Fire point detection algorithm combining radiation angle normalization and prior data set

Abstract

The invention relates to a radiation angle normalization and priori data set combined fire point detection algorithm, which comprises the following steps of simulating a difference ratio of an MODIS fire point detection wave band at different satellite scanning view angles to sub-satellite point radiance by using a radiation transfer model, and normalizing atmospheric top radiance at different satellite scanning view angles to an MODIS sub-satellite point; a priori data set is constructed, the priori data set comprises a sample database of fire points and non-fire points, a fire point detection wave band is automatically selected by comparing radiation characteristics of the fire points and the non-fire points, and an appropriate fire point detection threshold value is determined according to expected precision; kappa coefficients and weights of the Kappa coefficients under different fire point test conditions are calculated through the accuracy of fire point and non-fire point recognition, and fire point probability values are generated by fusing all fire point detection algorithms. According to the invention, the obvious influence of the satellite scanning view angle on the fire point can be effectively reduced, the fire point information contained in fire disasters in different wave bands is fully mined, and the fire point detection capability is improved.

Description

Technical Field

[0001] The invention relates to the technical field of fire point monitoring based on satellite remote sensing, and in particular to a fire point detection algorithm combining radiation angle normalization with a priori data set. Background Art

[0002] Forest fire is a natural disaster that is extremely destructive and sudden, causing huge losses to humans and ecosystems. In order to minimize the losses caused by fire, it is of great significance to timely and accurately discover and grasp fire information. With the continuous development and increasing maturity of aerospace remote sensing technology, the use of remote sensing satellites to monitor forest fires has become the mainstream trend of forest fire prevention. The high temporal and spatial resolution and large-area observation characteristics of remote sensing satellites can play a key role in fire monitoring. In addition, satellite remote sensing can also provide information such as fire location, temperature, area, etc. to help track the dynamic progress of fires.

[0003] The Moderate Resolution Imaging Spectroradiometer (MODIS) has 36 channels, covering the range from visible light to thermal infrared bands, with a resolution of 250 meters to 1 kilometer, and can play an important role in fire monitoring. At present, the most commonly used fire point detection algorithms are the threshold method and the context method, which use the temperature difference between the fire point and the normal temperature ground objects in the mid-infrared and thermal infrared bands to extract the fire point information. Traditional fire monitoring algorithms have two significant defects: First, small fires are easily missed. Since the burning area of ​​small fires is small, for satellite data with relatively low resolution such as MODIS, from the pixel scale, the temperature difference between fire pixels and background pixels is not obvious, and they cannot be effectively distinguished. In particular, the detection of low-intensity fires by the MODIS contextual algorithm is significantly affected by the satellite scanning angle. When the satellite observation angle is large, the spatial range obtained by the satellite sensor is large, which reduces the proportion of small fires in the overall radiation contribution, resulting in the accuracy of small fire detection decreasing with the increase of scanning angle; second, fixed heat sources on the surface are easily misidentified as fire points. The temperature of fixed heat sources is close to the temperature of fire points, and they have similar radiation characteristics to fire points, which causes great interference to fire point detection.

[0004] The current fire point detection algorithm is insufficient in its ability to mine remote sensing data information and cannot fully mine the fire point information contained in fires burning in different bands. The accuracy of fire point detection is still greatly limited, and there are problems such as missed detection of small fire points and false detection of pseudo fire points. Therefore, it is necessary to further improve the problems existing in the fire point detection algorithm. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a fire point detection algorithm that combines radiation angle normalization with a priori data set. The radiation angle of the top atmospheric data is normalized to enhance the small fire point information. At the same time, the priori data set is used to deeply mine the fire point information and determine the threshold of fire point detection. The algorithm includes the following steps:

[0006] 1. Normalize the MODIS top atmosphere radiation angle, use the radiation transfer model to simulate the difference ratio between the MODIS fire detection band at different satellite scanning angles and the sub-satellite point radiance, characterize the impact of angle changes on MODIS top atmosphere radiation, and normalize the top atmosphere radiance at different band scanning angles to the radiance observed at the sub-satellite point;

[0007] 2. Construct a priori data set, which includes the reflectivity and brightness temperature data of fire point samples and non-fire point samples. After preprocessing MODIS data such as geometric correction, radiation brightness calculation and brightness temperature, the U.S. wildfire geographic data and MODIS data are temporally and spatially aligned to form a data set that matches the surface and satellite.

[0008] 3. Select the fire point detection band, taking into account all bands from visible light to thermal infrared, and automatically select the fire point detection band by performing band calculations and statistical analysis on the fire point and non-fire point sample data, mining the fire information contained in different bands and band combinations;

[0009] 4. Select the fire point detection threshold, based on the fire point and non-fire point prior data sets, use computer automatic analysis technology to automatically compare the radiation characteristics of the fire point and non-fire point, and automatically select the optimal threshold of the fire point test condition according to the expected accuracy;

[0010] 5. Fusion of fire point detection algorithms: Based on the prior data set, the fire point detection results corresponding to the single-band, multi-band and context band algorithms are obtained respectively. The Kappa coefficients of different fire point test conditions are calculated by the accuracy of fire point and non-fire point identification. The respective weights are calculated according to the Kappa coefficients of all test conditions. The weights of each band test condition and the detection results are substituted into the fusion algorithm formula for fire point detection to generate a fire point probability value. The higher the probability value, the more reliable the detection result of the fire point pixel.

[0011] As a preferred implementation, the input of the MODTRAN model includes control operation parameters, sensor parameters, atmospheric parameters, observation geometry and surface parameters. The previously known atmospheric profile type is used to simulate the top atmosphere radiation difference ratio r(θ) between the sub-satellite point and different satellite scanning angles. The calculated radiation difference ratio is combined with the observation zenith angle data of the MODIS sensor, and the top atmosphere radiance observed from the satellite scanning angle θ is normalized to the MODIS satellite track directly below:

[0012]

[0013] Where: Rad obs (θ) is the top-of-atmosphere radiance received by MODIS at each satellite scanning angle, is the radiance normalized to the area directly below the satellite track.

[0014] As a preferred implementation, the fire point sample database should include fires of different types and sizes and various stages of fire burning. The fire point information includes the time of fire occurrence, the time of fire extinguishing, longitude, latitude, fire area, etc.; the establishment of the non-fire point sample database is based on multi-time series remote sensing images. The selection of non-fire point samples should not only consider terrain factors, including terrain features such as mountains, plains, and hills, but also various land use / land cover types, including four typical surfaces: artificial surface, vegetation, cloud, and water body. Among them, artificial surface includes cities, roads, industrial areas, etc., which are extracted using the normalized building index. Vegetation mainly includes farmland, forests, grasslands, etc., which are selected using the normalized vegetation index. Cloud pixels include thick clouds, thin clouds, broken clouds, etc. Water bodies include oceans, rivers, lakes, etc. Cloud and water pixels are selected using the cloud water mask threshold in the MODIS fire point algorithm. High-brightness surfaces such as agricultural greenhouses and smooth buildings are selected using a combination of near-infrared, short-wave infrared reflectivity and mid-infrared brightness temperature values.

[0015] As a preferred implementation, the radiation difference between fire points and non-fire points is analyzed based on the sample database, and the wavelength bands of 0.86μm, 1.65μm, 2.13μm, 3.8μm, 3.96μm, 11.03μm and 12.02μm of the MODIS sensor are selected. The single-band, multi-band difference and context band combination method are used, and the bubble method is used to automatically calculate the fire point pixel recognition accuracy to determine the band for fire detection;

[0016] As a preferred implementation, in the process of determining the optimal threshold, the minimum brightness temperature value and the maximum brightness temperature value of the fire point in the fire point prior data set are statistically analyzed, and the fire point detection threshold is changed from the minimum brightness temperature value of the fire point prior data set to the maximum brightness temperature value in steps of K. As the threshold changes, the fire point pixel recognition accuracy (the fire point pixel accuracy refers to the proportion of correctly identified fire point pixels to the total number of fire point pixels) and the surface misjudgment rate (the surface misjudgment rate refers to the proportion of misjudged non-fire point pixels to the total number of non-fire point pixels) also change accordingly. When the threshold reaches a certain value, the fire point pixel recognition accuracy increases slowly with the change of the threshold, but the non-fire point pixel misjudgment rate increases significantly with the increase of the threshold. Based on the principle of the bubble method, the computer automatically selects the threshold with the largest accuracy as the optimal threshold under the premise of limiting the non-fire point misjudgment rate a%.

[0017] As a preferred implementation, the calculation formulas for the Kappa coefficient and weight corresponding to each band algorithm are as follows:

[0018]

[0019] Among them, p a is the probability of correct classification, p e is the probability of correct classification when the relative frequency of fire points / non-fire points is known. TP is the number of fire point pixels correctly detected by the algorithm, TN is the number of non-fire point pixels correctly detected by the algorithm, FP is the number of fire point pixels misjudged by the algorithm, FN is the number of non-fire point pixels misjudged by the algorithm, P is the number of fire point pixels in the fire point sample database, F is the number of non-fire point pixels in the non-fire point sample database, and W i is the weight of each band algorithm, Kappa i is the Kappa coefficient of each band algorithm, and n is the total number of all band algorithms.

[0020] As a preferred implementation, the fire point detection algorithm is integrated, the confidence of the detection result is adjusted, and the calculation formula of the fire probability is adopted:

[0021]

[0022] Where G is the final result of fire point detection; F i Is the result of a single fire point test condition (if it is a fire point, F i =1, if it is not a fire point, F i =0); W i is the weight of the corresponding fire point test condition; N is the total number of fire test conditions. If the probability of the weighted synthesis result reaches 1, then this pixel is a complete fire pixel. Otherwise, if the probability drops to 0, then this pixel is a complete non-fire pixel.

[0023] Due to the adoption of the above steps and methods, the technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0024] (1) Since the radiation energy received by the satellite sensor decreases with the increase of the satellite scanning angle, the detection of low-intensity fires by the MODIS context algorithm is significantly affected by the satellite scanning angle. The accuracy of small fire point detection decreases with the increase of the scanning angle. Therefore, the present invention normalizes the MODIS atmospheric top layer radiation angle, normalizes the radiance of the fire point detection band at different satellite scanning angles to the radiance of the sub-satellite point, and improves the detection accuracy of small fire points.

[0025] (2) In order to make full use of the band information of MODIS and deeply explore the information differences between fire points and non-fire points, the present invention assists in determining the fire point detection threshold based on the U.S. wildfire geographic data, uses a computer to automatically analyze the differences in radiation characteristics between fire points and non-fire points, and automatically selects the optimal threshold for fire point detection based on the fire point recognition accuracy under the control of the desired accuracy, fully exploring the fire point information contained in different bands of fire and improving the ability of fire point detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] Figure 1 is a schematic diagram of the organizational structure of the technical roadmap according to the embodiment DETAILED DESCRIPTION

[0028] The present invention provides a fire point detection algorithm combining radiation angle normalization with a priori data set. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] Reference Figure 1 ,This embodiment provides a fire point detection algorithm that combines radiation angle normalization with a priori data set.

[0030] 1. Normalize the MODIS top atmosphere radiation angle, use the radiation transfer model to simulate the difference ratio between the MODIS fire detection band at different satellite scanning angles and the sub-satellite point radiance, characterize the impact of angle changes on MODIS top atmosphere radiation, and normalize the top atmosphere radiance at different band scanning angles to the radiance observed at the sub-satellite point;

[0031] The input of the MODTRAN model includes control operation parameters, sensor parameters, atmospheric parameters, observation geometry conditions and surface parameters. The angle of radiance is normalized for the fire point detection band. The difference ratio r(θ) between different observation angles and the sub-satellite observation radiation is simulated to characterize the influence of angle change on TOA radiation. The following formula is used:

[0032] r(θ)=[Rad(θ)-Rad nadir ] / Rad nadir

[0033] Among them, Rad(θ) and Rad nadir They are the top atmosphere radiance values ​​observed by MODIS at different satellite scanning angles and directly below the satellite track (subsatellite point);

[0034] The known atmospheric profile type is used to simulate the top atmosphere radiation difference ratio r(θ) between the sub-satellite point and different satellite scanning angles. The calculated radiation difference ratio is combined with the observation zenith angle data of the MODIS sensor, and the top atmosphere radiance observed from the satellite scanning angle θ is normalized to the bottom of the MODIS satellite track:

[0035]

[0036] Where: Rad obs (θ) is the top-of-atmosphere radiance received by MODIS at each satellite scanning angle, is the radiance normalized to the area directly below the satellite track.

[0037] 2. Construct a priori data set, which includes the reflectivity and brightness temperature data of fire point samples and non-fire point samples. After preprocessing MODIS data such as geometric correction, radiation calibration, and brightness temperature, the U.S. wildfire geographic data and MODIS data are temporally and spatially aligned to form a data set that matches the surface and satellite.

[0038] The fire point sample database should include fire information of different scales and different burning stages, and select fires of different area levels in the wildfire data. The fire point information includes the time of fire occurrence, fire extinguishing time, longitude, latitude, fire area, etc.

[0039] The establishment of the non-fire point sample database is based on multi-time series remote sensing images. The selection of non-fire point samples should not only consider terrain factors, including mountains, plains, hills and other terrain features, but also various land use / land cover types, including four typical surfaces: artificial surface, vegetation, cloud, and water body. Among them, artificial surface includes cities, roads, industrial areas, etc., which are extracted using the normalized building index. Vegetation mainly includes farmland, forest, grassland, etc., which are selected using the normalized vegetation index. Cloud pixels include thick clouds, thin clouds, broken clouds, etc. Water bodies include oceans, rivers, lakes, etc. Cloud and water pixels are selected using the cloud and water mask threshold in the MODIS fire point algorithm. High-brightness surfaces such as agricultural greenhouses and smooth buildings are selected using a combination of near-infrared, short-wave infrared reflectivity and mid-infrared brightness temperature values.

[0040] (1) Selection of building samples

[0041] The normalized building index NDBI is selected to detect buildings. It is generally believed that when the NDBI is greater than 0, the pixel point is a building.

[0042]

[0043] Where: 2.1 , 0.86Represent the apparent reflectance of the shortwave infrared band and the near infrared band, respectively.

[0044] (2) Highlight surface selection

[0045] The algorithm of the Visible Infrared Imaging Radiometer Suite (VIIRS) 375m active fire detection product was used as reference, and the apparent reflectance of the shortwave infrared and near infrared and the brightness temperature threshold of the 21 / 22 channel of the thermal infrared band were used for selection.

[0046] (ρ 2.1 >0.3&&ρ 2.1 >ρ 0.86 &&ρ 0.86 >0.25)&&BT4≤335K

[0047] Where: 2.1 , 0.86 They represent the apparent reflectance of the shortwave infrared and near-infrared bands respectively, and BT4 is the brightness temperature of the 21 / 22 channel in the mid-infrared band.

[0048] (3) Vegetation sample selection

[0049] Vegetation type selection uses the Normalized Difference Vegetation Index (NDVI) to set a threshold for selection. Pixels that meet this condition are considered to belong to the vegetation type.

[0050]

[0051] Where: 0.86 , 0.65 Represent the apparent reflectance of the near-infrared band and the red light band, respectively.

[0052] (4) Cloud pixel selection

[0053] The cloud detection method adopts the cloud detection algorithm in the latest version of MODIS C6 algorithm, and uses the apparent reflectance of MODIS red band and near infrared band and the brightness temperature of channel 32 of thermal infrared band to identify clouds:

[0054] (ρ 0.65 +ρ 0.86 >1.2)or(BT 12 <265K)or

[0055] (ρ 0.65 +ρ 0.86 >0.7andBT 12 <285K)

[0056] Where: 0.86 , 0.65are the apparent reflectances of the near-infrared band and the red light band, BT 12 is the brightness temperature of 32 bands.

[0057] (5) Water pixel selection

[0058] Water body pixel detection uses the apparent reflectance of the MODIS near-infrared band and short-wave infrared band and the normalized difference vegetation index (NDVI) for extraction:

[0059]

[0060] Where: 0.86 , 0.65 Represent the apparent reflectance of the near-infrared band and the red light band, respectively.

[0061] 3. Select the fire point detection band, taking into account all bands in the visible light to thermal infrared range, and through band calculations and statistical analysis of fire point and non-fire point sample data, dig out the fire information contained in different bands and band combinations, and automatically select the fire point detection band. For single band and multi-band, based on the sample database, the computer uses the bubbling method to automatically calculate the accuracy of fire point recognition based on the reflectivity and brightness temperature data of the pre-selected single band and multi-band, and select the fire point detection band. For the bands of the context algorithm, the bubbling classification method is used to automatically calculate the accuracy of the context band in judging fire, so as to select the context band according to the fire point discrimination effect;

[0062] 4. Select the fire point detection threshold, based on the fire point and non-fire point prior data sets, use computer automatic analysis technology to automatically compare the radiation characteristics of the fire point and non-fire point, and automatically select the optimal threshold of the fire point test condition according to the expected accuracy;

[0063] Based on the sample database, the radiation difference between fire points and non-fire points was analyzed, and the wavelength bands of MODIS sensor were selected, including 0.86μm, 1.65μm, 2.13μm, 3.8μm, 3.96μm, 11.03μm and 12.02μm. The bubble method was used to automatically calculate the accuracy of fire point pixel recognition and determine the band for fire detection by using single-band, multi-band difference and context band combination.

[0064] In the process of determining the optimal threshold, the minimum and maximum brightness temperature values ​​of the fire points in the fire point prior data set are counted, and the fire point detection threshold is changed from the minimum brightness temperature value of the fire point prior data set to the maximum brightness temperature value in steps of K. As the threshold changes, the fire point pixel recognition accuracy (the fire point pixel accuracy refers to the proportion of correctly identified fire point pixels to the total number of fire point pixels) and the surface misjudgment rate (the surface misjudgment rate refers to the proportion of misjudged non-fire point pixels to the total number of non-fire point pixels) also change accordingly. When the threshold reaches a certain value, the fire point pixel recognition accuracy increases slowly with the change of the threshold, but the non-fire point pixel misjudgment rate increases significantly with the increase of the threshold. Based on the principle of the bubble method, the computer automatically selects the threshold with the largest accuracy as the optimal threshold under the premise of limiting the non-fire point misjudgment rate a%.

[0065] 5. Fusion of fire point detection algorithms: Based on the prior data set, the fire point detection results corresponding to the single-band, multi-band and context band algorithms are obtained respectively. The Kappa coefficients of different fire point test conditions are calculated by the accuracy of fire point and non-fire point identification. The respective weights are calculated according to the Kappa coefficients of all test conditions. The weights of each band test condition and the detection results are substituted into the fusion algorithm formula for fire point detection to generate a fire point probability value. The higher the probability value, the more reliable the detection result of the fire point pixel.

[0066] As a preferred implementation, the calculation formulas for the Kappa coefficient and weight corresponding to each band algorithm are as follows:

[0067]

[0068] Among them, p a is the probability of correct classification, p e is the probability of correct classification when the relative frequency of fire points / non-fire points is known. TP is the number of fire point pixels correctly detected by the algorithm, TN is the number of non-fire point pixels correctly detected by the algorithm, FP is the number of fire point pixels misjudged by the algorithm, FN is the number of non-fire point pixels misjudged by the algorithm, P is the number of fire point pixels in the fire point sample database, F is the number of non-fire point pixels in the non-fire point sample database, and W i is the weight of each band algorithm, Kappa i is the Kappa coefficient of each band algorithm, and n is the total number of all band algorithms.

[0069] The confidence level of the test results is adjusted using the fire probability calculation formula:

[0070]

[0071] Where G is the final result of fire point detection; F i Is the result of a single fire point test condition (if it is a fire point, F i=1, if it is not a fire point, F i =0); W i is the weight of the corresponding fire point test condition; N is the total number of fire test conditions. If the probability of the weighted synthesis result reaches 1, then this pixel is a complete fire pixel. Otherwise, if the probability drops to 0, then this pixel is a complete non-fire pixel.

[0072] The algorithm reduces the false fire point misjudgment rate and improves the detection accuracy of small fire points. Compared with MODIS fire point products, it has lower misclassification errors and omission errors, and lower false detection rates than VIIRS fire point products. Based on MODIS data, a method of using radiation angle normalization is proposed to improve the accuracy of small fire points, and based on the analysis of prior data sets, the fire point information is deeply mined, the appropriate monitoring threshold is set, and the accuracy of fire point monitoring is improved.

[0073] It should be noted that the above description is only some embodiments of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other.

Claims

1. A fire point detection algorithm combining radiation angle normalization with a priori data set, characterized in that: The following steps are involved: (1) The MODIS top atmosphere radiation angle was normalized, and the radiation transfer model was used to simulate the difference ratio between the MODIS fire detection band at different satellite scanning angles and the sub-satellite point radiance, to characterize the impact of angle changes on the MODIS top atmosphere radiation, and the top atmosphere radiance at different band scanning angles was normalized to the radiance observed at the sub-satellite point; (2) Construct a priori dataset, which includes the reflectivity and brightness temperature data of fire point samples and non-fire point samples. After preprocessing the MODIS data by geometric correction, radiation brightness calculation and brightness temperature, the U.S. wildfire geographic data and MODIS data are temporally and spatially aligned to form a dataset that matches the surface and satellite. (3) Select the fire point detection band, taking into account all bands from visible light to thermal infrared, and through band calculation and statistical analysis of fire point and non-fire point sample data, mining the fire information contained in different bands and band combinations, and automatically selecting the fire point detection band; (4) Selecting the fire point detection threshold, based on the prior data set of fire points and non-fire points, using computer automatic analysis technology to automatically compare the radiation characteristics of fire points and non-fire points, and automatically selecting the optimal threshold of the fire point test condition according to the expected accuracy; (5) Fusion of fire detection algorithms: Based on the prior data set, the fire detection results corresponding to the single-band, multi-band and context band algorithms are obtained respectively. The Kappa coefficients of different fire point test conditions are calculated by the accuracy of fire point and non-fire point identification. The respective weights are calculated according to the Kappa coefficients of all test conditions. The weights of each band test condition and the detection results are substituted into the fusion algorithm formula for fire point detection to generate a fire point probability value. The higher the probability value, the more reliable the detection result of the fire point pixel.

2. According to the fire point detection algorithm of a radiation angle normalization combined with a priori data set described in claim 1, the input of the MODTRAN model includes control operation parameters, sensor parameters, atmospheric parameters, observation geometry conditions and surface parameters, and the atmospheric top layer radiation difference ratio r(θ) at the sub-satellite point and different satellite scanning angles is simulated by using the atmospheric profile type known in advance, and the calculated radiation difference ratio is combined with the observation zenith angle data of the MODIS sensor, and the atmospheric top layer radiance observed from the satellite scanning angle θ is normalized to just below the MODIS satellite track: Where: Rad obs (θ) is the top-of-atmosphere radiance received by MODIS at each satellite scanning angle, is the radiance normalized to the area directly below the satellite track.

3. According to the fire point detection algorithm combining radiation angle normalization with a priori data set as described in claim 2, the fire point sample data set should contain fires of different types and sizes and various stages of fire burning, and the fire point information includes the time of fire occurrence, fire extinguishing time, longitude, latitude, fire area size, etc.; the establishment of the non-fire point sample database is based on multi-time series remote sensing images, and the selection of non-fire point samples should not only consider terrain factors, including terrain features such as mountains, plains, and hills, but also various land use / land cover types, including human Four typical surfaces: construction site surface, vegetation, cloud, and water body. Artificial surfaces include cities, roads, industrial areas, etc., which are extracted using the normalized building index. Vegetation mainly includes farmland, forest, grassland, etc., which are selected using the normalized vegetation index. Cloud pixels include thick clouds, thin clouds, broken clouds, etc. Water bodies include oceans, rivers, lakes, etc. Cloud and water pixels are selected using the cloud and water mask threshold in the MODIS fire point algorithm. High-brightness surfaces such as agricultural greenhouses and smooth buildings are selected using a combination of near-infrared, short-wave infrared reflectivity and mid-infrared brightness temperature values.

4. According to the fire point detection algorithm combining radiation angle normalization with a priori data set described in claim 3, the radiation difference between fire points and non-fire points is analyzed based on the sample database, and the bands of 0.86μm, 1.65μm, 2.13μm, 3.8μm, 3.96μm, 11.03μm and 12.02μm of the MODIS sensor wavelength are selected. The single-band, multi-band difference and context band combination method are used, and the bubble method is used to automatically calculate the fire point pixel recognition accuracy to determine the band for fire detection; 5. According to the fire point detection algorithm combining radiation angle normalization and a priori data set described in claim 4, in the process of determining the optimal threshold, the minimum brightness temperature value and the maximum brightness temperature value of the fire point in the fire point a priori data set are counted, and the fire point detection threshold is changed from the minimum brightness temperature value of the fire point a priori data set to the maximum brightness temperature value in steps of K. With the change of the threshold, the fire point pixel recognition accuracy (the fire point pixel accuracy refers to the proportion of correctly recognized fire point pixels to the total number of fire point pixels) and the surface misjudgment rate (the surface misjudgment rate refers to the proportion of misjudged non-fire point pixels to the total number of non-fire point pixels) also change accordingly. When the threshold reaches a certain value, the fire point pixel recognition accuracy increases slowly with the change of the threshold, while the non-fire point pixel misjudgment rate increases significantly with the increase of the threshold. Based on the principle of the bubble method, the computer automatically selects the threshold with the largest accuracy as the optimal threshold under the premise of limiting the non-fire point misjudgment rate a%.

6. According to the fire point detection algorithm combining radiation angle normalization with a priori data set described in claim 5, the calculation formulas of the Kappa coefficient and weight corresponding to each band algorithm are as follows: in, p a is the probability of correct classification, p e is the probability of correct classification when the relative frequency of fire points / non-fire points is known. TP is the number of fire point pixels correctly detected by the algorithm, TN is the number of non-fire point pixels correctly detected by the algorithm, FP is the number of fire point pixels misjudged by the algorithm, FN is the number of non-fire point pixels misjudged by the algorithm, P is the number of fire point pixels in the fire point sample database, F is the number of non-fire point pixels in the non-fire point sample database, and W i is the weight of each band algorithm, Kappa i is the Kappa coefficient of each band algorithm, and n is the total number of all band algorithms.

7. According to the fire point detection algorithm combining radiation angle normalization with a priori data set described in claim 5, the fire point detection algorithm is integrated and the confidence of the detection result is adjusted. The calculation formula of the fire probability is as follows: Where G is the final result of fire point detection; F i Is the result of a single fire point test condition (if it is a fire point, F i =1, if it is not a fire point, F i =0); W i is the weight of the corresponding fire point test condition; N is the total number of fire test conditions. If the probability of the weighted synthesis result reaches 1, then this pixel is a complete fire pixel. Otherwise, if the probability drops to 0, then this pixel is a complete non-fire pixel.

Images