Remote sensing fire point detection sample generation method

Abstract

The invention relates to the technical field of fire point detection, and discloses a remote sensing fire point detection sample generation method, which comprises the following steps of: 1, extracting fire points with medium and high confidence coefficients by taking a fire point product as an initial data source; 2, screening a fire burn area by using a difference normalized combustion index, and extracting burned area information; according to the remote sensing fire point detection sample generation method, real fire points and pseudo fire points are screened in multiple steps, misjudged samples are effectively eliminated, a high-quality fire point detection sample data set is generated, medium and high confidence fire points are used for extraction, the reliability of initial data is ensured, a fire burn area is screened through a difference normalization combustion index, and burned area information is extracted. The authenticity of the fire point is further verified, the real fire point is accurately identified by combining calculation of the area proportion of the open fire area in the fire point pixel, the pseudo fire point is screened based on the spatial-temporal characteristics, and the multi-dimensional screening mechanism ensures the purity and the accuracy of the sample data set.

Description

Technical Field

[0001] The present invention relates to the technical field of fire point detection, and in particular to a method for generating remote sensing fire point detection samples. Background Art

[0002] Fires are sudden, destructive, and difficult to control, posing a serious threat to human life and socio-economic development. The huge smoke and dust produced by combustion not only pollutes the air and water, but also exacerbates the global greenhouse effect, causing a huge impact on the ecological environment and bringing immeasurable economic losses. Under the general trend of global warming, the frequency of global fires is on the rise, making it particularly important to detect and locate fires in a timely and accurate manner.

[0003] Since its application in the 1970s, satellite remote sensing technology has been widely used in fire detection due to its comprehensiveness, high timeliness, repeatability, and low cost. Many researchers and scholars at home and abroad have proposed remote sensing fire detection algorithms based on physical mechanisms, achieving the classification and identification of fire data. Currently, the more mature traditional algorithms can be summarized as threshold methods, context methods, and multi-temporal methods. These algorithms, supported by theories such as Planck's radiation law and the Stefan-Boltzmann law, extract fire location information by setting thresholds for remote sensing image pixel brightness, reflectivity, brightness temperature, and other indicators. In recent years, with the rapid development of artificial intelligence and cloud computing technologies, remote sensing fire detection technology has further room for in-depth research and application prospects. Using techniques such as machine learning and deep learning, fire data can be classified and identified more accurately, rapidly, and comprehensively, improving the accuracy and automation of fire detection technology, and achieving major breakthroughs and improvements in development direction and application expansion. Compared with traditional fire detection algorithms based on physical mechanisms, deep learning models have demonstrated higher detection accuracy and robustness.

[0004] Although various algorithms have made great progress in remote sensing fire detection, the following problems still exist: Various fire point products have shown that artificial heat sources on the urban surface, such as solar reflectors, industrial mining areas, and self-built roofs made of special materials, will interfere with remote sensing fire point detection, resulting in a large number of "pseudo-fire points" in the detection process. The spatial differences of fixed heat source points in a single scene may show similar characteristics to fire points, but their characteristics are long-term high temperature and no temporal changes in spatial differences, which can be used to eliminate false positives.

[0005] In addition, for deep learning remote sensing fire point detection, the importance of sample datasets is self-evident. It directly affects the performance, practicality and test results of the model. However, the currently publicly available fire point datasets for deep learning training and testing are generally insufficient. There are also a large number of pseudo fire points in the officially released fire products, which cannot be directly used as fire point samples for model training. Summary of the Invention

[0006] In view of the fact that the above-mentioned existing publicly available fire point data sets for deep learning training and testing are generally insufficient, and there are also a large number of pseudo fire points in the officially released fire products, which cannot be directly used as fire point samples for model training, the present invention is proposed.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a remote sensing fire point detection sample generation method, comprising the steps of: first, extracting medium and high confidence fire points using fire point products as an initial data source;

[0008] Step 2: Use the difference normalized burn index (dNBR) to screen the fire burn area and extract the fire scar information;

[0009] Step 3: Screen the real fire points by the ratio of the open fire area in the fire point pixel;

[0010] Step 4: Filter out false fire points based on spatiotemporal characteristics. Multi-source remote sensing data can provide information with different spatiotemporal resolutions.

[0011] Step 5: Construct other non-fire point datasets, including cloud, water, vegetation, building land, and bright surface samples;

[0012] Step 6: Generate a sample data set containing real fire points, pseudo fire points and other non-fire points.

[0013] As a preferred embodiment of the remote sensing fire point detection sample generation method of the present invention, the difference normalized burn index (dNBR) in step 2 is calculated by the following formula:

[0014] dNBR=NBR before fire - NBR after fire

[0015] The calculation formula of the normalized combustion index is:

[0016]

[0017] NIR is the near-infrared band reflectivity, and SWIR is the short-wave infrared band reflectivity.

[0018] As a preferred solution of the remote sensing fire point detection sample generation method of the present invention, the area ratio of the open fire area within the fire point pixel in step 3 is calculated by the following formula:

[0019]

[0020] R iMix =P·R iF +(1-P)·R iB

[0021] Where ΔT is the brightness temperature increment of the pixel containing the open fire, T iMix is the brightness temperature of the pixel containing bright fire, T iB is the background pixel brightness temperature, h is Planck's constant, c is the speed of light in vacuum, k is the Boltzmann constant, λ i is the wavelength of band i, R iMix is the radiance of the pixel containing the bright fire, R iF is the radiance of the open fire area, R iB is the radiance of the non-open fire area, and P is the area ratio of the open fire area.

[0022] As a preferred solution of the remote sensing fire point detection sample generation method of the present invention, the false fire point screening in step 4 is based on spatiotemporal characteristics and is achieved by the following steps:

[0023] S1, establish a buffer zone with the latitude and longitude of the candidate pseudo-fire point as the center;

[0024] S2, extracting areas where the buffer overlap frequency is greater than a preset threshold within a certain time range;

[0025] S3, record all points in the area as pseudo-fire points.

[0026] As a preferred solution of the remote sensing fire point detection sample generation method of the present invention, the other non-fire point data sets in step 5 are constructed by the following steps:

[0027] S1, extract cloud pixels using cloud detection threshold;

[0028] S2, extracting water pixels using spectral analysis;

[0029] S3, extracting vegetation pixels using the Normalized Difference Vegetation Index (NDVI);

[0030] S4, extracting building land pixels using the normalized difference building index (NDBI);

[0031] S5, extracting bright surface pixels using a multi-channel thresholding method;

[0032] S6, expand non-fire point samples by random sampling.

[0033] As a preferred solution of the remote sensing fire point detection sample generation method described in the present invention, the fire point product is a VNP14IMG fire product with a spatial resolution of 375m, which contains the latitude and longitude, confidence and pixel type information of the fire point pixel.

[0034] As a preferred solution of the remote sensing fire point detection sample generation method described in the present invention, the multi-source remote sensing data includes VIIRS data and Sentinel-2 data, wherein the VIIRS data is used to extract fire point information, and the Sentinel-2 data is used to calculate dNBR and extract burnt area information.

[0035] Compared with the prior art, the present invention has at least the following beneficial effects:

[0036] 1. The present invention uses multiple steps to screen real fire points and false fire points, effectively eliminates misjudged samples, and generates a high-quality fire point detection sample data set. First, medium and high confidence fire points are extracted to ensure the reliability of the initial data. Secondly, the difference normalized combustion index (dNBR) is used to screen the fire burn area, extract the fire scar information, and further verify the authenticity of the fire point. Finally, combined with the calculation of the area ratio of the open flame area within the fire point pixel, the real fire point is accurately identified. This multi-dimensional screening mechanism ensures the purity and accuracy of the sample data set, provides a reliable training basis for the deep learning model, and thus significantly improves the accuracy of fire point detection.

[0037] 2. By incorporating pseudo-fire points and other non-fire point samples into the data set, the present invention helps the algorithm better learn the difference between fire points and non-fire points, significantly improving the robustness and generalization ability of the model. Pseudo-fire point screening is based on spatiotemporal characteristics, which can identify fixed heat source points and avoid misjudgment. At the same time, other non-fire point data sets constructed cover a variety of surface types such as clouds, water bodies, vegetation, building land and bright surfaces, ensuring the diversity of samples. This diversity and comprehensiveness enable the model to adapt to various complex fire scenarios, reduce the misjudgment rate, and improve the reliability of fire monitoring.

[0038] 3. The present invention combines multi-source remote sensing data, such as VIIRS and Sentinel-2, to fully leverage the advantages of both. VIIRS data is used to extract fire point information and has high sensitivity and a spatial resolution of 375m, enabling the detection of smaller-scale fires. Sentinel-2 data is used to calculate dNBR and extract burnt area information. It has high spatial resolution and rich spectral bands, and can provide detailed surface feature information. This fusion of multi-source data makes the method applicable to different surface types and fire scenarios, adapting to complex and changing environmental conditions, meeting the fire monitoring needs of different regions, and ensuring the stability and reliability of sample data.

[0039] 4. The algorithm-based automated processing flow of the present invention can efficiently generate sample data sets and significantly reduce the cost of manual intervention. Through the calculation of the difference normalized combustion index, the screening of the area ratio of the open fire area, the screening of the spatiotemporal characteristics of pseudo-fire points, and the multi-method construction of non-fire point samples, the entire process is highly automated. This automated processing not only improves the efficiency of sample generation, but also ensures the consistency and accuracy of the sample data. The high-quality sample data set provides a reliable training basis for the deep learning model, further improves the efficiency of fire monitoring, and provides strong support for the timely detection of fires and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0041] Figure 1 This is the overall technical roadmap of the remote sensing fire point detection sample generation method of the present invention.

[0042] Figure 2 This is a brightness temperature increment diagram caused by different open fire area ratios in the remote sensing fire point detection sample generation method of the present invention.

[0043] Figure 3 This is the real fire point map screened out by the remote sensing fire point detection sample generation method of the present invention (the yellow rectangle is the fire point pixel, and the red area is the burned area).

[0044] Figure 4 Schematic diagram of pseudo-fire point screening for the remote sensing fire point detection sample generation method of the present invention (red dots represent fire point distribution, blue rectangles represent fire point buffer zones, and red gradient portions represent overlapping areas; darker colors indicate higher overlap frequencies).

[0045] Figure 5 This is a pseudo-fire point map of industrial heat sources for the remote sensing fire point detection sample generation method of the present invention (a is a true color image acquired by the Sentinel-2 satellite on April 29, 2021, b is a true color image acquired by the Sentinel-2 satellite on May 9, 2021, and c is a true color image acquired by the Sentinel-2 satellite on May 24, 2021. The yellow rectangles are pseudo-fire point pixels). DETAILED DESCRIPTION

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0048] Reference Figure 1-5 , as an embodiment of the present invention, provides a remote sensing fire point detection sample generation method, the remote sensing fire point detection sample generation method comprises the steps of: first, using the fire point product as the initial data source, extracting medium and high confidence fire points;

[0049] Step 2: Use the difference normalized burn index (dNBR) to screen the fire burn area and extract the fire scar information;

[0050] Step 3: Screen the real fire points by the ratio of the open fire area in the fire point pixel;

[0051] Step 4: Filter out false fire points based on spatiotemporal characteristics. Multi-source remote sensing data can provide information with different spatiotemporal resolutions.

[0052] Step 5: Construct other non-fire point datasets, including cloud, water, vegetation, building land, and bright surface samples;

[0053] Step 6: Generate a sample data set containing real fire points, pseudo fire points and other non-fire points.

[0054] 1. Introduction to VIIRS Data

[0055] The Visible Infrared Imaging Radiometer (VIIRS) onboard the Suomi-NPP satellite collects data using 22 channels in the 0.4-12 range, including five high-resolution image channels (I1-I5 bands, with a resolution of 375m), 16 medium-resolution image channels (M1-M16 bands, with a resolution of 750m), and a panchromatic day-night observation channel (DNB, with a resolution of 750m). Specific information for the 375m resolution I band is shown in the table below. As the next generation of MODIS sensors, VIIRS expands and improves on the basis of inheriting the advantages of MODIS. It has extremely wide applications in the fields of environment, ecology, and meteorology. It can measure atmospheric cloud cover, invert aerosol concentration, land and ocean surface temperature, and fire point conditions. In terms of fire point monitoring applications, VIIRS has increased the spatial resolution of the fire point band data to 375m, which is theoretically more sensitive to small-area and small-amplitude thermal anomaly changes; the 375m resolution I band table is as follows:

[0056] Table 1VII RSI band data table

[0057]

[0058] 2. Fire point product introduction.

[0059] To quickly obtain a sufficient number of wildfire point data, this study used the VNP14IMG fire product as the initial fire point data source;

[0060] VNP14IMG is collected by the Visible Infrared Imaging Radiometer (VIIRS) sensor on the Suomi-NPP satellite and processed using a specialized fire detection algorithm to provide location and intensity information for thermal anomaly data at a spatial resolution of 375 meters. This product, based on an improved detection algorithm for MODIS thermal anomaly data products, is more sensitive to smaller fires and better maps the perimeters of large fires. Its data primarily includes information such as the latitude and longitude of fire point pixels, confidence levels, and pixel type. A detailed description of these data is provided in the table below:

[0061] Table 2 VNP14IMG data sheet

[0062]

[0063] 3. Introduction to Sentinel-2L2A data.

[0064] Sentinel-2 is an Earth observation mission under the European Copernicus program, a collaboration between the European Commission (EC) and the European Space Agency (ESA). It provides a variety of services related to land management and applications to European and national organizations. The Sentinel-2 mission consists of two polar-orbiting satellites, Sentinel-2A (launched on June 23, 2015) and Sentinel-2B (launched on March 7, 2017). They operate in the same sun-synchronous orbit at an average altitude of 786 km, 180° out of phase with each other, with a swath of 290 km and a complementary revisit period of five days. Sentinel-2 is a high-resolution multispectral imaging satellite equipped with an advanced multispectral instrument (MSI), covering 13 spectral bands at three spatial resolutions of 10 m, 20 m, and 60 m. It includes data from three bands in the red edge range, making it highly effective for monitoring vegetation health. The detailed introduction of relevant data is shown in the table below:

[0065] Table 3 Sentinel-2 data table

[0066]

[0067] Sample dataset construction

[0068] 1. Construction of real fire point dataset.

[0069] This study uses fire point products as the initial fire point data source, and adopts three steps to screen the real fire points in the product: extracting medium and high confidence fire points, screening the difference normalized combustion index, and screening the area ratio of the open fire area of the fire point pixel.

[0070] 1.1. Medium and high confidence fire point extraction

[0071] The VNP14IMG fire point product has a spatial resolution of 375m, which has a higher detection capability for relatively small fires. It provides fire point location information and divides fire confidence into three levels: low, medium, and high. To ensure the accuracy of fire points, this study only selected fire points with medium and high confidence to further extract true fire points.

[0072] 1.2. Difference Normalized Combustion Index Screening

[0073] Fires are often accompanied by significant temperature anomalies, and the physical properties of vegetation before and after a fire can also change dramatically. Traditional fire spot screening methods rely primarily on manual visual interpretation, judging the authenticity of fire spots by observing burn marks in images before and after the fire. This method is inefficient and subject to subjective factors. Using remote sensing data to extract burn marks from vegetation before and after combustion to screen fire spots has significant advantages, improving the objectivity and reliability of fire spot identification. This study used high-spatial-resolution Sentinel-2 imagery to calculate the Normalized Burn Ratio (NBR) to assess the burned area and burn severity. The NBR index can effectively enhance the characteristic information of the fire area and can be used to assess the condition of vegetation in the burned area. Its calculation formula is as follows:

[0074]

[0075] Among them, represents the reflectance in the near-infrared band and represents the reflectance in the short-wave infrared band. In this study, they correspond to bands B8 and B12 of the Sentinel-2L2A data, respectively. Healthy vegetation generally has high reflectance in the near-infrared (NIR) band and low reflectance in the short-wave infrared (SWIR) band. Therefore, in the absence of fire, the NBR value is generally high. When vegetation is burned by fire, the reflectance in the NIR band will decrease significantly, while the reflectance in the SWIR band will increase significantly due to the loss of vegetation water and the production of charcoal and ash, resulting in a decrease in the NBR value.

[0076] By comparing the changes in NBR values before and after a fire, that is, calculating the difference normalized burn index (dNBR), the burn area can be effectively identified and the burn scar information can be extracted. The calculation formula is as follows:

[0077] dNBR=NBR pre-fire -NBR post-fire

[0078] Where represents the NBR value before the fire, and represents the NBR value after the fire. A larger dNBR value indicates a more severe combustion. Based on experience, a dNBR threshold range of 0.2-1.2 achieves higher accuracy.

[0079] This study used the Google Earth Engine (GEE) platform to further screen fires. The time, latitude, and longitude coordinates of the medium- and high-confidence fire data extracted in the previous step were uploaded to the GEE platform. Based on the time and location of the fires, two cloud-free Sentinel-2 images within 15 days before and after the fires were searched for the closest fire times to calculate the dNBR. The Sentinel-2 L2A scene classification map was used to mask water bodies, clouds, and snow. The spatial resolution of the B12 band was upsampled to 10 meters to allow direct calculations with the B8 band, generating a dNBR image with a 10-meter spatial resolution. Based on the uploaded fire location information, a 375-by-375-meter buffer zone was established, centered on the longitude and latitude of each fire. When the buffer zone fell within the burned area, the fire was marked as a candidate true fire.

[0080] 1.3. Screening of the proportion of open fire area by fire point image

[0081] To more accurately screen fire spots, this study considered the proportion of burnt areas within the fire pixel. Because there are significant differences in temperature and radiation characteristics between an open fire and its surroundings, it can be assumed that the radiance of a fire pixel is a linear combination of the radiances of the open fire area and the non-open fire area. The calculation formula is as follows:

[0082] R iMix =P·R iF +(1-P)·R iB

[0083] Among them, is the radiance of the i-band of the pixel containing open fire, is the radiance of the i-band of the open fire area, is the radiance of the i-band of the non-open fire area, and is the ratio of the area of the open fire area in the pixel to the area of the pixel.

[0084] Based on Planck's law, the brightness temperature increment of a pixel containing an open fire can be expressed as:

[0085]

[0086] Wherein, is the brightness temperature increment of the i-band pixel containing bright fire, is the brightness temperature of the i-band pixel containing bright fire, is the brightness temperature of the i-band pixel containing background, h is Planck's constant, h = 6.626 × 10^(-34) J·s, c is the speed of light in vacuum, k is the Boltzmann constant, k = 1.38 × 10^(-23) J / K, λ i is the wavelength of the i band.

[0087] For VIIRS images, the wavelength of the I4 mid-infrared band at 375m resolution is 3.74, which can sensitively monitor surface thermal anomalies. Assuming that the brightness temperature of the surface background pixel in this band is 295K, during the burning of biomass such as forests or grasslands, the flame brightness temperature is usually between 600K and 1000K. The larger the proportion of the open fire area, the more significant the brightness temperature difference caused. Figure 2 The figure shows the brightness temperature increment caused by different percentages of open fire area in the I4 mid-infrared band. When the brightness temperature of the open fire area is 600K and the area ratio is 0.5%, the brightness temperature increment is approximately 40K, indicating a significant thermal anomaly. Therefore, after extracting the burn scars, this study selected fire points with a burned area ratio greater than 0.5% within the fire point pixel as the final true fire point.

[0088] 1.4 Introduction to Fire Point Screening

[0089] The three steps above can be used to screen out the real fire points in the fire point product. Using the MODIS land cover type product (MCD12Q1), the types provided by the International Geosphere-Biosphere Program (IGBP) can be used to classify fire points into agricultural fires, forest fires, and grassland fires. Figure 3As shown, a, b, and c are selected farmland fires. The fire coordinates are 34.38357°N, 107.6972733°E, and the fire time is 05:56 on June 29, 2022. a is a true color image acquired by the Sentinel-2 satellite on June 15, 2022, and b is a true color image acquired by the Sentinel-2 satellite on June 30, 2022. d, e, and f are selected forest fires. The fire coordinates are 25.13401°N, 109.9468231°E, and the fire time is 17:55 on April 6, 2022. d is a true color image acquired by the Sentinel-2 satellite on April 5, 2022, and e is a true color image acquired by the Sentinel-2 satellite on April 7, 2022. Figures g, h, and i represent selected grassland fires. The fire coordinates are 40.95688°N, 114.6431732°E, and the fire time was 04:55 on April 24, 2022. g is a true-color image acquired by the Sentinel-2 satellite on April 12, 2022, and h is a true-color image acquired by the Sentinel-2 satellite on April 25, 2022. The yellow rectangles in c, f, and i represent fire pixels, and the red areas are burn scars extracted using the difference normalized burn index. Comparing the images before and after the fire reveals clear burn scars, which dNBR can effectively extract. A fire is considered a true fire if the burned area accounts for more than 0.5% of the fire pixel area.

[0090] 2. Construction of pseudo-fire point dataset

[0091] Traditional satellite fire monitoring algorithms primarily identify fires based on their abnormally high temperatures and brightness temperature differences compared to surrounding pixels. Fixed surface heat sources often exhibit radiation signatures similar to those of fires, leading to misidentification and the creation of numerous "pseudo-fires" (false fires). These include industrial heat sources, custom-built rooftops made of special materials, and bright surface reflections. By constructing false fire samples and incorporating them into non-fire samples, the algorithm can better learn to distinguish between fires and non-fires, thereby enhancing the model's generalization capabilities.

[0092] Fixed heat source has the characteristics of position invariance and long-term high temperature, but biomass burning is concentrated and extinguished in a short time, and multiple fires will not occur in the same place. Figure 4 As shown in Figure 2, this study uses fire point products as the initial data source based on spatiotemporal characteristics. After excluding confirmed real fire points according to the above method, pseudo fire points are extracted from the remaining fire point products. A 375m×375m buffer is established with the latitude and longitude of each candidate pseudo fire point as the center. Areas with a buffer overlap frequency of more than five points within a certain time range (e.g., 30 days) are extracted, and all points within these areas are recorded as pseudo fire points.

[0093] This method can effectively screen out false fire points in VIIRS fire point products, such as Figure 5 As shown, the coordinates of this industrial heat source pseudo-fire point are 35.35124588°N, 115.9872971°E. During May 2021, this location was misidentified as a fire point by the VIIRS fire point product for a total of 17 days.

[0094] 3. Construction of other non-fire point datasets

[0095] To ensure that non-fire samples contain information that effectively distinguishes burned areas from other non-burned areas, the samples must be diverse and cover a wide range of surface types. In addition to using pseudo-fires as non-fire samples, clouds, water bodies, vegetation, built-up areas, and bright surfaces must also be included.

[0096] (1) Cloud pixel selection

[0097] The cloud detection threshold in the VIIRS fire point algorithm has been verified to be able to identify large and low-temperature clouds. The specific formula is as follows:

[0098] BT5<265K||ρ1+ρ2>0.9&&BT5<295K

[0099] ||ρ1+ρ2>0.7&&BT5<285K

[0100] (2) Water pixel selection

[0101] The water pixels are extracted using a spectral analysis method based on the reflectance of the VIRSI1, I2, and I3 bands. The specific formula is as follows:

[0102] ρ1>ρ2>ρ3

[0103] (3) Vegetation pixel selection

[0104] The Normalized Difference Vegetation Index (NDVI) is widely used to assess vegetation cover and vegetation health. It reflects vegetation growth by calculating the difference in reflectance between red and near-infrared bands. The specific formula is as follows:

[0105]

[0106] (4) Selection of building land pixels

[0107] The Normalized Difference Building Index (NDBI) is a remote sensing index used to detect and quantify urban built-up areas or artificial surfaces. Positive values indicate a high probability of built-up or urbanized areas, while negative values indicate the possibility of vegetation or bare soil. The specific formula is as follows:

[0108]

[0109] (5) Selection of bright surface pixels

[0110] Bright surface pixels are extracted using a multi-channel threshold method. The reflection of the bright surface will cause high brightness temperature values in the daytime data, which will be confused with fire point pixels. The specific formula is as follows:

[0111] BT4<335K||ρ1+ρ2>0.6&&BT5<285K||ρ3>0.3

[0112] &&ρ3>ρ2&&ρ2>0.25&&BT4<335K

[0113] (6) Random sampling

[0114] In order to further improve the non-fire point dataset, this study also uses random sampling to expand the non-fire point samples. Random sampling is performed in remote sensing images, but potential fire points are also eliminated at the same time. The specific formula is as follows:

[0115] BT4<325K&&BT 45 <25K

[0116] Among them, ρ1, ρ2, ρ3 are the reflectances of I1, I2, and I3 bands, BT4 and BT5 are the brightness temperatures of I4 and I5 bands, and BT 45 is the brightness temperature difference between I4 and I5 bands.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for generating remote sensing fire point detection samples, characterized in that: The following steps are involved: Step 1: Using the fire point product as the initial data source, extract the medium and high confidence fire points; Step 2: Use the difference normalized combustion index to screen the fire burn area and extract the fire scar information; Step 3: Screen the real fire points by the ratio of the open fire area in the fire point pixel; Step 4: Filter out false fire points based on spatiotemporal characteristics. Multi-source remote sensing data can provide information with different spatiotemporal resolutions. Step 5: Construct other non-fire point datasets, including cloud, water, vegetation, building land, and bright surface samples; Step 6: Generate a sample data set containing real fire points, pseudo fire points and other non-fire points.

2. The remote sensing fire point detection sample generation method according to claim 1, characterized in that: The difference normalized combustion index in step 2 is calculated using the following formula: dNBR=NBR before fire - NBR after fire The calculation formula of the normalized combustion index is: NIR is the near-infrared band reflectivity, and SWIR is the short-wave infrared band reflectivity.

3. The remote sensing fire point detection sample generation method according to claim 1, characterized in that: The area ratio of the open fire area within the fire point pixel in step 3 is calculated by the following formula: P iMix =P·R iF +(1-P)·R iB Where ΔT is the brightness temperature increment of the pixel containing the open fire, T iMix is the brightness temperature of the pixel containing bright fire, T iB is the background pixel brightness temperature, h is Planck's constant, c is the speed of light in vacuum, k is the Boltzmann constant, λ i is the wavelength of band i, R iMix is the radiance of the pixel containing the bright fire, R iF is the radiance of the open fire area, R iB is the radiance of the non-open fire area, and P is the area ratio of the open fire area.

4. The remote sensing fire point detection sample generation method according to claim 1, characterized in that: The false fire point screening in step 4 is based on spatiotemporal characteristics and is achieved through the following steps: S1, establish a buffer zone with the latitude and longitude of the candidate pseudo-fire point as the center; S2, extracting areas where the buffer overlap frequency is greater than a preset threshold within a certain time range; S3, record all points in the area as pseudo-fire points.

5. The remote sensing fire point detection sample generation method according to claim 1, characterized in that: The other non-fire point datasets in step 5 are constructed by the following steps: S1, extract cloud pixels using cloud detection threshold; S2, extracting water pixels using spectral analysis; S3, extracting vegetation pixels using the normalized vegetation index; S4, extracting building land pixels using normalized building index; S5, extracting bright surface pixels using a multi-channel thresholding method; S6, expand non-fire point samples by random sampling.

6. The remote sensing fire point detection sample generation method according to claim 1, characterized in that: The fire point product is the VNP14IMG fire product, which has a spatial resolution of 375m and contains the latitude and longitude, confidence and pixel type information of the fire point pixel.

7. The remote sensing fire point detection sample generation method according to claim 6, characterized in that: The multi-source remote sensing data includes VIIRS data and Sentinel-2 data, wherein VIIRS data is used to extract fire point information, and Sentinel-2 data is used to calculate dNBR and extract burnt area information.

Images