Forest fire smoke identification method and system for Sentinel-2 satellite under cloud interference

Abstract

The invention discloses a forest fire smoke identification method and system for a Sentinel-2 satellite under cloud interference. The forest fire smoke identification method comprises the following steps: step 1, downloading and preprocessing a Sentinel-2 image; step 2, carrying out Sentinel-2 image segmentation and carrying out pixel reflectivity reassignment; step 3, a forest fire smoke index (FFSI) is constructed, and a forest fire smoke index (FFSI) is constructed; and step 4, determining an FFSI threshold value and identifying forest fire smoke based on the FFSI. According to the method, the defects that when a pixel-oriented identification method is purely adopted, the reflectivity difference of pixels in a smog area is obvious, and mistaken classification is easily caused are overcome. Meanwhile, forest fire smoke extraction can be realized only through a single image and one-time threshold judgment, the difficulty of forest fire smoke recognition based on a multiband reflectivity threshold method is reduced, and the problem that forest fire smoke recognition based on a multi-temporal method is easily limited by multi-temporal data accessibility is solved. And the availability of cloud remote sensing images in a multi-cloud area or a multi-cloud season is improved.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing imagery, specifically relating to a method and system for identifying forest fire smoke under cloud interference using the Sentinel-2 satellite. Background Technology

[0002] Due to climate change and extreme droughts, large-scale forest fires have frequently occurred around the world in recent years. These fires have not only devoured vast tracts of precious forest resources and damaged fragile ecosystems, but have also seriously threatened the lives of residents in many regions. Satellite remote sensing technology has significant application value in the field of forest fire monitoring. Utilizing satellite remote sensing technology to quickly and accurately locate fires is a prerequisite for achieving early detection and suppression of fires and reducing the damage caused by wildfires.

[0003] Forest fire remote sensing identification involves two approaches: one is using thermal infrared remote sensing to detect thermal anomalies caused by the fire, and the other is using optical remote sensing to detect smoke signals. The former uses thermal infrared remote sensing data with a much lower spatial resolution than the latter uses optical remote sensing data. In the early stages of a fire, low-spatial-resolution thermal infrared remote sensing struggles to capture thermal anomalies within a small area, while higher-spatial-resolution optical remote sensing can capture smoke signals. Therefore, forest fire smoke optical remote sensing identification is the best way to detect wildfires in their earliest stages.

[0004] Clouds and forest fire smoke share similar spectral response characteristics in optical remote sensing images, and clouds are ubiquitous in optical satellite remote sensing images. Therefore, the presence of clouds significantly increases the difficulty of forest fire smoke identification. Multi-band reflectance thresholding and multi-temporal methods are two common approaches to forest fire smoke identification. Multi-band reflectance thresholding requires multiple indicators and multiple threshold judgments to identify forest fire smoke, making it complex to operate and the identification results highly sensitive to threshold values. Multi-temporal methods require cloudless images before the fire as a reference; however, in cloudy areas or during cloudy seasons, the presence of clouds significantly reduces the availability of multi-temporal remote sensing data. Furthermore, both multi-thresholding and multi-temporal methods are pixel-based forest fire smoke identification. The reflectance of different forest fire smoke pixels is not entirely the same across different bands. For example, a patch of smoke consists of numerous smoke pixels, and due to the influence of smoke concentration, composition, and lighting conditions, the reflectance of different smoke pixels is not entirely the same, and may even differ significantly. Therefore, when using a pixel-oriented approach to identify forest fire smoke, it is easy for some pixels in a patch of smoke to be mistakenly classified as non-smoke.

[0005] The commonly used smoke detection methods are as follows:

[0006] (1) Forest fire smoke identification based on multi-band reflectivity threshold method

[0007] The principle of forest fire smoke recognition based on multi-band reflectance thresholds is that the reflectance of forest fire smoke differs from that of other objects across various bands. By analyzing a large amount of smoke reflectance data, reflectance thresholds for distinguishing smoke from non-smoke in each band can be determined, and a series of threshold conditions can be combined to differentiate smoke pixels from non-smoke pixels. This method is simple and intuitive in principle, has clear physical meaning, and is computationally efficient and fast. However, this method requires multiple thresholds, and setting the reflectance thresholds for distinguishing smoke from clouds is difficult, highly sensitive, and lacks universality. Fixed thresholds are difficult to adapt to forest fire smoke recognition under complex environments such as different seasons, regions, and weather conditions.

[0008] (2) Forest fire smoke identification based on multi-temporal method

[0009] The core idea of ​​multi-temporal smoke discrimination for forest fires is that the spectral response characteristics of remote sensing images change significantly before and after a fire. For example, smoke from a forest fire will cover the surface information at the corresponding location, thus altering the spectral response characteristics of the remote sensing image. In practical applications, based on remote sensing images before and during a fire, a change detection method and a change threshold are designed by comparing the changes in spectral response characteristics at the same location (pixel). Pixels exceeding a certain change threshold are considered forest fire smoke pixels. However, while multi-temporal smoke identification relies on the rapid and significant changes in the spectral characteristics of remote sensing images before and after a fire to detect smoke, cloud cover significantly reduces the availability of multi-temporal remote sensing data in cloudy areas or seasons, thus limiting the widespread application of this algorithm. Summary of the Invention

[0010] To address the aforementioned issues, this invention discloses a method and system for identifying forest fire smoke under cloud interference using the Sentinel-2 satellite.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] A method for identifying forest fire smoke under cloud interference using Sentinel-2 satellite includes the following steps:

[0013] Step 1: Downloading and preprocessing Sentinel-2 satellite imagery:

[0014] Download Sentinel-2 satellite L2A level data and convert the DN values ​​of the L2A level data into true reflectance to obtain reflectance data of Sentinel-2 satellite imagery;

[0015] Step 2: Input the reflectance data into the image segmentation tool of the image processing software to automatically segment it into N objects of varying sizes and irregular boundaries; calculate the average reflectance R of all pixels within each object in the i-th band. (i,m)i = 1, 3, or 9; where 1 represents aerosol band B1, 3 represents green light band B3, and 9 represents water vapor detection band B9; the average reflectance R of the m-th object is used as the reference value. (i,m) The reflectance of all pixels of the m-th object is replaced by the reflectance of the m-th object.

[0016] Step 3: Construct the Forest Fire Smoke Index (FFSI):

[0017] FFSI=f(s)×f(c) (2)

[0018]

[0019] Among them, R (1,m) R (3,m) and R (9,m) The average reflectance of the m-th object is characterized in aerosol band B1, green light band B3, and water vapor detection band B9, respectively.

[0020] Step 4: Identify objects with a positive FFSI that are not greater than a preset threshold as forest fire smoke.

[0021] A further improvement is made to the method of converting the DN value into true reflectance in step one, as follows:

[0022] R = DN / 100

[0023] In the formula: R is the true reflectivity; DN is the original DN value of the Sentinel-2 satellite L2A product.

[0024] As a further improvement, the image processing software is ArcGIS Pro.

[0025] As a further improvement, the preset threshold is 14.

[0026] A forest fire smoke identification system under cloud interference for Sentinel-2 satellite, wherein the forest fire satellite remote sensing image reconstruction system is used to operate the above-mentioned forest fire smoke identification method under cloud interference for Sentinel-2 satellite.

[0027] Advantages of this invention:

[0028] This invention leverages the reflectance differences of forest fire smoke and clouds in the aerosol, visible light, and water vapor bands of Sentinel-2 imagery to construct a forest fire smoke recognition method that combines pixel-oriented and object-oriented approaches. Forest fire smoke extraction requires only a single image and a single threshold judgment. This invention overcomes the drawbacks of purely pixel-oriented recognition methods, which suffer from significant differences in reflectance among pixels within large smoke areas, leading to misclassification. Furthermore, this invention reduces the difficulty of forest fire smoke recognition based on multi-band reflectance thresholding methods by requiring only a single image and a single threshold judgment. It also avoids the limitations imposed by the availability of multi-temporal data in forest fire smoke recognition based on multi-temporal methods, improving the usability of remote sensing imagery of cloudy areas or cloudy seasons. This method can also be applied to forest fire smoke recognition in other areas with satellite data in the aerosol and water vapor bands. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention.

[0030] Figure 2 This paper presents a comparison between the results of extracting the spatial distribution of forest fire smoke from Sentinel-2 satellite imagery and pixel-oriented spatial distribution results. (a) shows the Sentinel-2 satellite true-color composite image and segmentation boundaries; (b) and (c) show the superposition results of forest fire smoke extracted by different methods and the true-color composite image (green represents forest fire smoke, with a transparency of 40%). Specifically, (b) shows the forest fire smoke distribution result extracted based on the average reflectance of each object obtained in step 2 and FFSI; (c) shows the forest fire smoke distribution result extracted based on the original reflectance in step 1 and FFSI. The comparative analysis shows that when using the same FFSI model and threshold to extract forest fire smoke, this invention, by using the average reflectance of each object instead of the original reflectance of the pixel, can obtain a complete forest fire smoke outline with complete continuity within the smoke, effectively suppressing the misclassification problem caused by differences in smoke pixel reflectance. However, the results of forest fire smoke identification obtained by using the original reflectance of each pixel are very fragmented. A large number of smoke pixels in the smoke are misclassified as other types. Moreover, a large number of non-smoke pixels are misclassified as smoke pixels. This phenomenon is mainly caused by a few non-smoke pixels having similar spectral characteristics to forest fire smoke pixels. When the present invention uses the average reflectance instead of the original reflectance, the phenomenon of non-smoke pixels being misclassified as smoke pixels is greatly reduced. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Example 1

[0033] like Figure 1This invention presents a method for identifying forest fire smoke under cloud interference using Sentinel-2 satellite imagery. This method leverages the differences in reflectance characteristics of forest fire smoke and clouds in the aerosol, visible light, and water vapor bands of Sentinel-2 satellite imagery. It constructs a forest fire smoke identification method that combines object-oriented and pixel-oriented approaches, applicable to cloud-covered satellite imagery. Forest fire smoke extraction can be achieved with only a single image and a single threshold judgment. This invention overcomes the drawbacks of simply using pixel-oriented identification methods, which suffer from significant differences in reflectance within large smoke areas and are prone to misclassification. It reduces the difficulty of forest fire smoke identification based on multi-band reflectance thresholding methods and avoids the limitations imposed by the availability of multi-temporal data in forest fire smoke identification based on multi-temporal methods, thus improving the usability of remote sensing imagery of cloudy areas or cloudy seasons. The main steps are as follows: Step 1, Sentinel-2 image download and preprocessing; Step 2, Sentinel-2 image segmentation and pixel reflectance reassignment; Step 3, Forest Fire Smoke Index (FFSI) construction; Step 4, FFSI threshold determination and FFSI-based forest fire smoke identification.

[0034] Step 1: Downloading and Preprocessing Sentinel-2 Satellite Imagery

[0035] Download Sentinel-2 L2A-level data that includes both clouds and wildfire smoke. Sentinel-2 imagery contains 13 bands: aerosol band B1, blue band B2, green band B3, red band B4, red-edge bands 1-B5, red-edge bands 2-B6, red-edge bands 3-B7, near-infrared band B8, and narrow near-infrared band B... 8A Water vapor detection band B9, cloud detection band B 10 Shortwave infrared band 1-B 11 and shortwave infrared band 2-B 12 Although the Sentinel-2 satellite L2A level data has undergone atmospheric correction, its DN value is not the true reflectance. It still needs to be converted by a coefficient to obtain the true reflectance. The conversion method between reflectance and DN value is shown in formula (1):

[0036] R = DN / 100 (1)

[0037] In the formula: R is the reflectivity, which ranges from 0% to 100%, and DN is the original value of the Sentinel-2 satellite L2A level data.

[0038] Step 2: Sentinel-2 satellite image segmentation and pixel reflectance reassignment

[0039] A patch of smoke is composed of numerous smoke pixels. Influenced by smoke concentration, composition, and lighting conditions, the reflectance of different smoke pixels is not entirely the same, and can even vary significantly. Therefore, when using a pixel-oriented approach for forest fire smoke identification, it is easy for some pixels within a patch of smoke to be mistakenly classified as non-smoke. To address this issue, a new framework for forest fire smoke identification under cloud interference is proposed, combining object-oriented and pixel-oriented approaches. First, the reflectance data obtained in step one is input. Using the image segmentation tool in ArcGIS Pro software, based on the spectral and textural features of the image, the Sentinel-2 satellite image is automatically segmented into N objects of varying sizes and irregular boundaries. Each object consists of several pixels, and at this point, all pixels within each object can be considered to be of the same type. Second, taking each object as a unit, the average reflectance R of all pixels contained within that object in the i-th band is calculated band by band. (i,m) The average reflectance is then used to replace the original reflectance of each pixel in the i-th band. This preserves the regularity of the spectral characteristics of objects of the same type while ensuring the homogeneity of the reflectance of all pixels within each object. Based on this, forest fire smoke recognition can eliminate the situation where some smoke pixels in the same smoke patch are mistakenly classified as non-smoke pixels due to spectral differences.

[0040] Step 3: Construction of the Forest Fire Smoke Index (FFSI)

[0041] Extensive experimental data show that the reflectance of forest fire smoke decreases significantly from the aerosol band (B1) to the green band (B3), while that of vegetation, bare land, and built-up areas increases significantly. Therefore, models can be constructed using the aerosol and green bands of Sentinel-2 satellite imagery to distinguish forest fire smoke from vegetation, bare land, and built-up areas. Furthermore, forest fire smoke and clouds show significant differences in the water vapor band; therefore, a water vapor band can be added to differentiate forest fire smoke from clouds. The calculation method for FFSI is shown in formula (2):

[0042] FFSI=f(s)×f(c) (2)

[0043]

[0044] In the formula: R (1,m) R (3,m) and R (9,m) The average reflectance of each object obtained in step 2 in aerosol band B1, green light band B3 and water vapor detection band B9 is characterized, rather than the original reflectance of each pixel.

[0045] For f(s), forest fire smoke has a positive value, while bare land, vegetation, and built-up areas have negative values. Clouds have both positive and negative values, so f(s) can be used to exclude vegetation, bare land, and built-up areas. Clouds have a much higher reflectivity in the water vapor band (B9) than forest fire smoke. Using an exponential form can further amplify the difference between clouds and smoke. Therefore, although both clouds and smoke have positive values ​​in f(c), the cloud value is much higher than the smoke value, which can be used to distinguish between smoke and clouds. When f(s) × f(c) is used to obtain the FFSI, forest fire smoke has a positive value in the FFSI, while bare land, vegetation, and built-up areas have negative values. Cloud values ​​are both positive and negative, but in the positive range, the cloud value is much higher than that of forest fire smoke. Therefore, after obtaining the FFSI, it is only necessary to select an appropriate threshold in the positive range to achieve forest fire smoke identification under cloud interference.

[0046] Step 4: FFSI Threshold Determination and FFSI-Based Forest Fire Smoke Identification

[0047] Based on the forest fire smoke index obtained in step 3, the distribution of forest fire smoke and cloud values ​​on the FFSI was statistically analyzed to determine the FFSI value range for identifying forest fire smoke. Experimental results show that the optimal value range for forest fire smoke on the FFSI is 0.

Claims

1. A method for identifying forest fire smoke under cloud interference using Sentinel-2 satellite, characterized in that, Includes the following steps: Step 1: Downloading and preprocessing Sentinel-2 satellite imagery: Download Sentinel-2 satellite L2A level data and convert the DN values ​​of the L2A level data into true reflectance to obtain reflectance data of Sentinel-2 satellite imagery; Step 2: Input the reflectance data into the image segmentation tool of the image processing software to automatically segment it into N objects of varying sizes and irregular boundaries; calculate the average reflectance R of all pixels within each object in the i-th band. (i,m) i = 1, 3, or 9; where 1 represents aerosol band B1, 3 represents green light band B3, and 9 represents water vapor detection band B9; the average reflectance R of the m-th object is used as the reference value. (i,m) The reflectance of all pixels of the m-th object is replaced by the reflectance of the m-th object. Step 3: Construct the Forest Fire Smoke Index (FFSI): FFSI=f(s)×f(c) (2) Among them, R (1,m) R (3,m) and R (9,m) The average reflectance of the m-th object is characterized in aerosol band B1, green light band B3, and water vapor detection band B9, respectively. Step 4: Identify objects with a positive FFSI that are not greater than a preset threshold as forest fire smoke.

2. The method for identifying forest fire smoke under cloud interference for Sentinel-2 satellite as described in claim 1, characterized in that, In step one, the method for converting the DN value into true reflectance is as follows: R = DN / 100 In the formula: R is the true reflectivity; DN is the original DN value of the Sentinel-2 satellite L2A product.

3. The method for identifying forest fire smoke under cloud interference for Sentinel-2 satellite as described in claim 1, characterized in that, The image processing software is ArcGIS Pro.

4. The method for identifying forest fire smoke under cloud interference for Sentinel-2 satellite as described in claim 3, characterized in that, The preset threshold is 14.

5. A forest fire smoke identification system under cloud interference for Sentinel-2 satellite, characterized in that, The forest fire satellite remote sensing image reconstruction system is used to run the forest fire smoke identification method under cloud interference for Sentinel-2 satellite as described in any one of claims 1-4.

Images