A high-resolution monitoring method for forest fire burning areas

Abstract

The present invention relates to the technical field of forest fire monitoring, and specifically to a high-resolution monitoring method for forest fire burning areas. Based on high-resolution satellite image data, the present invention predicts forest burning areas by using machine learning algorithms such as distance weighted functions of image pixels, and predicts the burning spatial range by using a threshold classification method, and combines accuracy verification to effectively monitor and predict the burning areas of forest fires, thereby providing a basis for rapid and large-scale monitoring of forest fires and providing strong support for subsequent post-disaster management and standard formulation.

Description

Technical Field

[0001] The invention relates to the technical field of forest fire monitoring, in particular to a high-resolution monitoring method for forest fire burning areas. Background Art

[0002] In recent years, forest fires have occurred frequently due to climate change and human activities, which not only seriously damaged forest resources and the ecological environment, but also caused great harm to people’s lives and property and public safety, and posed a huge threat to economic sustainable development and ecological security.

[0003] When a forest fire occurs, being able to accurately and quickly detect the burning area is crucial for timely extinguishing the fire, thereby reducing losses, prevention, and the scientific allocation of subsequent related fire prevention resources and the formulation of rules. However, in the forest, due to the obstruction of lush trees and the large area of ​​forest coverage, it is impossible to quickly detect the burning area only through manual means such as visual inspection, which affects the accuracy and timeliness of fire extinguishing. However, there is no effective solution to this problem in the existing technology.

[0004] Therefore, it is necessary to design a high-resolution monitoring method for forest fire burning areas. Through a fast, large-scale monitoring method with high spatial resolution related to forest fires, it can be used to identify the spatial extent of forest fire burning areas, providing positive and scientific technical support for post-disaster management and assessment. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-resolution monitoring method for forest fire burning areas. Through a high-spatial-resolution monitoring method related to forest fires that is fast and can be used to quickly and widely monitor the burning areas, the present invention is used to identify the spatial extent of forest fire burning areas, thereby providing positive and scientific technical support for post-disaster management and assessment.

[0006] In order to achieve the above object, the present invention provides a high-resolution monitoring method for forest fire burning areas, comprising the following steps:

[0007] S1, Satellite data processing:

[0008] S1-1, data download:

[0009] S1-1-1, download satellite image data of 13 spectral bands of the Sentinel-2 multispectral instrument, with spatial resolution covering the visible, near-infrared and short-wave infrared bands;

[0010] S1-1-2, download the series of image data from the Gaofen satellite, the resolution of the panchromatic band is 1m, and the resolution of other multispectral bands is 4m;

[0011] S1-1-3, all the above satellite image data are reprojected onto the Universal Transverse Mercator UTM projection based on WGS84, where the image data in S1-1-1 is used as input data for mapping forest fires, and the image data in S1-1-2 is used as reference data for evaluating the accuracy of forest fire assessment;

[0012] S1-2, data preprocessing:

[0013] S1-2-1, the satellite images were geometrically orthorectified and georeferenced, and the interference of clouds, cloud shadows and snow was removed using the Fmask (Function ofmask) mask function; then 100 control points were added to determine the spatial position correspondence between the Sentinel-2 images and the Gaofen satellite series images, and a first-order polynomial transformation was selected;

[0014] S1-2-2, calculate the normalized difference vegetation index NDVI to represent the forest growth status, as shown in the following formula:

[0015]

[0016] Among them, NIR and RED are the surface reflectance values ​​in the infrared band and near infrared band respectively;

[0017] S2, extraction of spatial distribution information of forest fire burning areas:

[0018] S2-1, reduce image noise:

[0019] The spectral features of 300 sample points of bare land category A, forest category B, and forest fire burn area category C were manually extracted from Sentinel-2 images. Using the random forest model, 30% of the 900 sample points were randomly selected as training samples and 70% as test samples. The trained random forest model was used to traverse the Sentinel-2 images to reduce image noise.

[0020] S2-2, Calculation of forest fire burning area probability:

[0021] For each pixel pi of the image, its probability of being in class C, i.e., the forest fire burn area, is predicted based on the distance weighted function of its j=1, 2, ..., J neighboring pixels, as shown in Equation 2:

[0022]

[0023] in is the probability that the image pixel pi is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire area, λ ij is the weight related to the distance, as shown in Formula 3:

[0024]

[0025] where h ij is the distance between pixel i and pixel j, α is the nonlinear parameter of the exponential model;

[0026] S2-3, Identification of the spatial extent of forest fire burning areas:

[0027] The probability of forest fire burning area for each image pixel is arranged in descending order and then classified. The classification standard is identified by threshold classification method, as shown in the following formula 4:

[0028]

[0029] Among them, P is the classification threshold, P max and P min are the maximum and minimum probability values ​​of the pixels in the forest fire burning area, respectively, and d B (x, y) is the distance from pixel x to the nearest forest fire burn area sample point y; when the image pixel probability ≥ P, it is set as a forest fire burn area; when the image pixel probability < P, it is set as a non-forest fire burn area;

[0030] The unburned areas in the forest fire burning area are excluded to identify the closed forest fire burning area, as shown in the following formula 5:

[0031]

[0032] in is the probability that the image pixel pj is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire burn area, and A is the probability;

[0033] S3, accuracy verification:

[0034] The accuracy is evaluated using overall accuracy, root mean square error, mean absolute error, and systematic error, where overall accuracy, root mean square error, and mean absolute error are used to calculate accuracy, while systematic error is used to quantify the estimation bias, as shown in Equations 6 to 8:

[0035] OA=A m / A;

[0036]

[0037] Where OA is the overall accuracy, A m is the area of ​​forest fire burnt area identified, A is the area of ​​forest fire burnt area actually reported, is the forest fire burning area of ​​sample i, f i is the forest fire burn area of ​​sample i manually interpreted, n is the total number of sample polygons, RMSE is the root mean square error, MAE is the mean absolute error, and SE is the systematic error.

[0038] The 13 spectral bands in S1-1-1 are 4 10-meter bands, 6 20-meter bands and 3 60-meter bands.

[0039] The panchromatic band in S1-1-2 is band 1, and the other multispectral bands are bands 2 to 5.

[0040] The empirical value of α in Formula 3 is 3.

[0041] σ p It is 10 kilometers.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] Based on high-resolution satellite image data, the present invention predicts forest burning areas through machine learning algorithms such as distance weighted functions of image pixels, and predicts the burning spatial range through threshold classification method, and combines with accuracy verification. It can effectively monitor and predict the burning areas of forest fires, provide a basis for rapid and large-scale monitoring of forest fires, and provide strong support for subsequent post-disaster management and standard setting. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the overall flow chart of the present invention;

[0045] Figure 2 A schematic diagram of a satellite for precision verification according to an embodiment of the present invention; DETAILED DESCRIPTION

[0046] The present invention will now be further described with reference to the accompanying drawings.

[0047] See also Figures 1-2 The present invention provides a high-resolution monitoring method for forest fire burning areas, comprising the following steps:

[0048] S1 satellite data processing:

[0049] S1-1 satellite data download:

[0050] The main satellite image data downloaded and used were from the Sentinel-2 Multispectral Instrument (MSI), with a total of 13 spectral bands: 4 10-meter bands, 6 20-meter bands, and 3 60-meter bands, with spatial resolution covering visible light, near infrared, and short-wave infrared bands. In addition, a series of image data from the Gaofen satellite were also used, including GF1, GF1B, GF1C, GF1D, and GF2, with a resolution of 1 meter for the panchromatic band (band 1) and 4 meters for other multispectral bands (bands 2-5).

[0051] Subsequently, all satellite images were reprojected to the Universal Transverse Mercator (UTM) projection based on WGS84. Among them, Sentinel-2 images were used as input data for mapping forest fires, and high-resolution series images were used as reference data for forest fire accuracy assessment.

[0052] S1-2 satellite data preprocessing:

[0053] (1) The satellite images were geometrically orthorectified and georeferenced, and the interference of clouds, cloud shadows and snow was removed using the Fmask (Function ofmask) mask function. Subsequently, 100 control points were added to determine the spatial position correspondence between the Sentinel-2 images and the high-resolution series images, and a first-order polynomial transformation was selected.

[0054] (2) Calculate the Normalized Difference Vegetation Index (NDVI) to represent the forest growth status:

[0055]

[0056] Where NIR and RED are the surface reflectance values ​​in the red light band and near infrared band, respectively.

[0057] S2 Extraction of spatial distribution information of forest fire burning area:

[0058] S2-1 reduces image noise:

[0059] We manually extracted spectral features of 300 sample points of bare land (category A), forest (category B), and forest fire burn area (category C) from Sentinel-2 images. Then, using the random forest model, 30% of the 900 sample points were randomly selected as training samples and 70% as test samples. Finally, the trained random forest model was used to traverse the Sentinel-2 images to reduce image noise.

[0060] S2-2 Probability calculation of forest fire burning area

[0061] For each image pixel pi, its probability of being in class C (i.e., forest fire burn area) is predicted based on the distance weighted function of its j=1, 2, ..., J neighboring pixels:

[0062]

[0063] in is the probability that the image pixel pi is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire area, λ ij is the distance-related weight:

[0064]

[0065] In the formula, h ij is the distance between pixel i and pixel j, and α is the nonlinear parameter of the exponential model. According to the reference of the empirical value, α is 3 according to the experience.

[0066] S2-3 Spatial range identification of forest fire burning area

[0067] The probability of forest fire burning area for each image pixel is arranged in descending order and then classified. The classification standard is identified by threshold classification method:

[0068]

[0069] Among them, P is the classification threshold, P max and P min are the maximum and minimum probability values ​​of the pixels in the forest fire burning area, respectively, and d B (x,y) is the distance from pixel x to the nearest forest fire burn area sample point y;

[0070] σ p It is generally set to 10 kilometers. When the image pixel probability is ≥ P, it is set as a forest fire burning area; when the image pixel probability is < P, it is set as a non-forest fire burning area; some unburned areas in the forest fire burning area (such as exposed rocks, etc.) are excluded to identify closed forest fire burning areas:

[0071]

[0072] in is the probability that the image pixel pj is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire burn area, and A is the probability;

[0073] S3 accuracy verification:

[0074] S3-1 accuracy verification indicators:

[0075] Three widely used indicators are used to evaluate accuracy, including overall accuracy, root mean square error, mean absolute error, and systematic error. Among them, overall accuracy, root mean square error, and mean absolute error are used to calculate accuracy, while systematic error is used to quantify estimation bias. The calculation formula is as follows:

[0076] OA=A m / A;

[0077]

[0078] Where OA is the overall accuracy, A m is the area of ​​forest fire burnt area identified, A is the area of ​​forest fire burnt area actually reported, is the forest fire burning area of ​​sample i, f i is the forest fire burn area of ​​sample i manually interpreted, n is the total number of sample polygons, RMSE is the root mean square error, MAE is the mean absolute error, and SE is the systematic error.

[0079] S3-2 accuracy verification results:

[0080] This embodiment uses 6 fire remains published by the Hunan Provincial Emergency Management Department as an example for verification (see the table below), and it is found that the error of the method of the present invention is within 5%, and the accuracy is higher than 92%, showing high applicability:

[0081]

[0082] The above are only preferred embodiments of the present invention, which are only used to help understand the method and core idea of ​​the present application. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the idea of ​​the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

[0083] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0084] The present invention solves the problem of lack of effective and accurate means for monitoring forest fire burning areas in the prior art. By predicting forest burning areas through machine learning algorithms such as distance weighted functions of image pixels and threshold classification methods to predict the burning spatial range, and combining with accuracy verification, the burning areas of forest fires can be effectively monitored and predicted, which provides a basis for rapid and large-scale monitoring of forest fires, and also provides strong support for subsequent post-disaster management and standard setting.

Claims

1. A high-resolution monitoring method for forest fire burning areas, characterized in that: The following steps are involved: S1, Satellite data processing: S1-1, data download: S1-1-1, download satellite image data of 13 spectral bands of the Sentinel-2 multispectral instrument, with spatial resolution covering the visible, near-infrared and short-wave infrared bands; S1-1-2, download the series of image data from the Gaofen satellite, the resolution of the panchromatic band is 1m, and the resolution of other multispectral bands is 4m; S1-1-3, all the above satellite image data are reprojected onto the Universal Transverse Mercator UTM projection based on WGS84, wherein the image data in S1-1-1 is used as input data for mapping forest fires, and the image data in S1-1-2 is used as reference data for evaluating the accuracy of forest fires; S1-2, data preprocessing: S1-2-1, the satellite images were geometrically orthorectified and georeferenced, and the interference of clouds, cloud shadows and snow was removed using the Fmask mask function; then 100 control points were added to determine the spatial position correspondence between the Sentinel-2 images and the Gaofen satellite series images, and a first-order polynomial transformation was selected; S1-2-2, calculate the normalized difference vegetation index NDVI to represent the forest growth status, as shown in the following formula: Among them, NIR and RED are the surface reflectance values ​​in the infrared band and near infrared band respectively; S2, extraction of spatial distribution information of forest fire burning areas: S2-1, reduce image noise: The spectral features of 300 sample points of bare land category A, forest category B, and forest fire burn area category C were manually extracted from Sentinel-2 images. Using the random forest model, 30% of the 900 sample points were randomly selected as training samples and 70% as test samples. The trained random forest model was used to traverse the Sentinel-2 images to reduce image noise. S2-2, Calculation of forest fire burning area probability: For each pixel pi of the image, its probability of being in class C, i.e., the forest fire burn area, is predicted based on the distance weighted function of its j=1, 2, ..., J neighboring pixels, as shown in Equation 2: in is the probability that the image pixel pi is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire area, λ ij is the weight related to the distance, as shown in Formula 3: where h ij is the distance between pixel i and pixel j, α is the nonlinear parameter of the exponential model; S2-3, Identification of the spatial extent of forest fire burning areas: The probability of forest fire burning area for each image pixel is arranged in descending order and then classified. The classification standard is identified by threshold classification method, as shown in the following formula 4: Among them, P is the classification threshold, P max and P min are the maximum and minimum probability values ​​of the pixels in the forest fire burning area, respectively, and d B (x, y) is the distance from pixel x to the nearest forest fire burn area sample point y; when the image pixel probability ≥ P, it is set as a forest fire burn area; when the image pixel probability < P, it is set as a non-forest fire burn area; The unburned areas in the forest fire burning area are excluded to identify the closed forest fire burning area, as shown in the following formula 5: in is the probability that the image pixel pj is a forest fire burning area, is the jth pixel belonging to the forest fire burn area, is the value of the jth pixel belonging to the forest fire burn area, and A is the probability; S3, accuracy verification: The accuracy is evaluated using overall accuracy, root mean square error, mean absolute error, and systematic error, where overall accuracy, root mean square error, and mean absolute error are used to calculate accuracy, while systematic error is used to quantify the estimation bias, as shown in Equations 6 to 8: OA=A m / A; Where OA is the overall accuracy, A m is the area of ​​forest fire burnt area identified, A is the area of ​​forest fire burnt area actually reported, is the forest fire burning area of ​​sample i, f i is the forest fire burn area of ​​sample i manually interpreted, n is the total number of sample polygons, RMSE is the root mean square error, MAE is the mean absolute error, and SE is the systematic error.

2. The high-resolution monitoring method for forest fire burning areas according to claim 1 is characterized in that: The 13 spectral bands in the S1-1-1 are 4 10-meter bands, 6 20-meter bands and 3 60-meter bands.

3. The high-resolution monitoring method for forest fire burning areas according to claim 1 is characterized in that: The panchromatic band in S1-1-2 is band 1, and the other multispectral bands are bands 2 to 5.

4. The high-resolution monitoring method for forest fire burning areas according to claim 1 is characterized in that: The empirical value of α in the formula (3) is 3.

5. The high-resolution monitoring method for forest fire burning areas according to claim 1 is characterized in that: The σ p It is 10 kilometers.

Images