Automatic rapid fire point position correction method based on image frequency domain information

Abstract

The invention discloses an automatic rapid fire point position correction method based on image frequency domain information, and belongs to the field of image processing. According to the method, GSHHG coastline and inland water body data provided by NOAA are taken as geometric reference data, and geometric correction is carried out on the position offset of a fire point by utilizing frequency domain information of a specific wave band of Himawari image data. Firstly, after screening of available control points is completed, the POC algorithm is used for calculating the offset of windows around the control points. Based on the distance, three control points are selected, the average value of the offset of the three control points serves as the offset of the pixel to be corrected, and the position of the fire point pixel is corrected. According to the method, the geometric position deviation problem of the Himawari-8 / 9 image is corrected, and the fire point position deviation problem is effectively relieved. According to the invention, the precision of fire point detection in southwest China is effectively improved.

Description

Technical Field

[0001] Aiming at the problem of position offset in Himawari fire point detection results, the present invention proposes an automatic and rapid fire point position correction algorithm based on image frequency domain information. Firstly, the position offset of each control point is calculated using the POC algorithm, and the position offset of the fire point is obtained by calculating the average offset of the three nearest control points for the positions of different fire points in the study area, thereby realizing the correction of the fire point position offset. Background Art

[0002] Existing geometric correction methods mainly include traditional correction algorithms, correction algorithms based on spatial information, and correction algorithms based on frequency domain information. Traditional manual geometric correction methods use polynomial fitting after constructing line-of-sight equations and lookup tables. These algorithms not only have a large room for improvement in correction accuracy, but also require manual selection of control points for matching, which makes it difficult to achieve automation. In addition, affected by the imaging principle of the sensor, the Himawari-8 / 9 image data has different degrees of geometric distortion, which leads to a certain offset in the position of the identified fire point. Since the southwest region of my country is located at the edge of the Himawari full-disk image imaging area, the geometric distortion is more obvious. The study area of ​​this paper is a high-incidence area of ​​fire. Affected by natural factors such as complex and changeable terrain conditions, the positioning offset of the fire point pixel will bring great difficulties to the subsequent firefighting work. Therefore, developing a fast automatic correction algorithm for fire points for Himawari-8 / 9 data is of great significance for fire fighting and prevention and control. Summary of the invention

[0003] The invention solves the problem of positional deviation in Himawari fire point detection results.

[0004] The technical solution is an automatic rapid fire point position correction method based on image frequency domain information, which includes:

[0005] Step 1: Preprocess the coastline and inland water data. The two-dimensional Fourier transform formulas corresponding to the preprocessed data are as follows:

[0006]

[0007] Where F(u,v) and G(u,v) are the frequency domain images corresponding to the GSHHG reference water body data in the 6-band data domain of the Himawari image obtained after periodic extension and discrete Fourier transform; f(x,y) represents the 6-band data of the Himawari image, N represents the image as an N×N square matrix, g(x,y) represents the GSHHG reference water body data, (u,v) represents the index in the frequency domain, corresponding to the frequency components in the horizontal and vertical directions, (x,y) represents the position index in the spatial domain, ranging from 0 to N-1, and i represents the imaginary unit;

[0008] Step 2: To eliminate the noise in the Fourier transformed image, use the Hamming window function to filter it. The Hamming window is defined as follows:

[0009] w(x,y)=w(x)*w(y) (3)

[0010]

[0011] Where a0 is the proportional coefficient, and w(x,y) is the two-dimensional Hamming window function;

[0012] Step 3: After obtaining the filtered frequency domain image data, calculate its normalized cross-power spectrum function R(u,v) as shown in formula (6);

[0013]

[0014] in is the complex conjugate function of G(u,v). To obtain the offset information of the image, perform inverse Fourier transform on it, as shown in formula (7);

[0015]

[0016] Where r(x,y) is an impulse function, which represents the offset between the Himawari image to be corrected and the GSHHG data in the selected window; the window correction is performed according to the offset;

[0017] Step 4: Correct the fire point position.

[0018] Furthermore, the specific method of step 4 is:

[0019] Step 4.1: Manually select control points, and select inland lakes that can be identified at a spatial resolution of 2 km in band 6 of the Himawari image data as control points;

[0020] Step 4.2: Using the Himawari-8 / 9 image of the phase to be corrected as the source data, set a window of size 81×81 around the control point; after completing the setting and selection of the window, perform cloud mask processing on it;

[0021] Finally, use the following formula to filter available control points:

[0022]

[0023] Where amount_cloud and amount_total are the number of cloud pixels in the window and the total number of pixels in the window respectively, and criteria is a fixed threshold;

[0024] After cloud mask screening, the offset of each control point is calculated using a window registration method based on the POC algorithm; if the row offset or column offset of a control point exceeds 5, the point will be removed from the set; all remaining available control points are recorded as matrix OFF, which stores the position information and row and column offset information of each point.

[0025] Step 4.3: After obtaining the offset information matrix OFF, the offset of each fire point pixel is calculated based on the distance information; first, the distance between two points is defined based on the longitude and latitude information as shown in formula (9);

[0026]

[0027] where lon fp With lon cp are the longitudes of the fire point pixel and the control point pixel, lat fp With lat cp are the latitudes of the fire point pixel and the control point pixel respectively; by calculating the distance between the fire point to be corrected and all the control points in the set OFF, find out the three control points with the smallest distance, which are recorded as p1, p2 and p3 respectively; if the number of available control points is less than three, in order to avoid wrong results, choose not to correct the fire point in the scene image;

[0028] Step 4.4: After obtaining the three nearest control points, use equations (10) and (11) to calculate the row offset and column offset of the fire point pixel at the position to be corrected;

[0029]

[0030] Among them, loff and coff are the row offset and column offset of the fire point pixel to be corrected. i With coff i are the row offset and column offset of the i-th control point itself;

[0031] Step 4.4: The corrected longitude and latitude coordinates of the fire point pixel can be obtained by using equations (12) and (13);

[0032] lat real =lat ori -0.02*loff (12)

[0033] lon real =lon ori +0.02*coff (13)

[0034] Among them, lat ori With lon ori are the original latitude and longitude of the fire point to be corrected, lat real With lon real They are the latitude and longitude of the fire point to be corrected after correction.

[0035] Furthermore, the preprocessing method in step 1 is:

[0036] The Himawari image data was masked with water bodies. The water bodies were extracted by combining the vegetation index information with the strong absorption of water bodies near the 2.3μm band. The standards are as follows:

[0037]

[0038] a6<0.05 and a4<0.15 and NDVI<0 (2)

[0039] Where a6 is the reflectance of band 6 of Himawari-8 / 9 data, a3 and a4 are the reflectance of band 3 and band 4 of Himawari-8 / 9 data respectively, NDVI is the normalized difference vegetation index calculated based on Himawari-8 / 9 image data, and the pixels satisfying formula (2) are marked as water pixels;

[0040] The water mask result is filled with binary values ​​of 0 and 1, and the binary image obtained is used as the final result after preprocessing and input into the POC algorithm for registration;

[0041] For the GSHHG data, a raster image with a spatial resolution of 2 km is obtained through preprocessing, and the water and non-water parts are filled with binary values ​​of 0 and 1.

[0042] Furthermore, in step 4.2, criterion is a fixed threshold value, which is set to 0.3.

[0043] Furthermore, in step 2, a0=0.53836.

[0044] Beneficial effects of the present invention: Based on the frequency domain information of band 6 of Himawari-8 / 9 image data, the present invention proposes an automatic geometric correction algorithm for correcting the position offset of Himawari fire point detection results. First, the coastline and inland water body data are preprocessed, and the offsets of all selected control points are calculated based on the POC algorithm. For fire points in different locations, the distance between them and the control points is used as the selection criterion, and the position offset of the fire point is obtained by calculating the offset values ​​of the three nearest control points. In most cases, the algorithm can effectively realize the position correction of the Himawari fire point detection results and effectively improve the positioning accuracy of the fire point pixels. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 : Schematic diagram of the location and topography of the southwest study area, which marks the distribution location of the southwest study area, the topography of the study area, and the topography of the study area.

[0046] Figure 2 :GSHHG data L1, L2 layer example map. The spatial reference data selected by the present invention is the global self-contained hierarchical high-resolution geographic database (GSHHG) data provided by the National Oceanic and Atmospheric Administration (NOAA) of the United States. This is a set of hierarchical global high-resolution coastline data. The L1 and L2 layers are continental land and oceanic islands (except Antarctica) and lakes respectively.

[0047] Figure 3 :POC algorithm flow chart. The POC algorithm is an image registration algorithm based on image frequency domain information, and its core is the two-dimensional fast Fourier transform (FFT) of the image. The figure shows the specific process of the POC algorithm.

[0048] Figure 4 :An example of the POC algorithm running results. The results obtained by the POC algorithm are presented from multiple different angles.

[0049] Figure 5 : Schematic diagram of control point distribution, especially the partial display of inland water control points, the base map is a high-resolution remote sensing image.

[0050] Figure 6 :Example diagrams of window correction effects in the Qinghai Lake area at different phases. The upper figure is a schematic diagram of window correction in the Qinghai Lake area on March 17, 2021, and the lower figure is a display of the correction effect on January 24, 2023. The overlap between the Qinghai Lake area on the Himawari-8 / 9 image and the Qinghai Lake outline provided by the GSHHG inland water data before and after correction is shown.

[0051] Figure 7: An example of the geometric correction results of the fire point location is shown, with the base map being a high-resolution remote sensing image. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] 1: Window correction based on POC algorithm:

[0054] Flowchart as Figure 3 shown.

[0055] Before geometric correction of the fire point position, the POC algorithm must be used to calculate the offset of each control point. The data required are the 6-band data of the Himawari image and the GSHHG reference water data, which are recorded as f(x, y) and g(x, y) respectively. Before POC registration, the two types of data are preprocessed. The Himawari image data is masked with water. The present invention uses the strong absorption of water near the 2.3μm band and combines vegetation index information to extract water. The standards are as follows:

[0056]

[0057] a6<0.05 and a4<0.15 and NDVI<0 (2)

[0058] Where a6 is the reflectance of band 6 of the Himawari-8 / 9 data, and NDVI is the normalized difference vegetation index calculated based on the Himawari-8 / 9 image data. Pixels that satisfy formula (2) are marked as water pixels.

[0059] The water mask result is filled with binary values ​​of 0 and 1, and the binary image is used as the final result after preprocessing and input into the POC algorithm for registration. For the GSHHG data, the raster image with a spatial resolution of 2 km is obtained through preprocessing, and the water and non-water parts are also filled with binary values ​​of 0 and 1.

[0060] The two-dimensional Fourier transforms corresponding to the two types of preprocessed data are shown in equations (3) and (4) respectively.

[0061]

[0062] Where F(u,v) and G(u,v) are the frequency domain images obtained after periodic extension and discrete Fourier transform of the two images. There is some high-frequency noise in the frequency domain image obtained by transform. In order to eliminate this noise, the frequency domain information of the obtained image is filtered using the Hamming window function. The definition of the Hamming window is as shown in equations (5)-(7).

[0063] w(x,y)=w(x)*w(y) (5)

[0064]

[0065] The proportional coefficient a0 = 0.53836. w(x, y) is the two-dimensional Hamming window function. After obtaining the filtered frequency domain image data, the normalized cross-power spectrum function R(u, v) is calculated as shown in formula (8).

[0066]

[0067] in is the complex conjugate function of G(u,v). To obtain the offset information of the image, perform inverse Fourier transform on it, as shown in formula (9).

[0068]

[0069] Where r(x, y) should ideally be an impulse function, which represents the offset between the Himawari image to be corrected and the GSHHG data in the selected window.

[0070] 2: Fire point position correction based on control points:

[0071] Based on the selected control points of some coastlines, the present invention adds some control points located inside the study area. Some inland lakes that can be identified at the 2km spatial resolution of band 6 of the Himawari image data are selected as control points. Similarly, some coastline points that are close to the study area and less obscured by clouds are selected as supplements. The location information of the control points is shown in Table 1 and Figure 5 shown.

[0072] Table 1

[0073]

[0074]

[0075] Before correcting the offset of the fire point position, the control points in Table 1 need to be screened to obtain available control points.

[0076] First, the Himawari-8 / 9 image of the phase to be corrected is used as the source data, and a window of size 81×81 is set around all the control points in Table 1. The window set here is also the window used for matching when calculating the offset using the POC algorithm later.

[0077] After completing the window setting and selection, it needs to be subjected to cloud masking. After completing the cloud masking process, the available control points are screened using formula (10).

[0078]

[0079] Wherein amount_cloud and amount_total are the number of cloud pixels in the window and the total number of pixels in the window respectively, and the left side of inequality (10) is the ratio of cloud pixels in the window. criteria is a fixed threshold, which represents the acceptable degree of cloud contamination of the window around the control point. In the present invention, this threshold is set to 0.3. All pixels that do not satisfy equation (10) will be considered to be blocked by clouds and removed from the set of available control points.

[0080] After cloud mask screening, the offset of each control point is calculated using the window registration method based on the POC algorithm described in 1. If the row offset or column offset of a control point exceeds 5, the point will be removed from the set. All remaining available control points are recorded as matrix OFF, which stores the position information and row and column offset information of each point.

[0081] After obtaining the offset information matrix OFF, the offset of each fire point pixel is calculated based on the distance information. First, the distance between two points is defined based on the longitude and latitude information as shown in formula (11).

[0082]

[0083] where lon fp With lon cp are the longitudes of the fire point pixel and the control point pixel, lat fp With lat cp are the latitudes of the fire point pixel and the control point pixel, respectively. By calculating the distance between the fire point to be corrected and all the control points in the set OFF, the three control points with the smallest distance are found, which are recorded as p1, p2 and p3. It should be pointed out that if the number of available control points is less than three, it means that the basic information used to calculate the offset is seriously insufficient, and it is impossible to effectively simulate the overall distortion trend of the image. At this time, in order to avoid erroneous results, it is chosen not to correct the fire point in the image.

[0084] After obtaining the three nearest control points, use equations (12) and (13) to calculate the row offset and column offset of the fire point pixel at the correction position.

[0085]

[0086] Where loff and coff are the row offset and column offset of the fire point pixel to be corrected. i With coff i are the row offset and column offset of the ith control point itself. Then, the corrected longitude and latitude coordinates of the fire point pixel can be obtained through equations (14) and (15).

[0087] lat real =lat ori -0.02*loff (14)

[0088] lon real =lon ori +0.02*coff (15)

[0089] Among them, lat ori With lon ori are the original latitude and longitude of the fire point to be corrected, lat real With lon real They are the latitude and longitude of the fire point to be corrected after correction.

Claims

1. An automatic rapid fire point position correction method based on image frequency domain information, the method comprising: Step 1: Preprocess the coastline and inland water data. The two-dimensional Fourier transform formulas corresponding to the preprocessed data are as follows: Where F(u,v) and G(u,v) are the frequency domain images corresponding to the GSHHG reference water body data in the 6-band data domain of the Himawari image obtained after periodic extension and discrete Fourier transform; f(x,y) represents the 6-band data of the Himawari image, N represents the image as an N×N square matrix, g(x,y) represents the GSHHG reference water body data, (u,v) represents the index in the frequency domain, corresponding to the frequency components in the horizontal and vertical directions, (x,y) represents the position index in the spatial domain, ranging from 0 to N-1, and i represents the imaginary unit; Step 2: To eliminate the noise in the Fourier transformed image, use the Hamming window function to filter it. The Hamming window is defined as follows: w(x,y)=w(x)*w(y) (3) Where a0 is the proportional coefficient, and w(x,y) is the two-dimensional Hamming window function; Step 3: After obtaining the filtered frequency domain image data, calculate its normalized cross-power spectrum function R(u,v) as shown in formula (6); in is the complex conjugate function of G(u,v). To obtain the offset information of the image, perform inverse Fourier transform on it, as shown in formula (7); Where r(x,y) is an impulse function, which represents the offset between the Himawari image to be corrected and the GSHHG data in the selected window; the window correction is performed according to the offset; Step 4: Correct the fire point position.

2. The automatic rapid fire point position correction method based on image frequency domain information as claimed in claim 1, characterized in that: The specific method of step 4 is: Step 4.1: Manually select control points, and select inland lakes that can be identified at a spatial resolution of 2 km in band 6 of the Himawari image data as control points; Step 4.2: Using the Himawari-8 / 9 image of the phase to be corrected as the source data, set a window of size 81×81 around the control point; after completing the setting and selection of the window, perform cloud mask processing on it; Finally, use the following formula to filter available control points: Where amount_cloud and amount_total are the number of cloud pixels in the window and the total number of pixels in the window respectively, and criteria is a fixed threshold; After cloud mask screening, the offset of each control point is calculated using a window registration method based on the POC algorithm; if the row offset or column offset of a control point exceeds 5, the point will be removed from the set; all remaining available control points are recorded as matrix OFF, which stores the position information and row and column offset information of each point. Step 4.3: After obtaining the offset information matrix OFF, the offset of each fire point pixel is calculated based on the distance information; first, the distance between two points is defined based on the longitude and latitude information as shown in formula (9); where lon fp With lon cp are the longitudes of the fire point pixel and the control point pixel, lat fp With lat cp are the latitudes of the fire point pixel and the control point pixel respectively; by calculating the distance between the fire point to be corrected and all the control points in the set OFF, find out the three control points with the smallest distance, which are recorded as p1, p2 and p3 respectively; if the number of available control points is less than three, in order to avoid wrong results, choose not to correct the fire point in the scene image; Step 4.4: After obtaining the three nearest control points, use equations (10) and (11) to calculate the row offset and column offset of the fire point pixel at the position to be corrected; Among them, loff and coff are the row offset and column offset of the fire point pixel to be corrected. i With coff i are the row offset and column offset of the i-th control point itself; Step 4.4: The corrected longitude and latitude coordinates of the fire point pixel can be obtained by using equations (12) and (13); years real =years ori -0.02*loff (12) lon real = happy ori +0.02*memory (13) Among them, lat ori With lon ori are the original latitude and longitude of the fire point to be corrected, lat real With lon real They are the latitude and longitude of the fire point to be corrected after correction.

3. The automatic rapid fire point position correction method based on image frequency domain information as claimed in claim 1, characterized in that: The preprocessing method in step 1 is: The Himawari image data was masked with water bodies. The water bodies were extracted by combining the vegetation index information with the strong absorption of water bodies near the 2.3μm band. The standards are as follows: a6<0.05 and a4<0.15 and NDVI<0 (2) Where a6 is the reflectance of band 6 of Himawari-8 / 9 data, a3 and a4 are the reflectance of band 3 and band 4 of Himawari-8 / 9 data respectively, NDVI is the normalized difference vegetation index calculated based on Himawari-8 / 9 image data, and the pixels satisfying formula (2) are marked as water pixels; The water mask result is filled with binary values ​​of 0 and 1, and the binary image obtained is used as the final result after preprocessing and input into the POC algorithm for registration; For the GSHHG data, a raster image with a spatial resolution of 2 km is obtained through preprocessing, and the water and non-water parts are filled with binary values ​​of 0 and 1.

4. The automatic rapid fire point position correction method based on image frequency domain information as claimed in claim 1, characterized in that: In step 4.2, criterion is a fixed threshold value, which is set to 0.

3.

5. The automatic rapid fire point position correction method based on image frequency domain information as claimed in claim 1, characterized in that: In step 2, a0=0.53836.

Images