Wild fire monitoring method based on multi-source time sequence remote sensing image

Abstract

The invention discloses a wildfire monitoring method based on a multi-source time sequence remote sensing image, and relates to the technical field of remote sensing geoscience application, and the method comprises the steps: obtaining multi-source remote sensing data, constructing a time sequence waveband data set, and constructing a high-temperature thermal anomaly index through the atmospheric top reflectivity of a three-waveband spectrum; based on the high-temperature thermal anomaly index, extracting a first-level fire point pixel, a saturated pixel and a high-radiance pixel from the multi-source remote sensing data, and converting the saturated pixel and the high-radiance pixel into the first-level fire point pixel; extracting saturated pixels and high-radiance pixels from missed fire cluster cores in the multi-source remote sensing data; based on limiting conditions, building coverage area data and a high-frequency ground object mask, noise pixels in the first-level fire point pixels are removed, and second-level fire point pixels are obtained; transmitting to a local client and a cloud geographic space analysis platform, and creating a fire point list. Multi-source long-time sequence optical remote sensing data are combined, and the spatial and temporal distribution of the large-area wildfire is quickly and automatically identified.

Description

Technical Field

[0001] This invention relates to the field of remote sensing geosciences application technology, and more specifically, to a wildfire monitoring method based on multi-source time-series remote sensing images. Background Technology

[0002] Wildfires, a common biomass burning phenomenon in vegetated ecosystems, are experiencing a significant increase in frequency and intensity driven by global warming. Extreme weather events such as heat waves and droughts led to a 42% increase in wildfire burned area in the Arctic region from 2010 to 2020 compared to the previous decade, and in 2020, wildfires in Siberia released 1.5% of global anthropogenic carbon dioxide emissions. While traditional low-resolution remote sensing possesses high-frequency revisit capabilities, it suffers from a 43.41% omission rate for small fires smaller than a pixel, resulting in carbon emission estimation errors exceeding 30%. Accurate monitoring of the spatiotemporal distribution of wildfires is crucial for assessing ecological impacts and optimizing prevention and control strategies.

[0003] In medium- and high-resolution remote sensing images, the emitted radiation of wildfires in the near-infrared to short-wave infrared wavelength range increases exponentially with increasing wavelength. Among them, the reflectance of open flames and smoldering flames differs significantly in the far-short-wave infrared, near-short-wave infrared, and near-infrared bands. The intensity of thermal anomalies can be quantified by constructing the High Temperature Thermal Anomaly Index (TAI). By combining the band reflectance difference with spatial neighborhood analysis, wildfire pixels can be extracted and high-radiation / saturated pixels in the center of the fire cluster can be added to avoid missing "holes".

[0004] However, wildfire detection using large-scale, long-term imagery faces complex noise interference, including high-frequency ground features (such as building roofs and industrial heat sources), bare soil, snow and ice, cloud edges, and image data quality defects. Accuracy needs to be improved through spectral constraints and quality control. Simultaneously, processing massive amounts of high-resolution image data requires a hybrid "local + cloud" computing framework, combining parallel computing on the GEE platform with local batch processing to generate a high-precision wildfire inventory.

[0005] Currently, research on constructing large-scale wildfire inventories based on Sentinel-2 MSI and Landsat-8 OLI time-series imagery is relatively rare. This is mainly due to limitations in sensor bands (both Sentinel-2 and Landsat-8 lack mid- and late-wave bands sensitive to high-temperature fire sources, and Sentinel-2 lacks a thermal infrared band, so only the less sensitive short-wave infrared and near-infrared bands can be used for wildfire detection), the mixing of daytime radiation (daytime short-wave infrared radiation is composed of a mixture of solar radiation reflected from background objects and active radiation from high-temperature thermal anomalies), data update cycles and incomplete coverage (unlike low spatial resolution, the single-track coverage of Sentinel-2 and Landsat-8 is limited and cannot achieve complete daily coverage of the region, because wildfire ignition points are instantaneous phenomena, resulting in the omission of wildfires outside the track during transit; and cloud cover also affects the uncertainty of wildfire combustion detection), the complex underlying surface of medium- and high spatial resolution images, and the large amount of data brought about by medium- and high-resolution imagery.

[0006] Therefore, monitoring wildfires in large-scale, medium-to-high spatial resolution remote sensing images and accurately establishing time-series wildfire inventories remains a cutting-edge challenge that urgently needs to be overcome in the international remote sensing field. Summary of the Invention

[0007] In response to the problems in related technologies, this invention proposes a wildfire monitoring method based on multi-source time-series remote sensing images to overcome the aforementioned technical problems existing in the existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows: This invention provides a wildfire monitoring method based on multi-source time-series remote sensing imagery, comprising: S1. Acquire multi-source remote sensing data and construct a time-series band dataset, and use the atmospheric top reflectance of the three-band spectrum to construct a high-temperature thermal anomaly index. S2. Based on the high-temperature thermal anomaly index, extract first-level fire point pixels, saturated pixels and high emissivity pixels from multi-source remote sensing data, and convert saturated pixels and high emissivity pixels into first-level fire point pixels. S3. Extract saturated pixels and high emissivity pixels from the missing fire cluster cores in multi-source remote sensing data, and supplement the missing primary fire point pixels in the high-temperature combustion center of the fire point. S4. Based on pre-configured constraints, building coverage area data, and high-frequency ground cover masks, noise pixels are removed from the primary fire pixel to obtain the secondary fire pixel. S5. Transmit the secondary fire point pixels to the local client and the cloud geospatial analysis platform, and create a fire point list.

[0009] Furthermore, the multi-source remote sensing data includes optical images from land imagers and optical images from multispectral imagers; The three-band spectrum includes far-shortwave infrared, near-shortwave infrared, and near-infrared bands; The high-temperature thermal anomaly index refers to the ratio of the reflectance difference between the far-shortwave infrared band and the near-shortwave infrared band to the reflectance of the near-infrared band.

[0010] Furthermore, based on high-temperature thermal anomaly indicators, primary fire point pixels, saturated pixels, and high emissivity pixels are extracted from multi-source remote sensing data, and the saturated pixels and high emissivity pixels are converted into primary fire point pixels, including: S21. Based on the high temperature thermal anomaly index, extract first-level fire point pixels that are greater than or equal to the set threshold, and saturated pixels and high emissivity pixels that are less than the set threshold from the optical images of the land imager and the optical images of the multispectral imager. S22. Convert saturated pixels and high emissivity pixels into first-order fire pixels.

[0011] Furthermore, for the missing fire cluster cores in multi-source remote sensing data, saturated pixels and high emissivity pixels are extracted, including: S31. Based on the saturated pixels of the first-level preset band, supplement the missing fire cluster cores; S32. Based on the high emissivity pixels of the secondary preset band, supplement the missing fire cluster cores.

[0012] Furthermore, based on the saturated pixels of the first-level preset band, the missing fire cluster cores are supplemented, including: S311. Based on saturated pixels with a value greater than 0.4 in the far-shortwave infrared band and greater than 1 in the near-shortwave infrared band, supplement the missing fire cluster cores.

[0013] Furthermore, based on the high emissivity pixels of the secondary preset band, the missing fire cluster cores are supplemented, and the primary fire point pixels missing in the high-temperature combustion center of the fire point include: S321. Based on high emissivity pixels where the ratio of near-shortwave infrared band to far-shortwave infrared band is greater than 2, and the ratio of near-shortwave infrared band to far-shortwave infrared band is greater than 1, missing fire cluster cores are supplemented.

[0014] Furthermore, removing noisy pixels from primary fire pixels based on pre-configured constraints includes: S41. Add a first-level constraint to remove edge anomalies caused by differences in the imaging range of different band data in the image boundary area. S42. Add secondary restrictions to remove data packet loss that leads to missing or damaged pixels in the image; S43. Add three-level restrictions to eliminate false alarms at cloud edges that meet the criteria of near-shortwave infrared band being smaller than the coastal / aerosol band and near-infrared band. S44. Add a fourth-level constraint to remove bare soil noise that meets the condition that the far-shortwave infrared band is greater than the near-shortwave infrared band and the far-shortwave infrared band is greater than 0.2. S45. Add five levels of restrictions to remove ice and snow noise that meets the conditions that the far-shortwave infrared band is smaller than the near-shortwave infrared band and the near-infrared band is less than 0.65.

[0015] Furthermore, based on pre-configured building coverage area data and high-frequency ground feature masks, the formula for removing noisy pixels from primary fire point pixels is as follows: ; In the formula, Mask HFO Indicates a high-frequency ground feature mask; F ( x,y,t i ) indicates in i Fire pixels were detected on the date; n Indicates within the valid time range n Number of fire pixels detected; Object j Represents a planar vector object generated using a point statistical raster. j .

[0016] Furthermore, the secondary fire point pixels are transmitted to the local client and the cloud-based geospatial analysis platform, including: S51. The secondary fire point pixels are transmitted to the local client. The local client uses standardized data as an index to obtain image number information from the cloud geospatial analysis platform and transmits it to the local image task list. S52. Based on standardized data generated by the local client, the data is gradually calculated through the cloud geospatial analysis platform, and the fire point pixels are stored in the binary image. Combined with an automated processing mechanism, they are batch-converted into potential fire cluster objects in a four-neighbor connection manner. S53. Use the raster-to-polygon tool in geographic information software to convert the binarized image into a vector polygon. S54. Use geometric calculation tools to calculate the centroid of each vector polygon to obtain the daily fire point dataset.

[0017] Furthermore, creating a fire point list also includes: S55. Based on vector polygons and combined with daily fire point datasets, define the fire point burning area; S56. Using the field calculation tool in the geographic information software, count the number of pixels, detection date, latitude and longitude coordinates, and combustion type information within the combustion area of ​​each fire point.

[0018] The beneficial effects of this invention are as follows: 1) This invention combines massive amounts of multi-source long-term optical remote sensing data. The high spatial resolution enhances the robustness of remote sensing image detection, greatly improves the ability to monitor wildfires, and can quickly and automatically identify the spatiotemporal distribution of wildfires in large areas.

[0019] 2) The implementation steps of this invention are simple and easy to implement. By using a high-temperature thermal anomaly index based on three spectra and a hybrid calculation framework, it has a good effect on extracting wildfire hotspots in the Arctic and even globally.

[0020] 3) This invention uses long-term, large-spatial-coverage optical image data and an optimized and concise thermal anomaly index, combined with context analysis and false alarm suppression methods, to achieve accurate extraction of wildfires on a large spatial scale and effectively supplement the omissions of low spatial resolution fire point products.

[0021] 4) This invention helps to compile a detailed, timely and effective updated high-resolution wildfire inventory, and can provide scientific reference for global wildfire dynamic management, combustion pattern analysis and assessment of the potential impact of wildfires on the atmospheric environment and ecosystems. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is one of the flowcharts of a wildfire monitoring method based on multi-source time-series remote sensing images according to an embodiment of the present invention; Figure 2 This is a second flowchart of a wildfire monitoring method based on multi-source time-series remote sensing images according to an embodiment of the present invention; Figure 3 These are the blackbody radiation law and the characteristic curve of the wildfire spectrum; Figure 4 The detection performance of different IFOVs on a 100m² thermal anomaly target; Figure 5 This is a scatter plot of land cover types and wildfire burning sampling points in the Arctic region, and forest fire and shrub fire density on Sentinel-2 MSI ("blue-yellow" indicates "low-high" density). Figure 6This is a TAI scatter plot of land cover types and wildfire burning sampling points in the Arctic region, and forest and brush fire density on Landsat-8 OLI ("blue-yellow" indicates "low-high" density). Figure 7 This is a high radiation and saturation phenomenon in a typical wildfire combustion scenario of Sentinel-2 MSI; Figure 8 This is a high radiation and saturation phenomenon in a typical wildfire burning scenario in Landsat-8 OLI; Figure 9 The results are based on wildfire detection algorithms using MSI and OLI imagery; Figure 10 It's noise from bare soil and snow; Figure 11 It's a false alarm from MSI Cloud Edge and splicing lines; Figure 12 The difference lies in the spectral characteristics of cloud edge noise (red scatter) and wildfire pixels (density scatter). Detailed Implementation

[0024] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0025] According to an embodiment of the present invention, a wildfire monitoring method based on multi-source time-series remote sensing imagery is provided.

[0026] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 , Figure 2 As shown, a wildfire monitoring method based on multi-source time-series remote sensing imagery according to an embodiment of the present invention includes: Step S1: Acquire multi-source remote sensing data and construct a time-series band dataset, and use the top atmospheric reflectance (TOA reflectance) of the three-band spectrum to construct a high-temperature thermal anomaly index (TAI). Step S2: Based on the high temperature thermal anomaly index, extract primary fire point pixels, saturated pixels and high emissivity pixels from multi-source remote sensing data, and convert saturated pixels and high emissivity pixels into primary fire point pixels (i.e., identify wildfire pixels). Step S3: Extract saturated pixels and high emissivity pixels from the missing fire cluster cores in the multi-source remote sensing data, and supplement the missing primary fire point pixels in the high-temperature combustion center of the fire point. Step S4: Based on pre-configured constraints and building coverage area data (i.e., GHS construction land data) and high-frequency feature masks (i.e. Mask HFO Noise pixels are removed from the primary fire pixels to obtain secondary fire pixels (i.e. wildfire pixels). Step S5: Transmit the secondary fire point pixels to the local client (i.e., the local Python API) and the cloud geospatial analysis platform (i.e., the GEE cloud platform), and create a fire point list.

[0027] In this optional embodiment, the multi-source remote sensing data includes optical images from a land imager (i.e., Landsat-8 OLI) and optical images from a multispectral imager (i.e., Sentinel-2 MSI). The three-band spectrum includes far-shortwave infrared, near-shortwave infrared, and near-infrared bands; The high-temperature thermal anomaly index refers to the ratio of the reflectance difference between the far-shortwave infrared band and the near-shortwave infrared band to the reflectance of the near-infrared band.

[0028] It should be noted that for Landsat-8 OLI imagery, TAI = (ρ7 - ρ6) / ρ5, where ρ7 corresponds to the B7 band ( ), ρ6 corresponds to the B6 band ( ), ρ5 corresponds to the B5 band (NIR).

[0029] For Sentinel-2 MSI imagery, TAI = (ρ12 - ρ11) / ρ8a, where ρ12 corresponds to the B12 band ( ), ρ11 corresponds to the B11 band ( ), ρ8a corresponds to the B8A band (NIR).

[0030] Prepare multi-source remote sensing optical images and construct time-series band datasets for each. The multi-source remote sensing optical images include Landsat-8 OLI optical images and Sentinel-2 MSI optical images.

[0031] a. Wildfires are a common disturbance phenomenon in nature. The area burned by a wildfire generates a large amount of heat energy compared to the surrounding terrain, causing the wildfire to exhibit significantly different spectral characteristics compared to the surrounding terrain, which can then be monitored by passing remote sensing platforms. Based on Planck's blackbody radiation law, within the NIR to SWIR wavelength range, the emitted radiation of a wildfire increases exponentially with increasing wavelength. Wildfires in nature generally exhibit two forms: high-temperature open flames (1000–1200K) and low-temperature smoldering (600–800K). Figure 3The data shows that the radiance values ​​emitted by wildfires at different temperatures increase rapidly with increasing wavelength, with the most significant increase observed at 1200K.

[0032] b. The feasibility of using TOA reflectivity in monitoring thermal anomalies has been verified, and the response characteristics of wildfires to TOA reflectivity across different wavelengths are similar to those of radiance. Figure 3 As shown in b, this study screened the spectral response characteristics of wildfires and different land cover types in various bands from MSI / OLI imagery in the Arctic region. It is evident that wildfires exhibit characteristics opposite to other land cover types from the NIR to the SWIR bands: wildfire reflectance shows an exponential increase within the SWIR range, and the difference in reflectance between the two SWIR bands is significantly greater than that in the NIR band. This provides a theoretical basis for subsequent wildfire detection algorithms. Considering that GEE cloud stores complete TOA product data, subsequent wildfire detection algorithms are based on TOA reflectance inversion.

[0033] c. Spatial resolution, a key indicator for evaluating a sensor's ability to resolve subtle surface features, essentially reflects the precision with which a sensor captures information about small objects on the ground. Specifically, it characterizes the size of the area of ​​the ground imaged by the sensor's instantaneous field of view (IFOV), thus determining the sensor's detail resolution when observing small objects. In remote sensing, this is typically expressed as the linear size of the actual surface area represented by each pixel. In remote sensing applications monitoring flames or fire points, the detection capability of a satellite sensor depends not only on its spatial resolution but also on the energy emitted by the flame it captures. This means that the higher the sensor's spatial resolution, the smaller the ground unit it images, theoretically allowing for more accurate identification and location of small fire sources. All large fires originate from small fires, therefore, early detection of small fire sources and acquisition of crucial spatiotemporal location information are essential. Compared to low spatial resolution, satellite sensors with higher spatial resolution perform better in identifying small fires with subtle features because their instantaneous field of view occupies a larger proportion of the pixels, thereby improving detection accuracy. The IFOV of the Sentinel-2MSI (20×20m²) and Landst-8OLI (30×30m²) detectors is 2500 and 1111 times smaller than that of the 1×1km² MODIS detector, respectively. Therefore, for a 100m² thermal target, it occupies 25% of the MSI pixels, 11% of the OLI pixels, and only 0.01% of the MODIS IFOV. Assuming the thermal target is a blackbody (emissivity = 1), calculations using Planck's law show that only when the open flame temperature is sufficiently high (reaching 1000°C) can it be detected by the low-resolution MODIS IFOV. Figure 4 Correspondingly, MSI and OLI's IFOV can detect thermal targets with temperatures reaching 300°C via far- and short-wave infrared (FIR) signals. ) band detection, when the combustion temperature is 1000°C, can be simultaneously detected by far- and short-wave infrared ( ), near-shortwave infrared ( The MSI and NIR bands were detected simultaneously. Therefore, subsequent studies extracted wildfire combustion from MSI and OLI images based on the above three bands.

[0034] The high-temperature thermal anomaly index TAI= is proposed. ,in , , These correspond to the reflectance in the far-shortwave infrared, near-shortwave infrared, and near-infrared bands, respectively. This index quantifies the radiation characteristics of high-temperature thermal anomalies by utilizing the spectral differences in these three bands (i.e., the index quantifies the emission radiation contribution of wildfires by calculating the reflectance difference between the two fire-sensitive SWIR bands in the MSI and OLI).

[0035] Based on the above analysis of the spectral characteristics of wildfire combustion, the study uses near-shortwave infrared (NWI) spectra that are sensitive to fire. ) band (1.565~1.655μm), far and shortwave infrared ( A thermal anomaly index (TAI) was constructed by using the TOA reflectance of the SWIR band (2.1~2.5μm) and the NIR band (0.855~0.875μm). The TAI is based on the ratio of the reflectance difference between the two SWIR bands to the reflectance of the NIR band. Its expression is as follows: TAI= (1) In the formula, TAI represents the high-temperature thermal anomaly index; , , These correspond to MSI / OLI image bands B8A / B5, B11 / B6, and B12 / B7, respectively.

[0036] When extracting wildfire burning pixels based on TAI (Tracking Information Imagery), it is necessary to set reasonable thresholds for preliminary identification of fire pixels based on the actual scene. As an empirical analysis method, this study requires extensive sampling to determine the appropriate threshold size. Therefore, this study uses MODIS fire point products as an index to track and screen MSI / OLI tiles containing wildfire burning, using a visual discrimination method. Ultimately, the study sampled a total of 64 wildfire burning samples from multiple MSI / OLI images in the Arctic region, including 42 farmland fires and 22 bushfire scenes. The spatial distribution of the sampling points is shown below. Figure 6a,5a. The MSI sampling sample includes a total of 54,064 fire pixels ( Figure 5 b), including 33,231 forest fire pixels and 20,833 shrub fire pixels; the OLI sampling sample includes a total of 34,035 fire pixel samples ( Figure 6 (b) includes 30,012 forest fire pixels and 4,023 shrub fire pixels. Statistical analysis of the sampled scatter density map revealed that 95.2% of the sampled pixels exhibited a TAI value greater than or equal to 1, and the TAI values ​​of both forest and shrub wildfire pixels in both MSI and OLI imagery were greater than or equal to 0.45. Figure 5 Therefore, this study subsequently used a threshold of TAI ≥ 0.45 to perform preliminary extraction of clearly defined wildfire pixels.

[0037] In this optional embodiment, based on the high-temperature thermal anomaly index, primary fire pixels, saturated pixels, and high emissivity pixels are extracted from multi-source remote sensing data, and the saturated pixels and high emissivity pixels are converted into primary fire pixels, including: Step S21: Based on the high temperature thermal anomaly index, extract first-level fire point pixels with a value greater than or equal to the set threshold (i.e., 0.45) and saturated pixels and high emissivity pixels with a value less than the set threshold from the optical images of the land imager and the optical images of the multispectral imager. Step S22: Convert saturated pixels and high emissivity pixels into first-level fire pixels.

[0038] It should be noted that the threshold segmentation condition is TAI≥0.45, that is, TAI index greater than or equal to 0.45 is wildfire burning pixel.

[0039] Based on the Sentinel-2 QA60 and Landsat-8 QA_PIXEL bands, noise such as clouds and shadows is removed.

[0040] For Landsat-8 OLI and Sentinel-2 MSI images, pixels with TAI ≥ 0.45 were extracted to preliminarily identify potential wildfire pixels; This study used a threshold of TAI greater than or equal to 0.45 to identify most wildfire pixels in the Arctic region. However, when wildfires burned very intensely or the temperature was too high, wildfire pixels at the core of the fire clusters extracted using this threshold showed significantly low values ​​and were missed, resulting in a "hole" phenomenon. Figure 7 (8), which limits the integrity of wildfire detection.

[0041] In this optional embodiment, extracting saturated pixels and high emissivity pixels from the missing fire cluster cores in multi-source remote sensing data includes: Step S31: Based on the saturated pixels of the first-level preset band, supplement the missing fire cluster cores; Step S32: Based on the high emissivity pixels of the secondary preset band, supplement the missing fire cluster cores.

[0042] It should be noted that, 1) for Landsat-8 OLI and Sentinel-2 MSI images, pixels with TAI≥0.45 were extracted to preliminarily identify potential wildfire pixels; 2) When wildfires burn intensely or temperatures are excessively high, the wildfire pixels at the core of the fire clusters extracted using the threshold in 1) will show significantly low values ​​and be missed, resulting in a "hole" phenomenon. Therefore, based on the results in 1), extracting pixels that meet the threshold... >~ >0.4 and Saturated pixels >1 supplement potential missing fire pixels in the high-temperature combustion center of wildfires; 3) Similarly, since some high-emissivity fire pixels from wildfire combustion centers may be missed, based on the results of 1), extract those that meet the criteria... >2 and > High emissivity pixels >1 supplement potential missing fire pixels in the high-temperature combustion center of wildfires; 4) If there is a clearly defined wildfire pixel extracted in 1) in the eight-molar neighborhood of the saturated pixel and the high-radiation pixel, then add it to the result of 1); otherwise, it is considered as noise and is removed to effectively avoid missing the "hole" in the center of the fire cluster.

[0043] Based on a large number of MSI and OLI wildfire pixel samples and detailed spectral analysis, this study quantifies the radiative contribution of wildfires by calculating the reflectance difference between two fire-sensitive SWIR bands in the MSI and OLI. Furthermore, using the ratio between this difference and NIR band reflectance, with a fixed threshold of 0.45, we aim to extract distinct wildfire pixels in the Arctic region, as shown in the following formula: TAI ≥ 0.45; (2) In the formula, TAI represents the high-temperature thermal anomaly index.

[0044] To address the aforementioned "void" phenomenon—that is, some saturated and high-radiation fire pixels at the center of wildfire combustion are missed—the constraint condition of Equation 3 is used to detect these pixels. However, since some highly reflective objects may also satisfy this constraint condition, this study only labels pixels within the eight-molar neighborhood of the clearly identified fire pixels extracted by Equation 2 as saturated and high-radiation pixels of wildfires, as shown in the following formula: (3) In the formula, , , These correspond to the reflectance of the far-short-wave infrared, near-short-wave infrared, and near-infrared bands, respectively; AND represents logical AND; OR represents logical OR.

[0045] The first part of Equation 3 is used to label saturated pixels, while the second part is used to label high-emissivity pixels. Equations 2 and 3, used in combination, constitute the main body of the wildfire detection algorithm in this study, and have been visually verified. Figure 8 ).

[0046] In this optional embodiment, supplementing the missing fire cluster cores based on the saturated pixels of the first-level preset band includes: Step S311: Based on the saturation pixels where the far-shortwave infrared band is greater than 0.4 and the near-shortwave infrared band is greater than 1, supplement the missing fire cluster cores.

[0047] It should be noted that when discussing the detection capabilities of MSI and OLI imagery for wildfire scenes, their differences in radiosensitivity across different SWIR bands must be considered. Pixel saturation in MSI / OLI images is determined by their maximum resolvable radiance; the higher the radiance, the higher the achievable saturation brightness. Due to the small pixel area and specific system sensitivity of MSI and OLI sensors, when encountering large wildfires, the pixel corresponding to the center of the fire cluster... The wavelength often exceeds the maximum resolvable radiation value, leading to frequent pixel saturation in that band. For example... Figure 7 As shown in d and 8d, unlike high-emissivity pixels, saturated pixels... Compared to band In certain bands, because the lower maximum resolvable radiance value makes it easier to reach saturation, but the corresponding saturation brightness value will be lower. Bands, specifically represented by 1< < This leads to voids in the TAI (TAI < 0). It's worth noting that Landsat-8OLI differs somewhat from Sentinel-2MSI; due to saturation, OLI wildfire pixels may exhibit a unique pixel value folding phenomenon (DNfolding) in the band, meaning the maximum resolvable radiance reaches 24.3 W•m. -2 sr -1 μm -1 At the above time, the center of the fire cluster It will decrease significantly, reaching around 0.4 ( Figure 8 d).

[0048] In this optional embodiment, based on high emissivity pixels of the secondary preset band, the missing fire cluster core is supplemented. The supplementation of the primary fire point pixels missing in the high-temperature combustion center of the fire point includes: Step S321: Based on the high emissivity pixels where the ratio of the near short-wave infrared band to the far short-wave infrared band is greater than 2, and the near short-wave infrared band is greater than the far short-wave infrared band which is greater than 1, supplement the missing core of the fire cluster.

[0049] It should be noted that generally, the sensitivity to wildfire combustion is significantly higher, resulting in a higher TAI value, so wildfire pixels can be extracted ( Figure 7 c,8c). However, when the combustion of some fire cluster center pixels is too intense, the corresponding and bands both show high radiation energy, that is, > >1, but the reflectivity difference is significantly weaker than that of the surrounding fire pixels, resulting in a low TAI value and "holes" (0 < TAI < 0.45), and thus being missed ( Figure 6 d,7d).

[0050] As Figure 9 shown, the wildfire detection results based on the TAI algorithm can fully display the distribution details and characteristics of wildfires: 1) In the scenes of intense wildfire combustion, the fire points usually show a discontinuous linear distribution pattern surrounding into a circle, forming the so-called "fire line" ( Figure 9 a2-c2), so the wildfire detection results of MSI and OLI images can provide more accurate and detailed fire line position information. 2) Most of the fire clusters in the wildfire scene are composed of saturated / high emissivity fire pixels and the clear fire pixels around them, because the radiation energy of the outer fire pixels is generally lower than that of the flame core area and cannot reach the maximum resolvable radiation value. 3) The small and scattered wildfires around may be the diffusion direction of wildfires and the ignition source of the next large-scale fire.

[0051] In this optional embodiment, removing the noise pixels in the first-level fire point pixels based on the pre-configured limit conditions includes: Step S41: Add the first-level limit condition to remove the edge abnormal data of the imaging range difference of the data of different bands in the image boundary area; Step S42: Add the second-level limit condition to remove the missing or damaged pixels in the image caused by data packet loss; Step S43: Add the third-level limit condition to remove the cloud edge false alarms that meet the condition that the near short-wave infrared band is less than the coastal / aerosol band and the near infrared band; Step S44: Add the fourth-level limit condition to remove the bare soil noise that meets the condition that the far short-wave infrared band is greater than the near short-wave infrared band and the far short-wave infrared band is greater than 0.2; Step S45: Add the fifth-level limit condition to remove the ice and snow noise that meets the condition that the far short-wave infrared band is less than the near short-wave infrared band and the near infrared band is less than 0.65.

[0052] It should be noted that for bare soil noise: (Constraints to be added) >0.05, and >0.2, remove bare soil noise pixels; Ice and snow noise: add constraints: < and <0.65, remove pixels with ice and snow noise; While the TAI algorithm and saturation conditions described above can effectively identify wildfire pixels in most simple backgrounds, when detecting wildfires in large-scale, long-term image sequences, complex and diverse imaging conditions, underlying surface information, and inconsistent image quality can all generate atypical detection noise, thus affecting the accuracy of wildfire detection results. The noise sources for wildfire detection in large-scale, long-term MSI and OLI image sequences mainly include: Bare Soil and Snow: This study analyzed the spectral reflectance response curves of bare soil and snow in the built-in spectral library of ENVI, and found that in MSI and OLI... and The differences in reflectance across the bands are not significant. Furthermore, their reflectance is typically higher in the NIR band, and generally their TAI values ​​are lower. Figure 10 a, d). However, in long-term imagery wildfire detection, studies have found that similar noise resembling fire points can be generated on bare soil in some burned areas and on sunny slopes of snow-covered mountains. Figure 10 c, f). After sampling and spectral analysis ( Figure 10 b) Bare soil false alarm Usually slightly larger And most of them Since the value is greater than 0.2, the study added the following constraint 4 to the algorithm, which can effectively filter out bare soil noise (only 63 out of 8812 bare soil false alarm pixels were not filtered out). Therefore, the study added the following constraint 4 to the algorithm: (4) Figure 10 Scene f is a typical glacier landscape. Due to direct sunlight on the sunny slope, the near-SWIR reflectivity of the ice and snow surface is significantly enhanced, even exceeding the far-SWIR reflectivity. At the same time, the reflectivity in the NIR band is also significantly enhanced. A total of 354 ice and snow noises were detected. Figure 10 e), therefore, the study added constraint 5 to the algorithm, which can effectively remove noise from ice and snow. Therefore, the study added the following constraint 5 to the algorithm: (5) Cloud Edges and Linearity (MSI Optional): The MSI sensor consists of 12 detectors arranged in two horizontally staggered rows within the instrument, forming a wide field of view of 20.6°, achieving a scanning range of 290 km on the Earth's surface. Each detector is responsible for covering an area of ​​approximately 25 km. The key to the Sentinel-2L1C product lies in the data granules captured by these 12 detectors. These granules are stitched and reassembled using high-precision techniques to construct each complete MSI image. However, due to the unique arrangement of the detectors in the MSI, temporal deviations are unavoidable during data acquisition across different bands. Within the same image, since the imaging interval between data granules is approximately 2 seconds, when there is significant cloud cover between two adjacent granules, abnormal band reflectivity may appear in the stitching line area. This abnormal reflectivity is often misinterpreted as wildfire, resulting in linearly distributed false alarms and a large amount of linearly distributed noise. The study applied the aforementioned detection algorithm to the 08WNV image from July 13, 2021, detecting a total of 8,481 target pixels. After visual verification, 7,082 were identified as genuine wildfire pixels, while the remaining 1,399 were identified as noise. Figure 11 116 were removed from cloud band data, 1067 were near the cloud edge, and 216 came from splicing noise between data particles.

[0053] Comparative analysis of the spectral differences between cloud edge noise pixels and real fire pixels ( Figure 12 The study found significant differences in spectral characteristics between the two. Unlike fire pixels, cloud edge false alarm pixels... Typically smaller than ρ1 (coastal / aerosol band) and When both of the above conditions are met simultaneously, false alarms at cloud edges can be effectively filtered out. Therefore, the following constraint 6 was added to the algorithm: (6) To address the issue of mosaic line noise, the study found that Sentinel-2 L1C data provided by the European Space Agency, stored in ESA.SAFE format, includes metadata about the probe's trajectory. This metadata provides the WGS84 / UTM projection coordinates of each band's data particles before stitching, allowing for precise location of the mosaic lines. However, GEE cloud services do not provide access to this metadata in Sentinel-2 L1C products, while Amazon Web Services (AWS) allows users to download all data contained in the SAFE format. Therefore, by configuring Amazon S3, the study was able to batch retrieve the probe trajectory files corresponding to band 12 of daily MSI imagery of the Arctic region from the Amazon S3 data storage bucket to the local machine. By converting the data into vector files, the mosaic lines between data particles can be obtained. By merging the mosaic lines of band 12 with the same date daily and setting a buffer distance of 65m using the daily mosaic line database, most of the noise near the mosaic lines can be filtered out. Figure 11 Noise from all 216 splicing lines was filtered out.

[0054] Image data quality defects: MSI and OLI imagery contain a series of data quality issues that can lead to noise in wildfire detection. One issue is that outliers may appear in image boundary regions due to differences in the imaging range of different spectral bands. Since these boundary regions may contain null values ​​or missing data, they often appear as striped noise in the detection results. Therefore, constraint 7 was added to the algorithm: (7) Furthermore, during the transmission of MSI data to the ground, packet loss may occur, potentially leading to the loss or damage of certain pixels in the image, resulting in noticeable banded anomalies. Therefore, constraint 8 was added to the algorithm: (8) In the formula, i This indicates the different bands in MSI and OLI images. NoData This indicates missing data in the image.

[0055] In this optional embodiment, based on pre-configured building coverage area data and high-frequency ground feature masks, the formula for removing noise pixels from primary fire point pixels is as follows: (9) In the formula, Mask HFO Indicates a high-frequency ground feature mask; F(x,y,t i ) Indicates ini Fire pixels were detected on the date; n Indicates within the valid time range n Number of fire pixels detected per instance; Object j Represents a planar vector object generated using a point statistical raster. j .

[0056] In this optional embodiment, transmitting secondary fire pixels to a local client and a cloud-based geospatial analysis platform includes: Step S51: Transmit the secondary fire point pixels to the local client. The local client uses standardized data as an index to obtain image number information from the cloud geospatial analysis platform and transmits it to the local image task list. Step S52: Based on the standardized data generated by the local client, the fire point pixels are calculated step by step through the cloud geospatial analysis platform and stored in the binarized image; combined with the automated processing mechanism, they are batch converted into potential fire cluster objects in the form of four-neighbor connection. Step S53: Use the raster-to-polygon tool in the geographic information software to convert the binarized image into a vector polygon; Step S54: Calculate the centroid of each vector polygon using geometric calculation tools to obtain the daily fire point dataset.

[0057] It should be noted that, based on the index file, calculation requests are sent to the GEE cloud one by one, and the processing results will be returned in the form of binary raster and stored locally (0 represents non-burning pixels; 1 represents potential wildfire pixels).

[0058] In ArcGIS, the Raster To Polygon and Calculate Geometry tools are used to convert raster data to vector data and calculate centroids, ensuring positional accuracy.

[0059] To achieve efficient parallel computing of remote sensing big data, this study constructs a hybrid computing framework based on the GEE cloud computing platform and local computing. Figure 2 The detailed steps for accurately extracting Arctic wildfires from massive time-series imagery are as follows: 1) Creating a local imagery task list. The local client plays a crucial role in the data processing workflow. It can send customized requests to the GEE cloud platform, indexed by the names of all tiles in the Arctic region, to retrieve time-series imagery files for those tiles from 2016 to 2022. Once the requests are responded to, the corresponding imagery information is captured by the local client and organized into a task list to be processed (a total of 1,580,294 index files).

[0060] 2) GEE Algorithm Integration Processing. The study integrated a wildfire detection algorithm and a noise processing module into a local Python environment. Based on the index file generated in the previous step, computation requests were sent to the GEE cloud one by one. This method aims to avoid system overload that might be caused by concurrent uploads. Through the designed script, the advantages of parallel computing can be effectively utilized to achieve parallel processing of local tasks. Once the GEE cloud completes the computation, the processing results are returned in the form of a binary raster and stored locally (0 represents non-burning pixels; 1 represents potential wildfire pixels).

[0061] 3) Local batch processing to construct a daily wildfire inventory. For the stored binarized wildfire imagery, Arcpy (an automated processing mechanism) was used to batch convert it into potential fire clusters using a four-neighbor join method. It is worth noting that due to data overlap between adjacent tiles in the MSI and OLI images, a fusion process was performed after merging the daily areal fire clusters to avoid data redundancy and statistical errors. Then, the centroids of the daily areal data were calculated to obtain a daily potential fire point dataset. Daily mosaic line data was used to remove false alarms. Finally, the daily fire point data, after removing false alarms, was synthesized by year. High-resolution built-up area data and a generated MaskHFO high-frequency feature mask were used to remove noise from the high-frequency scene, successfully creating the first 20 / 30-meter spatial resolution wildfire fire point inventory in the Arctic region.

[0062] 4) For the stored binarized wildfire imagery, using ArcPy with the local environment, the images are batch-converted into potential fire clusters using a four-neighbor connection method. Then, the Raster To Polygon tool in ArcGIS is used to convert the binarized images into vector polygons, retaining only polygons with an attribute value of 1. The CalculateGeometry tool is used to calculate the centroid of each polygon, obtaining candidate wildfire locations, i.e., the daily potential fire point dataset. The daily fire point data, after removing false alarms from the mosaic lines, is synthesized by year, combined with high-resolution built-up area data and the generated MaskHFO high-frequency ground feature data. The masking tool in ArcGIS is then used to remove noise from the high-frequency scene.

[0063] In this optional embodiment, creating the fire point list further includes: Step S55: Based on vector polygons and combined with the daily fire point dataset, define the fire point burning area; Step S56: Using the field calculation tool in the geographic information software, count the number of pixels, detection date, latitude and longitude coordinates, and combustion type information within the combustion area of ​​each fire point.

[0064] It should be noted that: 1) Based on the vector polygon in step 7, combined with the centroid point in step 5, the accurate wildfire burning range is defined; 2) Use the Calculate Field tool in ArcGIS to add different fields to count the number of pixels, detection date, latitude and longitude coordinates, and burning type of each wildfire area.

[0065] To facilitate understanding of the above technical solutions of the present invention, the working principle or operation method of the present invention in actual process will be described in detail below.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wildfire monitoring method based on multi-source time-series remote sensing imagery, characterized in that, include: S1. Acquire multi-source remote sensing data and construct a time-series band dataset, and use the atmospheric top reflectance of the three-band spectrum to construct a high-temperature thermal anomaly index. S2. Based on the high-temperature thermal anomaly index, extract first-level fire point pixels, saturated pixels and high emissivity pixels from multi-source remote sensing data, and convert saturated pixels and high emissivity pixels into first-level fire point pixels. S3. Extract saturated pixels and high emissivity pixels from the missing fire cluster cores in multi-source remote sensing data, and supplement the missing primary fire point pixels in the high-temperature combustion center of the fire point. S4. Based on pre-configured constraints, building coverage area data, and high-frequency ground cover masks, noise pixels are removed from the primary fire pixel to obtain the secondary fire pixel. S5. Transmit the secondary fire point pixels to the local client and the cloud geospatial analysis platform, and create a fire point list.

2. The wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 1, characterized in that, The multi-source remote sensing data includes optical images from land imagers and optical images from multispectral imagers. The three-band spectrum includes far-shortwave infrared, near-shortwave infrared, and near-infrared bands; The high-temperature thermal anomaly index refers to the ratio of the reflectance difference between the far-shortwave infrared band and the near-shortwave infrared band to the reflectance of the near-infrared band.

3. The wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 2, characterized in that, The process of extracting primary fire point pixels, saturated pixels, and high emissivity pixels from multi-source remote sensing data based on high-temperature thermal anomaly indicators, and converting saturated pixels and high emissivity pixels into primary fire point pixels, includes: S21. Based on the high temperature thermal anomaly index, extract first-level fire point pixels that are greater than or equal to the set threshold, and saturated pixels and high emissivity pixels that are less than the set threshold from the optical images of the land imager and the optical images of the multispectral imager. S22. Convert saturated pixels and high emissivity pixels into first-order fire pixels.

4. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 3, characterized in that, The extraction of saturated pixels and high emissivity pixels from the missing fire cluster cores in multi-source remote sensing data includes: S31. Based on the saturated pixels of the first-level preset band, supplement the missing fire cluster cores; S32. Based on the high emissivity pixels of the secondary preset band, supplement the missing fire cluster cores.

5. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 4, characterized in that, The supplementation of missing fire cluster cores based on saturated pixels in a first-level preset band includes: S311. Based on saturated pixels with a value greater than 0.4 in the far-shortwave infrared band and greater than 1 in the near-shortwave infrared band, supplement the missing fire cluster cores.

6. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 5, characterized in that, The high emissivity pixels based on the secondary preset band supplement the missing fire cluster core, and the supplementary primary fire point pixels missing in the high-temperature combustion center of the fire point include: S321. Based on high emissivity pixels where the ratio of near-shortwave infrared band to far-shortwave infrared band is greater than 2, and the ratio of near-shortwave infrared band to far-shortwave infrared band is greater than 1, missing fire cluster cores are supplemented.

7. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 6, characterized in that, The process of removing noisy pixels from primary fire pixels based on pre-configured constraints includes: S41. Add a first-level constraint to remove edge anomalies caused by differences in the imaging range of different band data in the image boundary area. S42. Add secondary restrictions to remove data packet loss that leads to missing or damaged pixels in the image; S43. Add three-level restrictions to eliminate false alarms at cloud edges that meet the criteria of near-shortwave infrared band being smaller than the coastal / aerosol band and near-infrared band. S44. Add a fourth-level constraint to remove bare soil noise that meets the condition that the far-shortwave infrared band is greater than the near-shortwave infrared band and the far-shortwave infrared band is greater than 0.

2. S45. Add five levels of restrictions to remove ice and snow noise that meets the conditions that the far-shortwave infrared band is smaller than the near-shortwave infrared band and the near-infrared band is less than 0.

65.

8. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 1, characterized in that, The formula for removing noisy pixels from primary fire point pixels based on pre-configured building coverage area data and high-frequency ground feature masks is as follows: ； In the formula, Mask HFO Indicates a high-frequency ground feature mask; F(x,y,t i ) Indicates in i Fire pixels were detected on the date; n Indicates within the valid time range n Number of fire pixels detected per instance; Object j Represents a planar vector object generated using a point statistical raster. j .

9. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 1, characterized in that, The process of transmitting secondary fire point pixels to the local client and the cloud-based geospatial analysis platform includes: S51. The secondary fire point pixels are transmitted to the local client. The local client uses standardized data as an index to obtain image number information from the cloud geospatial analysis platform and transmits it to the local image task list. S52. Based on standardized data generated by the local client, the data is gradually calculated through the cloud geospatial analysis platform, and the fire point pixels are stored in the binary image. Combined with an automated processing mechanism, they are batch-converted into potential fire cluster objects in a four-neighbor connection manner. S53. Use the raster-to-polygon tool in geographic information software to convert the binarized image into a vector polygon. S54. Use geometric calculation tools to calculate the centroid of each vector polygon to obtain the daily fire point dataset.

10. A wildfire monitoring method based on multi-source time-series remote sensing imagery according to claim 9, characterized in that, The creation of the fire point list also includes: S55. Based on vector polygons and combined with daily fire point datasets, define the fire point burning area; S56. Using the field calculation tool in the geographic information software, count the number of pixels, detection date, latitude and longitude coordinates, and combustion type information within the combustion area of ​​each fire point.

Images