A method for early fire source positioning based on multispectral image features

Abstract

The application discloses a kind of early fire source positioning methods based on multispectral image features.The method includes: obtaining multispectral remote sensing image containing blue light band and identifying smoke mask;Determine the minimum area envelope ellipse of the smoke mask, and take its long axis as the diffusion main shaft;Calculate and compare the spectral mean of the two end regions of diffusion main shaft in blue light band, and determine the larger end as the initial fire point, and the smaller end as the initial plume tail;Based on initial fire point and initial plume tail, extract the real boundary features of smoke mask, and carry out position correction, determine the real fire point and smoke diffusion direction.The application realizes high-precision fire field source positioning and wind direction estimation without thermal infrared data.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing monitoring technology for forest and grassland fires, specifically involving an early fire source location method based on multispectral image features. Background Technology

[0002] Forest and grassland fires pose a significant threat to the ecological environment and the safety of life and property; early detection and early suppression are crucial for prevention and control. Satellite remote sensing technology is currently an effective means of large-scale monitoring of forest and grassland fires. When a fire occurs, the accompanying fire points and smoke will exhibit significant spectral differences from background features in satellite remote sensing images. In traditional monitoring, thermal infrared imagery from satellites is typically relied upon to identify high-temperature fire points and determine the location of the fire. Simultaneously, in medium- to high-spatial-resolution spectral images, smoke serves as another important indicator signal; its diffusion patterns and concentration distribution also contain rich information about the fire scene.

[0003] Existing technologies typically rely on thermal infrared imagery to identify fire points and determine their location, thereby identifying real-time information about the fire, such as the fire's head, tail, and wings, and ultimately determining whether a fire has occurred and its development. However, not all optical satellites are equipped with thermal infrared sensors, and thermal infrared imagery has relatively low spatial resolution, making it prone to missing or misidentifying fire points.

[0004] When a fire occurs, smoke often arises earlier than the ignition point and flows freely, penetrating the forest canopy unobstructed and facilitating observation. This is beneficial for detecting early small fires and smoldering fires. Therefore, how to deeply explore and utilize the key real-time fire signal of forest and grassland fire smoke, and achieve accurate estimation of the location of the fire ignition point and the real-time wind direction in the absence of external dynamic data support under single multispectral images, has become an urgent technical problem to be solved in the field of forest and grassland fire monitoring and early warning. Summary of the Invention

[0005] The purpose of this invention is to provide an early fire source location method based on multispectral image features, aiming to solve the technical problem of using forest and grassland fire smoke to locate fire sources and determine the direction of smoke diffusion in existing technologies. By using a technical solution of minimum area envelope ellipse orientation, end band mean polarization, and pixel-level correction by extracting real boundary features, the invention achieves the technical effect of accurately locking the real ignition point pixel and real-time wind direction in high-resolution multispectral images without the need for thermal infrared data.

[0006] To address the above problems, this invention provides an early fire source localization method based on multispectral image features, comprising the following steps: Acquire multispectral remote sensing images of the area to be monitored and identify the smoke mask from them, wherein the multispectral remote sensing images contain at least the blue light band; Based on the pixel distribution characteristics of the flue gas mask, the minimum area envelope ellipse of the flue gas mask is determined, and the major axis of the minimum area envelope ellipse is taken as the diffusion principal axis of the flue gas mask. Calculate and compare the spectral mean values ​​of the two end regions of the diffusion axis in the blue light band, determine the end with the larger spectral mean value as the initial fire source, and determine the end with the smaller spectral mean value as the initial plume tail. Within the monitored area, the true boundary features of the smoke mask are extracted based on the initial fire source and the tail of the initial smoke plume. The position of the initial fire source and the tail of the initial smoke plume is corrected according to the true boundary features to determine the true fire source and the direction of smoke diffusion.

[0007] Furthermore, in the above method, extracting the true boundary features of the flue gas mask includes: Based on the initial fire source and the initial plume tail, a first vector is constructed along the main diffusion axis, and a second vector is orthogonal to the first vector. Calculate the projections of each pixel of the smoke mask onto the first vector and the second vector.

[0008] Furthermore, in the above method, the step of correcting the position of the initial ignition point and the initial plume tail based on the true boundary features to determine the true ignition point and the direction of smoke diffusion includes: Among all pixels of the smoke mask, determine the extreme values ​​of the projection on the first vector and obtain the endpoints of the scanning interval; The tangent is translated along the first vector to the endpoint of the scanning interval. In the candidate pixel set intercepted by the tangent, the median point of the projection on the second vector is selected to obtain the corrected real fire source point and real smoke plume tail.

[0009] Further, in the above method, determining the minimum area envelope ellipse of the flue gas mask includes: A shape matrix is ​​constructed using the minimum area envelope ellipse fitting method, and an objective function is constructed using the determinant of the shape matrix. Solve for the optimal solution of the objective function to obtain the optimal shape matrix of all pixels enclosing the smoke mask, and determine the center of the minimum area envelope ellipse, the major axis and the minor axis based on the optimal shape matrix.

[0010] Furthermore, in the above method, solving for the optimal solution of the objective function includes: The cell coordinates of the flue gas mask are converted into homogeneous coordinates, and an initial weight is assigned to each cell of the flue gas mask to construct a weight vector; The default degree of each cell is calculated based on the current weight vector; The step size parameter is updated based on the maximum default degree, and the weight vector is iteratively updated until the convergence condition is met, thus obtaining the optimal shape matrix.

[0011] Furthermore, in the above method, the two end regions of the diffusion main shaft are divided in the following manner: Calculate the normalized projected coordinates of the pixels of the flue gas mask along the main diffusion axis; Obtain a preset end-cut ratio, and set the cut-off intervals corresponding to both ends of the diffusion main shaft based on the end-cut ratio; The pixels falling within the intercepted interval in the normalized projected coordinates are divided into two sets of end pixels, which serve as the two end regions of the diffusion principal axis.

[0012] Furthermore, in the above method, acquiring the multispectral remote sensing image of the area to be monitored includes: The multispectral remote sensing image is converted into multiband apparent reflectance data; The multispectral remote sensing image includes at least the green light band and the red light band.

[0013] Furthermore, in the above method, determining the direction of flue gas diffusion includes: Obtain the two-dimensional coordinates of the actual fire source and the actual plume tail; Based on the two-dimensional coordinates, the azimuth angle is calculated using the arctangent function, with true north as 0 degrees and clockwise normalized to the range of 0 to 360 degrees, and the azimuth angle is used as the direction of flue gas diffusion.

[0014] The core principle of this invention is as follows: First, by using a minimum area envelope ellipse, the highly irregularly shaped smoke mask is abstracted into a standard geometric shape, and the major axis is extracted as the diffusion principal axis to quantify the macroscopic smoke diffusion trend. Second, due to the symmetry of the pure geometric envelope, this invention cleverly utilizes the physical characteristic that the blue light band is highly sensitive to smoke concentration. By comparing only the average blue light spectrum at both ends (i.e., the end regions) of the diffusion principal axis, the symmetry is broken, giving the diffusion principal axis vector polarity and determining the initial fire source point. Finally, addressing the physical defect that the endpoints of the geometric envelope are easily detached from the actual smoke pixels, the actual boundary features are extracted based on the initial poles (by constructing an axial and normal orthogonal vector system and tangent translation scanning, etc.), pulling the suspended theoretical poles back to the outermost pixels of the actual smoke, thereby achieving precise positioning.

[0015] The above-described technical solution of the present invention has the following beneficial technical effects: (1) This invention corrects the position of the initial fire source by extracting real boundary features, effectively eliminating the phenomenon of theoretical poles being suspended due to edge expansion in conventional geometric envelope models, and forcing the algorithm to converge to the actual existing smoke boundary pixels, which greatly improves the absolute spatial accuracy of fire source positioning.

[0016] (2) When performing position correction, the boundary feature extraction mechanism (such as the selection logic of translation tangent combined with the median of normal projection in the embodiment) can act as a highly fault-tolerant edge filter, which perfectly eliminates the interference of smoke edge burrs, shape bifurcation or free noise (enclave pixels) caused by wind on the positioning coordinates, ensuring that the corrected position is the real fire source point.

[0017] (3) This scheme completely eliminates the dependence of traditional methods on thermal infrared temperature data. It only requires multispectral images provided by conventional optical satellites to achieve high-precision estimation of real-time wind direction and fire source through dual mining of geometric morphology and optical reflectivity characteristics. This not only solves the problem of missed and misidentified fires caused by the low spatial resolution of thermal infrared images, but also greatly expands the data source acquisition channels and applicable scenarios for high-precision monitoring of forest and grassland fires in the early stage. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method steps according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the flue gas mask extracted according to an embodiment of the present invention; Figure 3 This is a schematic diagram of minimum area envelope ellipse fitting and diffusion principal axis extraction for flue gas mask in an embodiment of the present invention; Figure 4 This is a schematic diagram of the initial ignition point, the initial plume tail, and the initial diffusion direction in an embodiment of the present invention; Figure 5 This is a schematic diagram of the actual fire source and the actual plume tail in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0020] refer to Figure 1 The flowchart illustrates the steps of an early fire source localization method based on multispectral image features according to this embodiment, specifically including the following steps: Step S100: Acquire multispectral remote sensing images of the area to be monitored and identify the smoke mask from them, wherein the multispectral remote sensing images contain at least the blue light band.

[0021] To ensure the availability of the basic band information required for smoke mask identification, it is necessary to restore the physical meaning of the original satellite payload data through data format conversion steps and spectral band definition features. In another embodiment of the invention, acquiring multispectral remote sensing images of the area to be monitored includes: converting the multispectral remote sensing images into multiband apparent reflectance data; wherein the multispectral remote sensing images contain at least green and red light bands.

[0022] Step S200: Based on the pixel distribution characteristics of the flue gas mask, determine the minimum area envelope ellipse of the flue gas mask, and take the major axis of the minimum area envelope ellipse as the diffusion principal axis of the flue gas mask.

[0023] Determining the minimum area envelope ellipse of the flue gas mask typically involves fitting using the minimum area envelope method. In another embodiment of the invention, determining the minimum area envelope ellipse of the flue gas mask includes: constructing a shape matrix using the minimum area envelope ellipse fitting method, and constructing an objective function using the determinant of the shape matrix. The specific calculation process is as follows: (Refer to...) Figure 2 First, obtain the pixel coordinates of all pixels within the flue gas mask, as well as the shape matrix and shape center of the flue gas mask. Then, construct the set constraints of the flue gas mask. The specific calculation formula is as follows: (1) in, For the set constraints of the flue gas mask, The coordinates of the shape center of the flue gas mask are: For shape matrix, here ,Right now It is a positive definite matrix. For the first inside the flue gas mask The pixel coordinates of each pixel. These are cell index values; the cell coordinates here are in two dimensions. It is a two-dimensional real number space.

[0024] Based on geometric properties, the area of ​​this envelope ellipse is related to the shape matrix of the flue gas mask. The square root of the determinant is inversely proportional to the given set constraints. To find the geometric envelope with the minimum area that satisfies the above set constraints, the geometric problem of finding the minimum area ellipse is equivalently transformed into an algebraic convex optimization problem that maximizes the matrix determinant. Based on this, an objective function for iterative solution is constructed, calculated as follows: (2) After optimizing formula (2), we get the following formula: (3) in, Let be the area of ​​the enclosing ellipse. For determinant, As constraints, Constraints .

[0025] Solve for the optimal solution of the objective function to obtain the optimal shape matrix of all pixels enclosing the smoke mask, and determine the center of the minimum area envelope ellipse, the major axis and the minor axis based on the optimal shape matrix.

[0026] Since the smoke mask acquired from remote sensing imagery contains a massive number of pixels, the traditional simple bounding rectangle method cannot accurately reflect the oblique diffusion characteristics of the smoke. To obtain the globally optimal envelope shape, this embodiment employs the minimum area envelope ellipse fitting method. First, a shape matrix is ​​constructed for the smoke mask. Since the area of ​​an ellipse can be mathematically proven to be proportional to a power of the determinant of the inverse of the shape matrix, the objective function is constructed using the determinant of the shape matrix. The task of finding the minimum area envelope ellipse is rigorously equivalent to finding the shape matrix that minimizes this objective function.

[0027] Subsequently, the optimal solution to the objective function is sought. During this process, the parameters of the shape matrix need to be continuously adjusted to continuously decrease the value of the objective function. Simultaneously, a rigid constraint must be satisfied: when the coordinates of all pixels in the flue gas mask are substituted into the shape matrix equation, the result must show that the pixel is inside or on the boundary of the ellipse. When a situation is found that satisfies all constraints and the objective function cannot be further reduced, the optimal shape matrix encompassing all pixels of the flue gas mask is obtained. Since the shape matrix itself contains all the geometric transformation information of the ellipse, eigenvalue decomposition or algebraic transformation is then performed on the optimal shape matrix to directly determine the coordinates of the center position of the minimum area enclosing ellipse, the major axis vector representing the diffusion direction, and the minor axis vector representing the diffusion width.

[0028] In another embodiment of the invention, solving for the optimal solution of the objective function includes: converting the cell coordinates of the flue gas mask into homogeneous coordinates, and assigning an initial weight to each cell of the flue gas mask to construct a weight vector; calculating the default degree of each cell based on the current weight vector; updating the step size parameter according to the maximum default degree, and iteratively updating the weight vector until the convergence condition is met to obtain the optimal shape matrix.

[0029] The pixel coordinates are homogenized to obtain the homogeneous vector of each pixel, calculated as follows: (4) in, For the first A homogeneous vector of pixels, It is a three-dimensional real number space.

[0030] Assign a non-negative real initial weight to each cell and construct a weight vector to obtain the homogeneous second-order matrix of the cell. The calculation formula is as follows: (5) in, Let be a homogeneous second-order matrix of pixels. Let be a weight vector, satisfying that all initial weights are non-negative and the sum of all initial weights is 1. This represents the total number of pixels in the smoke mask.

[0031] In each round of iterative optimization, based on the current weight vector and its corresponding homogeneous second-order matrix The default value of the homogeneous vector of a pixel in the set is calculated using the following formula: (6) in, For the first The default value of each pixel is used to measure how close each pixel is to the current ellipse, with the aim of selecting the largest value to update the corresponding weight vector and quickly approximate the minimum area envelope ellipse.

[0032] Next, an initial weight allocation strategy is set. In the initial stage of iteration, each cell is assigned equal geometric importance, that is, the weight vector is initialized as follows: After starting the iterative calculation, find the cell with the largest default value in each iteration, i.e., the [cell name missing]. In the next iteration Pixel default rate Maximum, set the stopping criterion for iteration, and obtain the spatial dimension of homogeneous coordinates. In this embodiment, it is a three-dimensional space, therefore Set tolerance threshold Determine whether the stopping criteria are met: If the condition is not met, continue to the next round of iteration calculation; if the condition is met, it means that all pixels have been perfectly enveloped by the current ellipse and the error is within the allowable range, and stop the iteration calculation.

[0033] If it is necessary to proceed to the next round of iteration calculation, the maximum default degree should be found based on the criteria. and spatial dimensions Calculate the step size parameter used to adjust the weights. The formula for updating the step size parameter is as follows: (7) The weight vector is iteratively updated, with the update criterion being: for the farthest... For a given pixel, its weight is increased; for the remaining pixels, their weights are decreased. The corresponding homogeneous second-order matrix is ​​updated synchronously. The formula for weight vector update is: (8) in, The farthest pixel in the current iteration. The standard basis vectors.

[0034] After the iterative calculations are complete, the ellipse can be fitted to obtain the final converged optimal weight vector. According to the optimal weight vector Calculate the geometric center coordinates of the minimum area envelope ellipse, which is also the shape center coordinates of the flue gas mask. The calculation formula is: (9) And the weighted covariance matrix, i.e., the scattering matrix of the flue gas mask, can describe the dispersion state of the flue gas, and its calculation formula is: (10) in, The distribution matrix of the flue gas mask. It is a matrix composed of all the pixels stacked together.

[0035] And the boundary equation of the minimum area envelope ellipse, the calculation formula is: (11) in, The boundary equation is for the minimum area envelope ellipse.

[0036] Eigenvalue decomposition yields the following formula: (12) in, For the corresponding unit eigenvector, Boundary equations for the minimum area envelope ellipse eigenvalues.

[0037] The minimum area envelope ellipse obtained is as follows: Figure 3 As shown, the physical lengths of the major and minor semi-axes of the minimum area enclosing ellipse are then calculated using the following formulas: (13) in, The physical length of the major semi-axis. This is the physical length of the minor semi-axis.

[0038] Finally, the deflection angle of the major axis is calculated using the following formula: (14) First, the pixel coordinates of the flue gas mask are converted to homogeneous coordinates. The center translation and scale rotation of the ellipse are handled together in a unified matrix expression, greatly simplifying the complexity of algebraic operations. Simultaneously, initial weights are assigned to each pixel of the flue gas mask to construct a weight vector. Initially, it can be assumed that all pixels contribute equally to the formation of the ellipse.

[0039] Next, we enter the iterative loop. The default score of each pixel is calculated based on the current weight vector. The current weight vector corresponds to a temporarily fitted ellipse. We check if each pixel's coordinates are inside this temporary ellipse. If a pixel is outside, its default score will be higher. To ensure the final ellipse encompasses all pixels, the algorithm needs to identify the most prominent problem in the current state—the pixel with the highest default score. To include this outermost pixel, the algorithm updates the step size parameter using specific mathematical rules. Then, the updated step size parameter is used to iteratively update the weight vector. The essence of the update is to increase the weight of the pixel with the highest default score, giving it a larger weight in the next fitting round, thus forcing the ellipse's boundary to expand towards that pixel. By repeatedly calculating the default score, finding the pixel with the highest default score, and updating the step size and weights, the ellipse gradually encompasses all the protruding outer pixels until the convergence condition is met—that is, all pixels are completely enclosed. The resulting weight vector is then used to directly calculate and generate the optimal shape matrix.

[0040] Step S300: Calculate and compare the spectral mean values ​​of the two end regions of the diffusion axis in the blue light band, determine the end with the larger spectral mean value as the initial fire source, and determine the end with the smaller spectral mean value as the initial plume tail.

[0041] In another embodiment of the invention, the two end regions of the diffusion axis are divided as follows: the normalized projection coordinates of the pixels of the flue gas mask along the direction of the diffusion axis are calculated; a preset end cropping ratio is obtained, and a cropping interval corresponding to both ends of the diffusion axis is set based on the end cropping ratio; the pixels in the normalized projection coordinates that fall into the cropping interval are divided into two end pixel sets, which are the two end regions of the diffusion axis.

[0042] The specific calculation process is as follows: Obtain the coordinates of the center pixel of the fitted minimum area envelope ellipse, and denote it as a vector with row and column direction components. Get the length of the semi-major axis of the ellipse. and the major axis unit vector along the diffusion principal axis direction The major axis unit vector contains the corresponding column direction component. and row direction components The coordinates of the two endpoints of the diffusion principal axis were calculated as follows: and .

[0043] For any current pixel within the flue gas mask, obtain its pixel coordinates. Calculate the current cell coordinates relative to the center cell coordinates. The difference is calculated, and this difference is then projected onto the major axis unit vector along the corresponding dimension to obtain the signed projected coordinates of the pixel along the diffusion principal axis. The calculation formula is: (15) Signed projection coordinates Divide by the length of the major semi-axis Obtain normalized projected coordinates This ensures that the projected coordinates of all pixels are mapped to a specific closed interval, i.e. .

[0044] Get the preset end cropping ratio Based on this end-truncation ratio, truncation intervals are set at both ends of the diffusion principal axis. All pixels are traversed, and pixels whose normalized projected coordinates fall within the corresponding truncation intervals are divided into two end-pixel sets. and The criteria for discrimination are: (16) (17) The two sets of end pixels mentioned above constitute the two end regions of the diffusion principal axis. The blue light band spectral value corresponding to each pixel in these two sets of end pixels is extracted. The spectral values ​​within each set are summed and then divided by the total number of pixels in the corresponding set. and The average blue light band spectrum of the two end pixel sets was calculated separately. and The calculation formulas are as follows: (18) (19) After obtaining the average blue light spectrum values ​​of the two end pixel sets, numerical comparison logic is executed. Since the concentration of smoke aerosols is higher and the scattering and reflection of the blue light band is stronger closer to the actual ignition point, the coordinates of the major axis endpoint corresponding to the end with the larger average spectrum value are determined as the initial fire source point, denoted as [reference needed]. Simultaneously, the coordinates of the major axis endpoint corresponding to the end with the smaller spectral mean are determined as the initial plume tail, denoted as... The coordinate assignment condition formulas are as follows: (20) (twenty one) In another embodiment of the invention, determining the flue gas diffusion direction includes: Obtain the two-dimensional coordinates of the actual fire source and the actual plume tail; Based on the two-dimensional coordinates, the azimuth angle is calculated using the arctangent function, with true north as 0 degrees and clockwise normalized to the range of 0 to 360 degrees, and the azimuth angle is used as the direction of flue gas diffusion.

[0045] Obtain the two-dimensional pixel coordinates of the true fire source point after extracting boundary features and correcting position. and the two-dimensional pixel coordinates of the actual plume tail. Calculate the difference vectors between the actual plume tail and the actual fire source on each coordinate axis.

[0046] Substitute the above coordinate difference into the bivariate arctangent function. In the middle, calculate the corresponding directional radians. Multiply this directional radian by a constant ratio. The result is converted to an angle measurement and modulo 360 degrees is performed to normalize the calculation to a closed interval between 0 and 360 degrees. Finally, the true azimuth angle of the flue gas diffusion direction, with true north as zero degrees and increasing clockwise, is obtained. The calculation formula is: (twenty two) like Figure 4 As shown, the initial fire source and initial plume tail determined at this time are only the endpoints of the major axis of the theoretical geometric envelope model, which are highly likely to deviate from the actual smoke pixel boundaries. Based on the above working principle, the embodiments of the present invention have the following technical effects: by using two-point difference combined with a bivariate arctangent function, the calculation overflow anomaly caused when the coordinate difference is zero is effectively avoided. At the same time, the strictly normalized azimuth angle of the modulo operation output is introduced, eliminating the ambiguity of multi-valued angles, so that the output results can be directly and seamlessly integrated with existing geographic information systems and fire fighting command platforms.

[0047] In actual forest and grassland fire monitoring, early-stage fire smoke may only be tens to hundreds of meters long. Using absolute pixel counts or physical distances to delineate the two ends of the main axis is not applicable to all fire smoke. Therefore, this embodiment first calculates the normalized projection coordinates of the smoke mask pixels along the diffusion main axis. By projection, the two-dimensional distribution is mapped onto the one-dimensional main axis, and through normalization, regardless of the physical length of the smoke, it is uniformly compressed into a standard-scale line segment along the main axis within the system. Subsequently, a preset end-truncation ratio is obtained, for example, 10%. Based on this ratio, interception intervals corresponding to the two ends of the diffusion main axis are set. This is equivalent to clearly defining the positions of the foremost 10% interval and the last 10% interval on this standard-scale line segment. Finally, pixel filtering and comparison are performed, selecting pixels whose normalized projection coordinates fall within the interception intervals and dividing them into two independent end pixel sets. These two sets of end pixels naturally constitute a local region containing a sufficient number of representative pixels, serving as the two end regions of the diffusion principal axis for subsequent spectral mean calculation and analysis.

[0048] By introducing normalized projected coordinates and an end-truncation ratio, a scale-independent adaptive end-region truncation mechanism is constructed. This completely eliminates the interference of absolute length differences in smoke when dividing the end regions. The method consistently and proportionally extracts statistically representative local regions at both ends. This standardized truncation significantly improves the stability and contrast of subsequent blue light spectral mean calculation, thereby ensuring a high success rate in initial pole determination. Those skilled in the art can flexibly set specific percentage values ​​corresponding to the end-truncation ratio, for example, between 5% and 20%, and this embodiment of the invention does not limit this.

[0049] Step S400: Within the area to be monitored, extract the true boundary features of the smoke mask based on the initial fire source and the initial smoke plume tail, and perform position correction on the initial fire source and the initial smoke plume tail according to the true boundary features to determine the true fire source and the direction of smoke diffusion.

[0050] In another embodiment of the invention, the step of extracting the true boundary features of the smoke mask includes: constructing a first vector along the diffusion axis and a second vector orthogonal to the first vector based on the initial fire source and the initial plume tail; and calculating the projection of each pixel of the smoke mask onto the first vector and the second vector.

[0051] The specific calculation process is as follows: Obtain the coordinates of the initial fire source determined in the previous steps. and the initial tail coordinates of the smoke plume ,in and These represent the column and row coordinates of a pixel, respectively. Calculate the pixel column increment from the initial fire source to the initial plume tail. with row increment Its algebraic expression is: (twenty three) (twenty four) Using the aforementioned row and column increments and row increments, the linear spatial distance between two points is calculated. The column and row increments are then divided by this spatial distance for magnitude normalization, constructing a unit spatial vector along the diffusion principal axis. This unit vector is used as the first vector. Simultaneously, based on the two-dimensional orthogonality rule, the row components of the first vector are inverted and their positions are swapped with the column components to construct a second vector that is strictly orthogonal to the first vector. First vector With the second vector The calculation formula is: (25) (26) Iterate through all the pixel sets extracted by the smoke mask to obtain the first... The coordinate vector of each pixel The pixel coordinate vector is compared with the first vector. The transpose matrix and the second vector The inner product operation is performed on the transpose of the matrix. This operation geometrically performs a projection transformation of the coordinate space, thereby calculating the nth... The projection scalar of each pixel onto the first vector and the projection scalar on the second vector The calculation formula is as follows: (27) (28) Among them, the projected scalar Used to characterize the longitudinal progress of pixels along the fluid diffusion direction, constructing a tangent index for translational scanning; projection scalar Used to characterize the degree of lateral dispersion of pixels from the diffusion principal axis.

[0052] The original coordinate system of remote sensing images is usually based on a geographic coordinate system or the row and column coordinate system of the image, oriented due north and due east. However, smoke from forest fires disperses along any wind direction. Continuing to use the original coordinate system to accurately extract the true boundary of the smoke along the wind direction would lead to exceptionally complex calculations and be easily affected by directional tilt. Therefore, this embodiment constructs a first vector along the main diffusion axis based on the already acquired initial fire source and initial plume tail. This first vector essentially establishes the horizontal axis of a local reference coordinate system for the smoke, which varies with the wind direction, such as... Figure 3 As shown. Simultaneously, a second vector orthogonal to the first vector is constructed as the vertical axis of this local reference coordinate system.

[0053] Subsequently, the projections of each pixel of the smoke mask onto the first and second vectors are calculated. Through this projection calculation, the originally chaotic set of two-dimensional pixels is endowed with entirely new spatial relative positional attributes. The projection value of a pixel onto the first vector reflects its distance from the fire source to the smoke tail along the entire diffusion line; the projection value of a pixel onto the second vector reflects its lateral distance from the center line of the diffusion axis. Through this process, the operation of extracting true boundary features is reduced in dimensionality, transforming the problem from dealing with complex two-dimensional image contours to dealing with two independent sets of one-dimensional numerical values.

[0054] In another embodiment of the invention, the step of correcting the position of the initial fire source and the initial plume tail based on the true boundary features to determine the true fire source and the direction of smoke diffusion includes: determining the extreme value of the projection on the first vector among all pixels of the smoke mask, and obtaining the endpoint of the scanning interval; translating the tangent along the first vector to the endpoint of the scanning interval, and selecting the median point of the projection on the second vector from the candidate pixel set intercepted by the tangent, and obtaining the corrected true fire source and the true plume tail respectively.

[0055] Obtain the projection scalar of the first vector corresponding to all pixels within the smoke mask. After setting the set, the global minimum and global maximum values ​​are extracted using an extremum search algorithm. The extracted minimum values ​​are then assigned to variables. Assign the maximum value to the variable The calculation formula is: (29) (30) The two extreme values ​​mentioned above and In physical space, the intercept positions of the first and last mask pixels to be touched and left are characterized when translating the tangent along the first vector direction, and these are established as the endpoints of the scanning interval used to lock the true source and tail.

[0056] Set a very small constant tolerance threshold Construct a scan tangent perpendicular to the first vector and translate it to the endpoint of the aforementioned minimum scan interval. The position of the first vector projection scalar is selected from all pixels. Satisfying inequalities The cells satisfying this inequality mean that the cell falls exactly within the extremely narrow spatial band intercepted by the tangent, and these cells are aggregated into a candidate cell set for correction.

[0057] Extract the second vector projection scalar corresponding to each pixel in the candidate pixel set. By calculating the spatial median operator argmedian, the above normal projection values ​​are sorted and filtered, and the pixel coordinates whose numerical distribution is in the middle are forcibly selected as the true fire source point after correction. Similarly, shift the tangent to the endpoint of the maximum value. At this location, the median point of the candidate pixel set on the second vector projection scalar is selected to obtain the true plume tail for correction. The calculation formula is: (31) (32) Combination Figure 4 and Figure 5 The comparison shows that by translating the tangent along the first vector direction and extracting the candidate pixel set, this embodiment successfully corrected the initially suspended extreme points to the actual smoke pixel boundaries. Furthermore, by taking the median point of the normal direction, it greatly avoided interference from edge spikes, thus locking in extremely robust real fire sources and real smoke plume tails. After completing pixel projection using the first and second vectors, it is necessary to mathematically find the true ends of the smoke. Among all pixels in the smoke mask, the projection extreme value on the first vector is determined. This extreme value represents the closest and farthest positions of the smoke pixel group spreading along the wind direction from the fire source. Obtaining this extreme value means obtaining the endpoints of the scanning interval. This step strictly limits the search range to the outermost edge of the actual smoke pixels, completely eliminating the initial fire source points previously suspended on the background.

[0058] Subsequently, the tangent is translated along the first vector to the endpoint of the scanning interval. Since the edge of the smoke plume is usually not a perfectly smooth point, but may exhibit irregular jagged edges or form multiple small branches due to wind dispersion, when the tangent reaches the outermost endpoint of the scanning interval, what is intercepted by the tangent is often not a single pixel, but a set of candidate pixels composed of multiple pixels. If a point is randomly selected from the candidate pixel set, or the mean is calculated, it is easily affected by individual free noise pixels or burr pixels that deviate from the main body, causing a lateral shift in the fire source location. Therefore, in this embodiment, the median point of the projection onto the second vector is selected from the candidate pixel set. The second vector represents normal expansion; taking the median of its projection means that the algorithm will ignore those extremely biased noise pixels on the outermost contact surface and stably select the pixel with the most central lateral position. In this way, the corrected true fire source point and the true smoke plume tail are obtained respectively.

[0059] Overall, the method provided by this invention includes image acquisition and mask recognition steps, envelope ellipse fitting and principal axis determination steps, end region spectral mean comparison and initial pole determination steps, and boundary feature extraction and position correction steps. Through these steps, the method achieves accurate extraction of the physical fire source location and quantification of the diffusion direction from an irregular fluid mask.

[0060] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for early fire source localization based on multispectral image features, characterized in that, Includes the following steps: Acquire multispectral remote sensing images of the area to be monitored and identify the smoke mask from them, wherein the multispectral remote sensing images contain at least the blue light band; Based on the pixel distribution characteristics of the flue gas mask, the minimum area envelope ellipse of the flue gas mask is determined, and the major axis of the minimum area envelope ellipse is taken as the diffusion principal axis of the flue gas mask. Calculate and compare the spectral mean values ​​of the two end regions of the diffusion axis in the blue light band, determine the end with the larger spectral mean value as the initial fire source, and determine the end with the smaller spectral mean value as the initial plume tail. Within the monitored area, the true boundary features of the smoke mask are extracted based on the initial fire source and the tail of the initial smoke plume. The position of the initial fire source and the tail of the initial smoke plume is corrected according to the true boundary features to determine the true fire source and the direction of smoke diffusion.

2. The method according to claim 1, characterized in that, The extraction of the true boundary features of the flue gas mask includes: Based on the initial fire source and the initial plume tail, a first vector is constructed along the main diffusion axis, and a second vector is orthogonal to the first vector. Calculate the projections of each pixel of the smoke mask onto the first vector and the second vector.

3. The method according to claim 2, characterized in that, The step of correcting the position of the initial fire source and the initial plume tail based on the actual boundary features, and determining the actual fire source and the direction of smoke diffusion, includes: Among all pixels of the smoke mask, determine the extreme values ​​of the projection on the first vector and obtain the endpoints of the scanning interval; The tangent is translated along the first vector to the endpoint of the scanning interval. In the candidate pixel set intercepted by the tangent, the median point of the projection on the second vector is selected to obtain the corrected real fire source point and real smoke plume tail.

4. The method according to claim 1, characterized in that, Determining the minimum area envelope ellipse of the flue gas mask includes: A shape matrix is ​​constructed using the minimum volume envelope ellipse fitting method (i.e., minimum area envelope ellipse in a two-dimensional scene), and the objective function is constructed using the determinant of the shape matrix. Solve for the optimal solution of the objective function to obtain the optimal shape matrix of all pixels enclosing the smoke mask, and determine the center of the minimum area envelope ellipse, the major axis and the minor axis based on the optimal shape matrix.

5. The method according to claim 4, characterized in that, Solving for the optimal solution of the objective function includes: The cell coordinates of the flue gas mask are converted into homogeneous coordinates, and an initial weight is assigned to each cell of the flue gas mask to construct a weight vector; The default degree of each cell is calculated based on the current weight vector; The step size parameter is updated based on the maximum default degree, and the weight vector is iteratively updated until the convergence condition is met, thus obtaining the optimal shape matrix.

6. The method according to claim 1, characterized in that, The two end regions of the diffusion main shaft are divided as follows: Calculate the normalized projected coordinates of the pixels of the flue gas mask along the main diffusion axis; Obtain a preset end-cut ratio, and set the cut-off intervals corresponding to both ends of the diffusion main shaft based on the end-cut ratio; The pixels falling within the intercepted interval in the normalized projected coordinates are divided into two sets of end pixels, which serve as the two end regions of the diffusion principal axis.

7. The method according to claim 1, characterized in that, The acquisition of multispectral remote sensing images of the area to be monitored includes: The multispectral remote sensing image is converted into multiband apparent reflectance data; The multispectral remote sensing image includes at least the green light band and the red light band.

8. The method according to claim 1 or 3, characterized in that, Determining the direction of flue gas diffusion includes: Obtain the two-dimensional coordinates of the actual fire source and the actual plume tail; Based on the two-dimensional coordinates, the azimuth angle is calculated using the arctangent function, with true north as 0 degrees and clockwise normalized to the range of 0 to 360 degrees, and the azimuth angle is used as the direction of flue gas diffusion.

Images