A Real-Time Wildfire Monitoring Method Based on Spatiotemporal Context Features of GK2A Data

Abstract

This invention discloses a real-time wildfire monitoring method based on the spatiotemporal context features of GK2A data, belonging to the field of remote sensing technology. A window is constructed centered on potential fire points, and the size of the background window is determined based on whether the number of effective pixels within the window satisfies a corresponding formula. Using the background information of the fire points, and through absolute brightness temperature change tests and spatiotemporal context fire point determination formulas, potential fire point pixels are identified as actual fire point pixels. Finally, the generated fire point data is further filtered based on land cover type, fire point pixel confidence, normalized vegetation index, and normalized combustion index to reduce the false alarm rate of the wildfire monitoring algorithm. This method designs a complete wildfire monitoring process and adjusts internal parameters, improving the sensitivity of GK2A data in early wildfire monitoring, effectively reducing missed detections of fire point pixels, and improving the accuracy of the wildfire monitoring algorithm.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology and relates to a method for real-time wildfire monitoring based on data from the Geo-Kompsat-2A geostationary meteorological satellite. Background Technology

[0002] With the increasing severity of global warming, the frequency and intensity of natural disasters and extreme weather events are constantly rising. In recent years, frequent wildfires have posed a serious threat to human life, health, and the Earth's ecological environment. Since July 8, 2019, Australia has experienced severe bushfires, burning 12 million hectares, killing approximately 1 billion wild animals, resulting in 33 deaths and damage to over 2,500 homes. On March 30, 2020, a sudden forest wildfire broke out in Xichang City, Sichuan Province, burning over 1,000 hectares, with over 80 hectares destroyed, resulting in 19 deaths. On August 8, 2023, Maui, Hawaii, suffered a severe wildfire, burning over 6,000 hectares, killing 97 people, damaging or destroying over 2,000 buildings, and causing economic losses of approximately US$6 billion. To maintain the stability of forest ecosystems and protect people's lives and property, achieving near-real-time, high-precision, and wide-coverage wildfire monitoring is urgently needed.

[0003] In recent years, with the rapid development of remote sensing technology, the application of different types of remote sensing satellites in the field of wildfire monitoring has made significant progress. Among them, geostationary meteorological satellites can provide near real-time, high temporal resolution ground information, thereby identifying wildfires more quickly and providing decision-making suggestions to relevant departments, enabling them to respond to rescue efforts in a timely manner.

[0004] Wildfire monitoring methods based on geostationary meteorological satellites primarily rely on satellite sensors to capture infrared and thermal infrared radiation characteristics associated with wildfires. By analyzing radiation changes, the location and intensity of fire points are accurately identified. These methods include traditional spatial context algorithms and machine learning models used to detect thermal signals generated by wildfires and other natural or anthropogenic heat sources. However, despite the significant advantages of geostationary meteorological satellites in terms of temporal resolution, their monitoring technology also faces certain challenges. First, the current satellite spatial resolution is insufficient, limiting the accurate identification of fire source locations. Second, existing wildfire monitoring algorithms suffer from high false alarm rates and low accuracy, severely impacting the effectiveness of real-time monitoring. Furthermore, cloud cover and atmospheric conditions can weaken sensor performance, thereby reducing monitoring accuracy. Therefore, it is essential to fully utilize data from the next-generation high-performance geostationary meteorological satellite GK2A, deeply mine the spatial, spectral, and temporal characteristics of wildfire pixels, develop real-time wildfire monitoring technologies based on GK2A data, and construct a rapid cloud masking algorithm suitable for the southwestern region to enhance monitoring capabilities and effectively support wildfire management. Summary of the Invention

[0005] This invention provides a real-time wildfire monitoring method based on spatial, spectral, and temporal features, applicable to Geo-Kompsat-2A satellite data, to achieve accurate and efficient wildfire monitoring in Southwest China. The method employs a random forest algorithm, using a water cloud mask to remove water and cloud pixels that significantly impact wildfire detection. Simultaneously, a background window concept is introduced, setting judgment conditions based on different wildfire types to filter out all potential fire point pixels. These potential fire point pixels are then judged using spatiotemporal context information, and the false alarm rate is reduced through four methods, ultimately generating wildfire fire point distribution results.

[0006] To ensure the wildfire monitoring algorithm is unaffected by clouds and water, cloud and water pixels in GK2A imagery need to be masked and removed. Cloud pixel removal employs a random forest algorithm to construct a high-precision, low-false-alarm-rate cloud detection model. Random forests reduce overfitting risk and improve the model's generalization ability to various cloud pixel characteristics by introducing randomness among multiple decision trees. Furthermore, the random forest algorithm provides feature importance assessment, helping to understand which spectral features are most critical for cloud pixel identification, thereby improving the algorithm's interpretability and accuracy. Water pixel removal is based on a water masking algorithm, determined according to relevant formulas.

[0007] The technical solution of this invention is: a real-time wildfire monitoring method based on spatiotemporal context features of GK2A data, the method comprising:

[0008] Step 1: Collect satellite image data of the target area and train a cloud detection model. Use the trained cloud detection model to detect clouds in the newly acquired images; then identify water pixels; finally, mask and remove cloud pixels and water pixels.

[0009] Step 2: Extract potential fire points;

[0010] Pixels that meet the following formula conditions are initially identified as potential fire points:

[0011] BT 07,i >305K and BT 07,i -BT 14,i >20K

[0012] Among them, BT 07,i For the observed brightness temperature value of the target pixel in band 7, BT 14,i This represents the observed brightness temperature value of the target image in band 14; K is the unit of temperature.

[0013] Step 3: Determine the background information of potential fire pixels;

[0014] The formula for determining the background window size of potential fire pixels is as follows:

[0015] Nv >w 2 *0.25orN v >8

[0016] In the formula, w is the size of the background window, and N v The effective number of background pixels; the initial size of the background window is set to the size of a rectangular window surrounding the potential fire point pixels, and the window size is expanded until the judgment formula is met or the maximum set window size is reached;

[0017] Step 4: Determine the fire point based on the spatiotemporal context;

[0018] Step 4.1: Perform an absolute brightness temperature change test;

[0019]

[0020] in, This represents the current brightness temperature. The value is the brightness temperature change value at the previous moment. If the brightness temperature change of the target pixel exceeds 4K in a short period of time, the pixel is determined to have a thermal anomaly.

[0021] Step 4.2: If step 4.1 determines otherwise, then determine the spatiotemporal context fire point information:

[0022]

[0023] Wherein, ΔBT 07,i The brightness temperature difference value of the target pixel before and after the time step. It is the absolute value of the average brightness temperature difference of the effective pixel in the background window before and after the time; if it meets the spatiotemporal context fire point information determination conditions, then the pixel is determined to have a thermal anomaly.

[0024] Step 5: Eliminate false alarms;

[0025] Step 5.1: Calculate Z using the following formula. 07 Z ΔT :

[0026]

[0027] Among them, Z 07 Z represents the absolute deviation of the brightness temperature value of the fire pixel in band 7 within the background window. ΔT This represents the absolute deviation of the brightness temperature difference between band 7 and band 14 of the fire point pixel within the background window. ΔT represents the brightness temperature difference between bands 7 and 14. This represents the average brightness temperature difference between bands 7 and 14; δ 07 δ represents the standard deviation of the brightness temperature of the 7th band of the effective background pixels. ΔT The standard deviation of the brightness temperature difference between effective background pixels in bands 7 and 14;

[0028] Step 5.2: Calculate the confidence level;

[0029] The confidence level of each pixel is composed of a combination of five sub-confidence levels labeled C1 to C5, with each sub-confidence level ranging from the lowest confidence level of 0 to the highest confidence level of 1.

[0030] C1 = S(BT) 07 ;310K,340K)

[0031] C2=S(Z 07 ;2.5,6)

[0032] C3=S(Z ΔT ;3,6)

[0033] C4=1-S(N ac ;3,6)

[0034] C5 = 1 - S(N) aw ;3,6)

[0035]

[0036] Where, N aw N represents the number of water pixels adjacent to the central fire pixel. ac S represents the number of cloud pixels adjacent to the central fire pixel; S() is the ramp function;

[0037] Step 5.3: Compare the normalized vegetation index and normalized fire index with the surrounding area to exclude fire pixels below a certain threshold. The calculation method is as follows:

[0038]

[0039] in,

[0040] and These represent the average values ​​of NDVI and NBR of the effective background pixels within the fire pixel background window, respectively, and RMSD. NDVI and RMSD NBR ρ represents the root mean square deviation of the effective background pixels NDVI and NBR within the background window, respectively; 3,i ρ 4,i and ρ 5,i These refer to the reflectance of the red band, near-infrared band, and short-wave infrared band, respectively.

[0041] Furthermore, in step 1, the cloud detection model is a random forest model. During the training process of the random forest model, the ratio of positive sample cloud pixels to negative sample non-cloud pixels is adjusted to 2:1, and the cloud pixel dataset is divided into training set and test set in a ratio of 7:3. The random forest model parameter tuning adopts a random search method, randomly selecting parameter combinations within the set parameter range.

[0042] The method for determining water pixels is as follows:

[0043] A pixel is identified as a water pixel if it meets the following conditions:

[0044] Daytime: ρ 6,i >0.012, or, at night: abs(ρ 3,i ) < 0.01 and abs(ρ 4,i <0.01

[0045] Where, ρ 3,i ρ 4,i ρ 6,i These are the reflectances of the 3rd, 4th, and 6th bands of the GK2A image, respectively. abs(·) represents the absolute value of the band reflectance, and i represents the i-th pixel.

[0046] Furthermore, in step 2, a detection window is first set up, and potential fire points are detected by pressing the window.

[0047] The initial background window range is set to 7×7. If the number of clear sky and land pixels in the window exceeds 8, the window is considered a valid background reference area. If the initial window does not meet this condition, the range of the background window is expanded outward from the target pixel until a sufficient number of valid pixels are obtained. The maximum background window range is set to 11×11.

[0048] Furthermore, the method for further identifying individual potential fire point pixels in small-scale wildfire events in step 2 is as follows:

[0049] BT 07,i >BT 07,min +5K and BT 07,i -BT 14,i > (BT) 07 -BT 14 ) min +5K

[0050] In the formula BT 07,min The lowest brightness temperature value in the background pixels, (BT) 07 -BT 14 ) min The minimum brightness temperature difference observed in background pixels in bands 7 and 14; if this condition is met, it is considered a small-scale wildfire event.

[0051] Furthermore, the method for further identifying a large number of potential fire point pixels in step 2 is as follows:

[0052] and

[0053] In the formula This represents the average brightness temperature of band 7 in the background window. This represents the average brightness temperature difference between band 7 and band 14 of the background window.

[0054] Furthermore, the method for further identifying large wildfire pixels in step 2 is as follows:

[0055] BT 07,i >310K and and

[0056] If the condition is met, it is determined to be a large wildfire pixel.

[0057] Furthermore, the ramp function in step 5.2 is:

[0058]

[0059] Where x, α, and β are the elements at the corresponding positions in the ramp function.

[0060] The beneficial effects of this invention: This invention proposes a real-time wildfire monitoring method based on the spatiotemporal context features of the Geo-Kompsat-2A geostationary meteorological satellite. Specifically, it involves comparing the brightness temperature changes of a target pixel with those of other surrounding effective pixels, and then using a cloud detection model constructed with a random forest algorithm. This model combines red and near-infrared band information, uses a water cloud mask to remove water and cloud pixels, and performs a series of filtering steps on the generated fire point data to reduce the false alarm rate of the wildfire monitoring algorithm. This invention improves the sensitivity of GK2A data in early wildfire monitoring, effectively reduces missed wildfire pixels, and enhances the accuracy of the wildfire monitoring algorithm. Attached Figure Description

[0061] Figure 1 A schematic diagram of the research area in southwestern my country;

[0062] In the image, 1 is evergreen coniferous forest, 2 is evergreen broad-leaved forest, 3 is deciduous coniferous forest, 4 is deciduous broad-leaved forest, 5 is mixed forest, 6 is closed shrub forest, 7 is open shrub forest, 8 is tropical savanna, 9 is tropical grassland, 10 is grassland, 11 is permanent wetland, 12 is cultivated land, 13 is urban and built land, 14 is cultivated land / natural vegetation mosaic, 15 is permanent ice and snow, 16 is wasteland, and 17 is water body.

[0063] Figure 2 This is a visualization of cloud detection results for GK2A cloud products, the fixed threshold method, and the random forest algorithm.

[0064] Figure 3 This is a graph showing the pixel brightness temperature variation and its rate of change.

[0065] Figure 4 A flowchart for spatiotemporal context wildfire monitoring.

[0066] Figure 5 This is a spatial distribution map of GK2A algorithm fire points and VIIRS fire points in the study area of ​​southwestern my country. Detailed Implementation

[0067] The present invention will now be further described with reference to the accompanying drawings.

[0068] (1) Cloud water detection

[0069] To ensure that the real-time wildfire monitoring algorithm of this invention is not affected by clouds and water during the calculation process, a random forest algorithm and a discriminant formula are used to mask and remove cloud and water pixels respectively. During the training of the random forest model, for the study area (… Figure 1 Due to the characteristics of frequent clouds and fog, the ratio of positive samples (cloud pixels) to negative samples (non-cloud pixels) was adjusted to 2:1, and the cloud pixel dataset was divided into training and test sets in a 7:3 ratio. Model parameter tuning employed a random search method, randomly selecting parameter combinations within the set parameter range to quickly find near-optimal parameter configurations. The results are as follows... Figure 2 As shown. Watermark removal is based on the following formula:

[0070] (ρ 6,i >0.012) or nighttime

[0071] Nighttime: (abs(ρ) 3,i )<0.01)and(abs(ρ 4,i <0.01)

[0072] Where, ρ 3,i ρ 4,i ρ 6,i These are the reflectance values ​​for bands 3, 4, and 6 of the GK2A imagery.

[0073] (2) Potential fire point extraction

[0074] The energy release peak during vegetation burning is concentrated in the 3.9 μm region, while the energy release of wildfires is relatively weak in the far-infrared band (i.e., the thermal infrared range). Based on this spectral characteristic, the following formula is used to identify potential fire point pixels:

[0075] BT07,i >305k&BT 07,i -BT 14,i >20k

[0076] In the formula BT 07,i For the observed brightness temperature value of the target pixel in band 7, BT 14,i This refers to the observed brightness temperature value of band 14 of the target image. The two thresholds, 305 and 20, are empirical thresholds determined by comprehensively considering factors such as the spectral characteristics of wildfire pixel combustion and the intensity of wildfire combustion in the southwestern study area.

[0077] The initial background window size is set to 7×7. If the number of clear sky and land pixels within the window exceeds 8, the window is considered a valid background reference area. If the initial window does not meet this condition, the background window size is expanded outward from the target pixel until a sufficient number of valid pixels are obtained, with the maximum background window size set to 11×11.

[0078] The formula for identifying potential fire points by combining spectral and spatial characteristics is as follows:

[0079] BT 07,i >BT 07,min +5K&BT 07,i -BT 14,i > (BT) 07 -BT 14 ) min +5K

[0080] In the formula BT 07,min The lowest brightness temperature value in the background pixels, (BT) 07 -BT 14 ) min This is the minimum brightness temperature difference observed in background pixels in bands 7 and 14. This discrimination formula can effectively identify individual potential fire points in small-scale wildfire events.

[0081]

[0082] In the formula This represents the average brightness temperature of band 7 in the background window. This is the average brightness temperature difference between band 7 and band 14 of the background window. This discrimination formula can efficiently identify and extract a large number of potential fire point pixels.

[0083]

[0084] This discrimination formula is used to determine the presence of large wildfires and to identify target pixels as potential fire points.

[0085] (3) Determination of fire point background information

[0086] The formula for determining the background window size of potential fire pixels is as follows:

[0087] N v >w 2 *0.25orN v >8

[0088] In the formula, w is the size of the background window, and N v The effective number of background pixels. The initial size of the background window is set to an 11×11 rectangular window surrounding the potential fire point pixels, and the window size is expanded until the judgment formula is satisfied or a 21×21 window size is reached.

[0089] (4) Spatiotemporal context fire point determination

[0090] First, an absolute brightness temperature change test was conducted. Figure 3 ):

[0091]

[0092] in This represents the current brightness temperature. The value is the brightness temperature change value at the previous moment. If the brightness temperature change of the target pixel exceeds 4K in a short period of time, after strict water cloud pixel removal, we determine that the pixel has a thermal anomaly.

[0093] If the absolute brightness temperature change test fails, then the spatiotemporal context fire point information is determined. Figure 4 The optimal determination formula for adjustments in the Southwest region is as follows:

[0094]

[0095] Where ΔBT 07,i The brightness temperature difference value of the target pixel before and after the time step. It represents the absolute value of the average brightness temperature difference between the effective pixels of the background window at different times.

[0096] (5) False alarm removal

[0097] Fire point pixels were examined using the MCD12Q1.006 land cover product.

[0098] Referring to MODIS fire point products, the confidence level is calculated as follows:

[0099]

[0100]

[0101] These two variables represent the fire pixel within the background window (BT). 07 and BT 14The absolute deviation is calculated using the ramp function:

[0102]

[0103] When calculating the confidence level, the confidence level of each pixel is composed of a combination of five sub-confidence levels labeled C1 to C5, with each sub-confidence level ranging from 0 (lowest confidence level) to 1 (highest confidence level).

[0104] C1 = S(BT) 07 ;310K,340K)

[0105] C2=S(Z 07 ;2.5,6)

[0106] C3=S(Z ΔT ;3,6)

[0107] C4=1-S(N ac ;3,6)

[0108] C5 = 1 - S(N) aw ;3,6)

[0109]

[0110] Where N aw N represents the number of water pixels adjacent to the central fire pixel. ac The number of cloud pixels adjacent to the central fire pixel; δ 07 The standard deviation of the brightness temperature of the 7th band of the effective background pixels; δ ΔT The standard deviation represents the brightness temperature difference between the 7th and 14th bands of the effective background pixels.

[0111] By comparing the normalized vegetation index and the normalized fire index with the surrounding area, fire pixels below a certain threshold are excluded. The calculation method is as follows:

[0112]

[0113] in and These represent the average values ​​of NDVI and NBR of the effective background pixels within the fire pixel background window, respectively, and RMSD. NDVI and RMSD NBR ρ represents the root mean square deviation of the effective background pixels NDVI and NBR within the background window, respectively. 3,i ρ 4,i and ρ 5,i These refer to the reflectance of the red band, near-infrared band, and short-wave infrared band, respectively.

[0114] This invention also introduces a potential fire point detection algorithm to comprehensively identify and extract all wildfire-related pixels and effectively exclude pixels without fire information. Based on the specific temperature attributes of vegetation combustion, and combining Planck's law, the Stefan-Boltzmann law, and Wien's displacement law, the matching relationship between the energy peak of vegetation combustion and the 7th band of the GK2A satellite is derived. By comparing the radiation differences between the mid-infrared and far-infrared bands, wildfires are effectively distinguished from other high-temperature sources. Furthermore, the introduction of the background window concept further improves the identification capability of potential fire points. The background window range is expanded centered on the target pixel until a sufficient number of effective pixels are obtained. Since fire point pixels and surrounding pixels usually have significant differences in brightness temperature, this method identifies potential fire points whose spectral features may be overlooked. The mathematical expression of Planck's law is:

[0115]

[0116] Where B(λ,T) is the radiation intensity at wavelength λ at temperature T; h is Planck's constant; c is the speed of light; k is Boltzmann's constant; T is the absolute temperature of the blackbody; and λ is the wavelength of the radiation.

[0117] The mathematical expression of the Stefan-Boltzmann law is:

[0118] E = σT 4

[0119] Where E is the radiant power per unit area (unit: watts per square meter, W / m²). 2 T is the absolute temperature of the blackbody; σ is the Stefan Boltzmann constant, which is approximately 5.67 × 10⁻⁶. -8 Wm -2 K -4 .

[0120] The mathematical expression for Wien's displacement law is:

[0121] λ max =b / T

[0122] Where λ max Let λ represent the wavelength of the maximum intensity of blackbody radiation, T represent the absolute temperature of the blackbody, and b be Wien's displacement constant, which has a value of approximately 2.897 × 10⁻⁶. -3 m·K.

[0123] Brightness temperature is calculated using the following formula:

[0124] T B =(hc / Kλ)·ln -1 (1+2hc 2 / B(λ,T)λ 5 )

[0125] Where T BThis is the brightness temperature; the other physical quantities have the same meaning as described above.

[0126] To determine whether a potential fire point pixel is a real fire point pixel, it is necessary to use the pixel information surrounding the potential fire point pixel to estimate the radiation signal of the potential fire point pixel when no wildfire occurs. Using the potential fire point pixel as the center of the background window, the background value is estimated based on the identified valid background pixels within the window. Pixels containing available observation data, non-cloud / non-water pixels, and non-potential fire point pixels within the background window are all considered valid background pixels. If there are enough valid background pixels within the background window—that is, 25% of the pixels within the window are valid background pixels or the number of valid pixels is at least 8—then the next step of calculating spatiotemporal context information is performed.

[0127] To extract qualified real fire points from a large number of potential fire points and meet the needs of real-time wildfire monitoring, this paper proposes a spatiotemporal context wildfire monitoring algorithm. This algorithm is mainly based on the statistical characteristics of brightness temperature values ​​of fire point detection channels at different times at the same point in time. Wildfire monitoring is achieved by comparing the brightness temperature changes of a target pixel before and after a certain time with the changes of other valid pixels in the background window before and after a certain time. The core of this method is that even if the brightness temperature values ​​of a specific target pixel differ at different times, the statistical properties and fluctuation range of these differences are predictable. Therefore, when the observed brightness temperature value of a target pixel exceeds the maximum value of the predicted fluctuation range, it is determined to be a fire point pixel. The brightness temperature change rate is calculated using the following formula:

[0128] ΔR=ΔT / Δu

[0129] Where ΔR is the brightness temperature change rate, ΔT is the brightness temperature difference of the same pixel at different times, and Δu is the time difference between different times.

[0130] To reduce the rate of false alarms during wildfire monitoring, this invention proposes several methods for filtering generated fire point data. These methods include land cover data verification, fire point confidence screening, and comparison of normalized vegetation index (NWRI) and normalized combustion index (NCI). By verifying the pixel values ​​of the corresponding coordinates in the land cover data against the coordinates of the fire point identified by the GK2A satellite, it is determined whether the fire point was caused by vegetation burning. Based on the spectral characteristics of the fire point pixels and the surrounding cloud and water conditions, the confidence level of the fire point pixels is calculated, and high-confidence fire points are selected. Furthermore, by comparing the fire point pixels with the NWRI and NCI of the surrounding area, if the results are below a certain threshold, the fire point is removed.

[0131] Figure 1The image shows the location of the study area in southwestern my country, and the distribution of land cover types such as forests and grasslands extracted according to the MCD12Q1.006 IGBP classification scheme. The study area specifically comprises Sichuan Province, Chongqing Municipality, Yunnan Province, and Guizhou Province in the southwest. In the image, 1 is evergreen coniferous forest, 2 is evergreen broad-leaved forest, 3 is deciduous coniferous forest, 4 is deciduous broad-leaved forest, 5 is mixed forest, 6 is closed shrubland, 7 is open shrubland, 8 is tropical savanna, 9 is tropical grassland, 10 is grassland, 11 is permanent wetland, 12 is cultivated land, 13 is urban and built-up land, 14 is cultivated land / natural vegetation mosaic, 15 is permanent snow and ice, 16 is wasteland, and 17 is water body.

[0132] Figure 2 The remote sensing image taken on April 28, 2023 at 04:40 (UTC) was selected as a case study. The classification results of the three methods were presented in black and white image format, where white areas represent clear sky pixels and black areas represent cloudy pixels. Quantitative evaluation results show that the random forest algorithm has higher accuracy in cloud detection than the GK2A cloud mask dataset and the fixed threshold method, and its classification results can be used for subsequent cloud pixel removal.

[0133] Figure 3 The upper and lower images show the brightness temperature variation and rate of change of the same pixel in band 7 within a 1400-minute (approximately one-day) timeframe in GK2A imagery. At the same location throughout the day, when the pixel is unaffected by clouds or fog, or has no thermal anomalies, its brightness temperature variation trend is relatively flat. Therefore, when the brightness temperature variation of a pixel exceeds a certain range across different time intervals, the brightness temperature variation rate can be used to determine the possible presence of thermal anomalies.

[0134] Figure 4 The document demonstrates the specific process of real-time monitoring of wildfires in spatiotemporal context. Its core lies in the processing of time-series remote sensing images, while obtaining thermal anomaly information through spatiotemporal context information.

[0135] Figure 5 The paper presents the spatial distribution of fire points generated by the wildfire monitoring algorithm proposed in this invention and high-confidence fire points from the VIIRS system in the study area of ​​Southwest my country. It shows that the fire points generated by this invention maintain good consistency with the VIIRS fire point products in six wildfire events, with their location distribution being quite similar to the VIIRS fire point products, and the overall false negative rate is low. The final distribution of wildfire fire points after false alarm removal is shown below. Figure 5 As shown, the fire point of this invention maintained good consistency with the VIIRS fire point product in 6 wildfire events, the location distribution was similar to that of the VIIRS fire point product, and the overall false negative rate was low.

Claims

1. A real-time wildfire monitoring method based on spatiotemporal context features of GK2A data, the method comprising: Step 1: Collect satellite image data of the target area and train a cloud detection model. Use the trained cloud detection model to detect clouds in the newly acquired images; then identify water pixels; finally, mask and remove cloud pixels and water pixels. Step 2: Extract potential fire points; Pixels that meet the following formula conditions are initially identified as potential fire points: ； in, The observed brightness temperature value of the target pixel in band 7. The value is the observed brightness temperature of the target pixel in band 14; K is the unit of temperature. Step 3: Determine the background information of potential fire pixels; The formula for determining the background window size of potential fire pixels is as follows: or ; In the formula The size of the background window. The effective number of background pixels; the initial size of the background window is set to the size of a rectangular window surrounding the potential fire point pixels, and the window size is expanded until the judgment formula is met or the maximum set window size is reached; Step 4: Determine the fire point based on the spatiotemporal context; Step 4.1: Perform an absolute brightness temperature change test; ； in, This represents the current brightness temperature. The value is the brightness temperature change value at the previous moment. If the brightness temperature change of the target pixel exceeds 4K in a short period of time, the pixel is determined to have a thermal anomaly. Step 4.2: If step 4.1 determines otherwise, then determine the spatiotemporal context fire point information: ； ； in, The brightness temperature difference value of the target pixel before and after the time interval. It is the absolute value of the average brightness temperature difference of the effective background pixel before and after the time step; if it meets the spatiotemporal context fire point information determination conditions, the pixel is determined to have thermal anomalies. Step 5: Eliminate false alarms; Step 5.1: Calculate using the following formula , : ； ； in, This represents the average brightness temperature of band 7 in the background window. This represents the absolute deviation of the brightness temperature value of the fire pixel in band 7 within the background window. This represents the absolute deviation of the brightness temperature difference between the fire pixel and bands 7 and 14 within the background window. This indicates the brightness temperature difference between bands 7 and 14. This represents the average brightness temperature difference between bands 7 and 14; The standard deviation of the brightness temperature of the 7th band of the effective background pixels. The standard deviation of the brightness temperature difference between effective background pixels in bands 7 and 14; Step 5.2: Calculate the confidence level; The confidence level of each pixel is defined by the label. arrive It consists of five sub-confidence levels, each ranging from the lowest confidence level of 0 to the highest confidence level of 1; ； in, This represents the number of water pixels adjacent to the central fire pixel. This represents the number of cloud pixels adjacent to the central fire pixel. It is a ramp function; Step 5.3: Compare the Normalized Difference Vegetation Index (NDVI) and Normalized Burn Index (NBR) with the surrounding area to exclude fire pixels below a specific threshold. The calculation method is as follows: ； ； in, , ; and These represent the average values ​​of NDVI and NBR of the effective background pixels within the fire pixel background window, respectively. and These represent the root mean square deviations of the effective background pixels NDVI and NBR within the background window, respectively. , and These refer to the reflectance of the red band, near-infrared band, and short-wave infrared band, respectively.

2. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, In step 1, the cloud detection model is a random forest model. During the training process of the random forest model, the ratio of positive cloud pixels to negative non-cloud pixels is adjusted to 2:1, and the cloud pixel dataset is divided into training set and test set in a ratio of 7:

3. The random forest model parameter tuning adopts a random search method, randomly selecting parameter combinations within the set parameter range. The method for determining water pixels is as follows: A pixel is identified as a water pixel if it meets the following conditions: daytime: Or, at night: ; in, , , These are the reflectance values ​​for bands 3, 4, and 6 of the GK2A imagery. denoted by , where i represents the absolute value of the band reflectivity.

3. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, In step 2, a detection window is first set up, and potential fire points are detected by pressing the window. The initial background window range is set to If the number of clear sky and land pixels within the window exceeds 8, the window is considered a valid background reference area. If the initial window does not meet this condition, the background window range is expanded outward from the target pixel until a sufficient number of valid pixels are obtained. The maximum background window range is set to... .

4. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, The method for further identifying individual potential fire point pixels in small-scale wildfire events in step 2 is as follows: ； In the formula The lowest brightness temperature value in the background pixels. The minimum brightness temperature difference observed in background pixels in bands 7 and 14; if this condition is met, it is considered a small-scale wildfire event.

5. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, The method for further identifying a large number of potential fire point pixels in step 2 is as follows: ； In the formula This represents the average brightness temperature of band 7 in the background window. This represents the average brightness temperature difference between band 7 and band 14 of the background window.

6. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, The method for further identifying large wildfire pixels in step 2 is as follows: ； If the condition is met, it is determined to be a large wildfire pixel.

7. The wildfire real-time monitoring method based on spatiotemporal context features of GK2A data as described in claim 1, characterized in that, The ramp function in step 5.2 is: ； in, These are the elements at the corresponding positions in the ramp function.

Images