides with the reference data position, and $totalfp\left(res\right)$ is the total number of fire point pixels in the algorithm detection result. The omission error rate and commission error rate can intuitively characterize the accuracy of the algorithm through a large number of sample analyses. From the machine learning perspective, the omission error rate and commission error rate are defined by Equations (29) and (30), respectively.

$$omi=1-\frac{TP}{TP+FN}$$

(29)

$$com=1-\frac{TP}{TP+FP}$$

(30)

where TP, FP, and FN represent the number of true positive samples, false positive samples, and false negative samples, respectively, which can be obtained by constructing a confusion matrix. The omission error rate is 1-Recall, and the commission error rate is 1-Precision. In addition, to comprehensively evaluate the accuracy of the model, F1 score is introduced as an evaluation index. F1 score is defined by Equation (31). It can be regarded as a harmonic average of model precision and recall, and it comprehensively represents the performance of fire detection algorithms in both omissions and false alarms.

$$F1=\frac{2\times (1-com)\times (1-omi)}{2-com-omi}$$

(31)

To enhance the persuasiveness of the verification results, previous fire cases are selected to verify the algorithm. The primary purpose of this process is to test whether the algorithm can achieve forest fire monitoring in practical applications.

3. Results

3.1. Wildfire Dataset and Machine Learning Model Analysis

The feature selection in the wildfire dataset is based on the feature’s important parameters, which are derived from the random forest model and represent the degree to which each feature affects the model. The initial dataset contains more information, but some of its features are of low importance. Therefore, some almost unimportant features are removed based on computational efficiency considerations. After filtering, the importance of each feature in the dataset is shown in Figure 4. The brightness temperature of band 7 and the difference in the brightness temperature between band 7 and band 14 accounts for more than half of the importance of the model, which is the same as the basic principle.

To further evaluate the performance of the random forest model, a comparison to CNN (convolutional neural network) is used. The CNN is simplified from the model proposed by Hong et al. [18]. It consists of two convolutional layers, which have a combined output and input to an ANN consisting of nine layers of decreasing size. This study uses the image of 20 April 2021 at 9 am (UTC) in the test. As there are no VIIRS data at the corresponding time, WLF is used as reference data to evaluate whether the model can extract potential fire pixels effectively. As shown in

Table 3. Machine learning model performance evaluation with WLF for reference.

Model/Product	Total Number	Number of Same Pixels to Reference
WLF	25	25
Random forest model	45	23
CNN	24	14

, the random forest model extracts 45 potential fire pixels, of which 23 are also detected by WLF. However, CNN only detects 14 pixels in the WLF product, and the total amount of fire pixels detected is also the lowest, which means that it has a worse ability to extract potential fire pixels. CNN requires much more time and better equipment to run. Therefore, the random forest model is used in the proposed algorithm to identify potential fire pixels.

3.2. Algorithm Application Effect Demonstration

Most pixels provided by wildfire products are not caused by fire but by thermal anomalies caused by various reasons. Therefore, to further illustrate the reliability of the algorithm, actual fire cases are selected (

Table 4. Fire cases detected.

Number	Date	Latitude (N)	Longitude (E)	Site	Detected or Not
1	24 April 2021	30.4	100.24	Ganzi	detected
2	29 November 2022	32.92	102.94	Hongyuan	detected
3	16 February 2022	29.04	101.46	Jiulong	detected
4	15 March 2022	29.04	103.92	Muchuan	detected
5	28 March 2020	27.96	101.26	Muli	detected
6	20 April 2021	28.5	102.2	Mianning	detected
7	23 April 2021	26.72	99.72	Lijiang	detected
8	27 March 2021	27.32	100.82	Lijinag	detected
9	14 March 2021	25.22	102.72	Kunming	detected
10	30 March 2020	27.84	102.18	Xichang	detected
11	3 January 2021	30.24	101.68	Ganzi	missed
12	13 March 2021	28.6	101.94	Mianning	missed
13	19 March 2021	28.68	102.24	Mianning	missed
14	2 May 2021	28.46	104.46	Yibin	missed

) for verification. The algorithm proposed in this study detected ten fire cases between 2020 and 2021. The results are shown in Figure 5. However, the algorithm also missed some fire cases. All the missed fire cases have a burning area of less than 6 hectares, which is just 1.5 percent of the area of a pixel. Therefore, the abnormal brightness temperature they caused is insufficient to cause an increase in the brightness temperature of the entire pixel, which demonstrates its difficulty in being detected. Moreover, to verify the detection performance of the algorithm, a time dimension verification was conducted for the fire that occurred in Mianning on 20 April 2021 (as shown in Figure 6). The algorithm detected many fire spots when the fire happened at 16:30 (UTC + 8). Over time, the fire gradually became under control and the number of fire spots gradually decreased.

3.3. Algorithm Accuracy Evaluation

By comparing the algorithm results with the WLF data provided by JAXA, the ability of the algorithm to extract fire points is evaluated, and thereby a preliminary evaluation of the accuracy of the algorithm is achieved, as shown in

Table 5. Accuracy evaluation with WLF for reference.

Data to Be Evaluated	Reference Data	Omission Error Rate	Commission Error Rate	F1 Score
Result of the proposed algorithm	WLF	0.30	0.26	0.72

.

Using the WLF product for reference, the omission error and commission error rates of the proposed algorithm are 0.3 and 0.26, respectively. Table 5 illustrates that the detection result of the algorithm is the same as that of the WLF wildfire product for most pixels. However, due to the limitation of satellite spatial resolution, the wildfire product provided by WLF itself needs to be corrected. Therefore, comparing the two can only roughly prove the reliability of the algorithm, and it is not possible to prove whether there is an improvement in the detection results of the algorithm in the omission error rate or commission error rate. The high spatial resolution VIIRS wildfire product is chosen as a reference to solve this problem. The spatial resolution of VIIRS data is 375 m, which has high accuracy and fire sensitivity. Comparative experiments using VIIRS data as a reference can objectively evaluate the detection effect of the algorithm. The result of the accuracy evaluation is shown in

Table 6. Accuracy evaluation with VIIRS for reference.

Data to Be Evaluated	Reference Data	Omission Error Rate	Commission Error Rate	F1 Score
WLF	VIIRS	0.7	0.24	0.43
Random forest model in this study	VIIRS	0.64	0.47	0.42
Result of the proposed algorithm	VIIRS	0.66	0.23	0.47

.

Compared to the VIIRS data, the omission error and commission error rates of the WLF data are relatively high. This is mainly affected by the low spatial resolution of Himawari-8 satellite data. The omission error rate of the random forest model is reduced (0.64), whereas the commission error rate is high. After filtering, the omission error rate of the proposed algorithm (0.66) is not significantly increased compared to the random forest model, but it still decreased compared to WLF products. The commission error rate (0.23) is lower than the random forest model and the WLF product. The F1 score is used as an indicator to evaluate the comprehensive performance of the algorithm. The F1 score of the proposed algorithm (0.47) is higher than WLF, indicating that the overall accuracy of the algorithm is improved compared to the WLF product.

To further illustrate the improvements in the proposed algorithm, a comparison to other algorithms is used. Algorithms B and C are the spatial contextual algorithms proposed by Giglio et al., in 2003 [41] and 2016 [26], respectively, without any further process. As shown in

Table 7. Accuracy comparison with VIIRS for reference.

Data to Be Evaluated	Reference Data	Omission Error Rate	Commission Error Rate	F1 Score
WLF	VIIRS	0.7	0.24	0.43
Result of the proposed algorithm	VIIRS	0.66	0.23	0.47
Results of algorithm B	VIIRS	0.82	0.13	0.29
Results of algorithm C	VIIRS	0.73	0.14	0.41

, algorithms B and C have lower commission error rates in the table, but they missed more fire pixels. This is because these methods are used for a high spatial resolution MODIS imager, and may not be so suitable for the AHI of Himawari-8 satellite without any improvement. Compared with them, the proposed algorithm has a lower omission error rate and higher F1 score, and can achieve fire detection relatively accurately.

4. Discussion

Based on the Himawari-8 image data, this study improves the traditional spatial contextual algorithm by adding a random forest model to detect potential fire pixels. The algorithm performs better than the existing high-temporal resolution wildfire product WLF with a high accuracy of detection, including both the lower omission error rate and lower commission error rate [28]. Unlike these algorithms that detect wildfire with temporal information [12,17], once the image data at the corresponding time is published by JAXA, the wildfire can be detected. This means that it can achieve near-real-time active fire detection with a low publish latency of Himawari-8 data.

Due to the low spatial resolution, wildfire detection based on Himawari-8 satellite usually has a higher omission error rate and commission error rate than those found on high spatial resolution sensors. This is mainly because a larger pixel area means that the fire temperature needs to be higher to cause temperature anomalies at the pixel [42]. A wider range of background pixels may not correctly represents the environment around the fire pixel or there may not be enough background pixels in the best range proposed [41].

This study adopts various methods to reduce omission errors and false alarms for wildfire detection. To reduce omission errors, this study introduces a random forest model. The random forest model was used to detect wildfires with few fixed thresholds [37], which may lead to weak robustness and strong dependence on the dataset. Because filtering wildfire pixels just by a deep learning model needs to take both precision and recall of the model into consideration [30,31], the omission of some wildfire pixels is inevitable [18]. The proposed algorithm uses strategies to improve the recall of the machine learning model as much as possible regardless of precision, with the result that fewer wildfire pixels will be omitted. An enhanced training set is added and an attention mechanism is introduced during the training process. The fire samples in the enhanced dataset are all relatively low-temperature fire points, which can make the model more sensitive to low-temperature fires. The introduction of an attention mechanism can help improve the accuracy of the model by making it pay more attention to fire samples [9].

To reduce false alarms, the results from the random forest model are first filtered through spatial contextual information. Then, three false alarm rejection algorithms are used to eliminate false alarms further. Most spatial contextual algorithms applied to a geostationary satellite use the variance of ambient background pixels to assess whether the potential fire pixel has a significant difference in the brightness temperature compared to the background [23,27,39]. However, the variance may not be enough to describe the dispersion degree of background pixels. Therefore, the mean deviation, which is used in wildfire detection algorithms for polar-orbit satellites [26,41], is introduced to better describe the distribution of the brightness temperature around potential fire pixels. The results show that this introduction makes the algorithm perform better than those that only use the variance. The false alarm rejection algorithms were first proposed to remove false alarms in MODIS Collection 6 active fire product [26,41]. Considering the environment of the study area, this study makes some improvements to it for better performance. Compared to those without false alarm rejection, the results show that these algorithms, especially the desert boundary rejection algorithm, play an important role in the algorithm [40,43].

However, the ability of the random forest model to detect small fires can be improved, and other kinds of machine learning or deep learning models [8,18] may perform better. The false alarm rejection algorithms remove some fire pixels by mistake, which shows that these methods can be improved. Based on the above discussion, there are some areas of research that should be finished in the future. The first one is to exploit the potential of the random forest models to detect fire spots further while improving the robustness of the model by adding more samples to the dataset. At the same time, the performance of deep learning models, such as artificial neural networks and convolutional neural networks, in fire detection based on remote sensing data will be explored. The second one is to optimize the false alarm rejection algorithm. The method can effectively remove false alarms while mistakenly identifying a few real fire points as false alarms. With further improvement, the false alarm rejection algorithm can minimize the false alarm pixels that were erroneously deleted.

5. Conclusions

This study proposes a near-real-time fire detection algorithm that combines a random forest model with a spatial contextual algorithm. The proposed algorithm has a lower omission error rate and commission error rate compared to existing high-temporal wildfire detection products, making it more suitable for near-real-time wildfire detection.

First, we built a high-quality wildfire dataset by integrating the VIIRS product and the WLF product. An additional portion of positive samples that were not detected by WLF but detected by VIIRS were added to enhance the sensitivity of the model to small fires. Then, we used the dataset to train a random forest model, which will be used to screen potential fire spots. Finally, we adopted the improved spatial contextual algorithm to filter potential fire spots, and introduced false alarm rejection methods to remove false alarms. The proposed algorithm has better accuracy than the WLF product. Therefore, it has practical application value in southwest China.