Forest and grassland surface fire detection method based on multispectral characteristic response
By combining multispectral sensors and support vector machine models, the problem of existing fire monitoring systems being unable to identify fires and interference factors has been solved, enabling accurate fire monitoring and early warning, and improving the accuracy of fire classification and identification.
Patent Information
- Application Number
- CN202410972751.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2044-07-19
AI Technical Summary
Existing fire monitoring systems lack the ability to comprehensively analyze multispectral information, cannot effectively identify and distinguish fires from other interfering factors, cannot achieve accurate fire location and identification, and lack the ability to proactively analyze fire characteristics or predict fire development.
Multispectral feature response information is acquired using mid-wave infrared sensors, ultraviolet sensors, and CMOS sensors. Time-domain and frequency-domain feature vector data are extracted through a feature extraction module. The data is then standardized and normalized using a support vector machine model. Classification is performed using polynomial kernel functions and radial basis kernel functions to achieve accurate fire identification and early warning.
It achieves efficient and reliable monitoring and identification of fires, significantly reduces the probability of misclassification, accurately distinguishes between fires and interference factors, and enables precise fire location and early warning.
Smart Images

Figure CN119169444B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of spectral analysis and image processing technology, and in particular to a method for detecting forest and grassland surface fires based on multispectral feature responses. Background Technology
[0002] Mid-wave infrared sensors and ultraviolet sensors are used to acquire infrared and ultraviolet response spectra, respectively. These sensors can be installed on towers or monitoring equipment at a certain height above the ground to achieve real-time monitoring of large-scale forest and grassland areas. According to the National Forestry and Grassland Administration's Forest Fire Prevention Plan 2016-2025, the video surveillance coverage rate in forest areas nationwide will reach over 95% by 2025, with a large portion of the monitoring equipment using infrared and ultraviolet sensors. Traditional fire monitoring systems typically rely on a single spectrum or manual inspection, lacking intelligent and comprehensive analysis capabilities, and are unable to detect and warn of fires in a timely and accurate manner.
[0003] Most existing fire monitoring systems can only perform simple infrared monitoring and lack the ability to comprehensively analyze multispectral information. They are unable to effectively identify and distinguish fires from other interfering factors, such as direct sunlight, reflected light, and electromagnetic interference. Although some systems have automatic alarm functions, the false alarm rate is high, making accurate fire location and identification impossible. Furthermore, traditional monitoring systems often only passively record data and cannot actively analyze fire characteristics or predict fire development.
[0004] The shortcomings of existing technologies are: lack of comprehensive analysis capabilities of multispectral information, inability to effectively identify and distinguish fire from other interfering factors; inability to achieve accurate fire location and identification, and inability to proactively analyze fire characteristics or predict fire development. Summary of the Invention
[0005] This invention provides a forest and grassland surface fire detection method based on multispectral feature response, which improves the accuracy of surface fire classification and identification.
[0006] To achieve the above objectives, this invention provides a forest and grassland surface fire detection method based on multispectral characteristic response, the key of which includes the following steps:
[0007] Step 1: The mid-wave infrared sensor acquires the infrared response spectrum a in real time and transmits it to the feature extraction module; the ultraviolet sensor acquires the ultraviolet response spectrum b in real time and transmits it to the feature extraction module; the CMOS sensor acquires the image response spectrum c in real time and transmits it to the feature extraction module.
[0008] Step 2: The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the infrared response spectrum a, forming an infrared spectral feature vector set A, and then passes it to the preprocessing module;
[0009] The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the ultraviolet response spectrum b, forming an ultraviolet spectral feature vector set B, which is then passed to the preprocessing module.
[0010] The feature extraction module extracts image feature vector data from the image response spectrum c, forms a time series image feature vector set C, and passes it to the preprocessing module.
[0011] Step 3: The preprocessing module performs standardization and normalization on the infrared spectral feature vector set A, the ultraviolet spectral feature vector set B, and the time series image feature vector set C, respectively, to obtain the infrared spectral normalized vector set A', the ultraviolet spectral normalized vector set B', and the time series image normalized vector set C'. Then, the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' are passed to the first support vector machine model SVM1, and the time series image normalized vector set C' is passed to the second support vector machine model SVM2.
[0012] Step 4: The first support vector machine model SVM1 classifies the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' using a polynomial kernel function, classifying them into interference data X and suspected surface fire data Y, and then sends the suspected surface fire data Y to the second support vector machine model SVM2.
[0013] Step 5: The second support vector machine model SVM2 classifies the normalized vector set C' of the time series image and the suspected surface fire data Y using the radial basis kernel function, classifying them into interference data X and surface fire data Z. If surface fire data Z is not classified, it means that no fire has occurred, and proceed to step 1; otherwise, it means that a fire has occurred, and a fire alarm is triggered.
[0014] Through the above design, multispectral sensors combined with image processing technology can comprehensively analyze and identify fire characteristics. Mid-wave infrared and ultraviolet response spectra provide characteristic information such as flames and smoke, while image processing captures dynamic changes at the fire scene. By comprehensively analyzing these three types of spectral information, fire can be accurately distinguished from other interfering factors, achieving precise fire monitoring and early warning.
[0015] The preprocessing module is used to standardize and normalize the feature vector set. The purpose of normalization is to standardize the data of different features so that they have the same dimensions and distribution range, thereby eliminating the dimensional differences between different features and facilitating the subsequent training and classification of the support vector machine (SVM) model.
[0016] Preferably, in step 2, the time-domain feature vector data of the infrared response spectrum a includes duration T, zero-crossing rate ZCR, energy E, envelope, and variance σ. 2 Kurtosis K, autocorrelation ACF, cross-correlation CCF, and filter complexity;
[0017] Where, duration T is the total time the signal strength exceeds a certain threshold, zero-crossing rate ZCR is the number of times the signal crosses zero per unit time, energy E is the integral of the sum of squares of the signal, and variance σ 2 Let K be the variance of the signal strength, K be the kurtosis of the signal, ACF be the autocorrelation of the signal itself, CCF be the cross-correlation between two signals, and C bevel complexity be the complexity of the signal.
[0018] The energy E is calculated using the following expression:
[0019] E=∫|A(t| 2 dt
[0020] The variance σ 2 The calculation expression is as follows:
[0021]
[0022] The formula for calculating the kurtosis K is as follows:
[0023]
[0024] The calculation expression for the autocorrelation ACF is as follows:
[0025]
[0026] The calculation expression for the cross-correlation CCF is as follows:
[0027]
[0028] Where A(t) represents the amplitude of the infrared response signal with respect to continuous time t, A i A represents the amplitude of the infrared response signal at the i-th discrete time point. i+τ Let A(i) represent the amplitude of the infrared response signal at the i-th discrete time point, τ represent the average value of the infrared response signal, σ represent the standard deviation of the infrared response signal, and τ represent the time lag; let A(i) represent the amplitude of the infrared response signal at the i-th discrete time point, B(i+τ) represent the amplitude of the ultraviolet response signal at the i-th discrete time point, and N represent the N discrete time points.
[0029] The time-domain feature vector data of the ultraviolet response spectrum b includes duration, zero-crossing rate, energy, envelope, variance, kurtosis, autocorrelation, cross-correlation, and filter complexity.
[0030] The calculation expression for the time-domain eigenvector data of the ultraviolet response spectrum b is the same as the calculation formula for the time-domain eigenvector data of the infrared response spectrum a, only the input is changed to the ultraviolet response signal.
[0031] Preferably, in step 2, the frequency domain feature vector data of the infrared response spectrum a includes the zero-frequency component F0 and the non-zero-frequency component F nz Amplitude D, phase φ, power spectral density PSD;
[0032] Wherein, amplitude D is the magnitude of the frequency domain signal, and phase φ is the angle of the frequency domain signal.
[0033] The feature extraction module converts the time-domain signal A(t) into the frequency-domain signal A(f) using a fast Fourier transform, as shown in the following expression:
[0034]
[0035] The expression for calculating the amplitude D is as follows:
[0036]
[0037] The expression for calculating the phase φ is as follows:
[0038] φ = arg(A(f))
[0039] The formula for calculating the power spectral density (PSD) is as follows:
[0040]
[0041] Where A(t) represents the time-domain signal of infrared response spectrum a, and A(f) represents the frequency-domain signal of infrared response spectrum a;
[0042] The frequency domain eigenvector data of the ultraviolet response spectrum b includes the zero-frequency component, non-zero-frequency component, amplitude, phase, and power spectral density.
[0043] The calculation expression for the frequency domain eigenvector data of the ultraviolet response spectrum b is the same as that for the frequency domain eigenvector data of the infrared response spectrum a, only the input is changed to the ultraviolet response signal.
[0044] Preferably, in step 2, the image feature vector data includes a color histogram, a texture binary pattern map, a highlight shrinkage and oscillation time series, a highlight target area, and a highlight boundary curvature.
[0045] Among them, the color histogram is the histogram distribution of each color channel of the image; the texture binary pattern map is the local binary pattern used to describe the image texture; the highlight shrinkage and swing time series is the time series change of the highlight area; the highlight target area is the area of the highlight area in the image; and the highlight boundary curvature is the curvature of the boundary of the highlight area.
[0046] This invention significantly reduces the probability of misclassification by utilizing the dynamic features of the highlighted areas in image time series, making it more reliable and effective than using only the bispectral algorithm.
[0047] Preferably, the infrared response spectrum has a spectral range of 3.0 μm to 5.0 μm, the ultraviolet response spectrum has a spectral range of 180 nm to 300 nm, and the image response spectrum has a spectral range of 350 nm to 1000 nm.
[0048] Preferably, in step 4, the expression for the polynomial kernel function is as follows:
[0049] K(A′ j ,B′ j )=(γ <A′ j ,B′ j >+r) d
[0050] Where, K(A′) j ,B′ j ) represents the polynomial kernel function, A′ j Let B′ represent the j-th infrared spectral normalized vector sample point. j This represents the j-th normalized vector sample point of the ultraviolet spectrum. <A′ j ,B′ j > represents sample point A′ j and B′ j The inner product, i.e., the dot product; γ represents the kernel function parameter that controls the degree of influence of the inner product; r represents the offset or constant term kernel function parameter; d represents the order of the polynomial, which determines the nonlinearity of the polynomial kernel.
[0051] The role of a multinomial kernel function is to map the input data to a higher-dimensional space, making the data linearly separable in this higher-dimensional space, thereby enabling more complex classification tasks.
[0052] Preferably, in step 5, the expression for the radial basis kernel function is as follows:
[0053] G(C′ j ,Y j )=exp(-γ∥C′ j -Y j ∥ 2 )
[0054] Among them, G(C′ j ,Y j ) represents the radial basis function kernel function, C′ j Y represents the normalized vector sample point of the j-th time series image. j Let $C$ represent the j-th suspected surface fire data sample point, $\overrightarrow{C}$. j -Y j ∥ 2 Represents C′ j and Y j The square of the Euclidean distance between them, exp() represents the natural exponential function.
[0055] The characteristic of the radial basis function (RBF) kernel is that the greater the distance between two sample points, the closer the kernel function value is to zero; conversely, the closer the distance between two sample points, the closer the kernel function value is to one. The advantage of the RBF kernel is that it can map the original input data to an infinite-dimensional feature space, thus handling highly nonlinear data distributions. It is suitable for many complex classification and regression problems and can improve classification accuracy.
[0056] Preferably, in steps 4 and 5, the input expression for the first support vector machine model SVM1 or the second support vector machine model SVM2 is as follows:
[0057] I = {(x1,y1),(x2,y2),…,(x n ,y n )}
[0058] The output expression of the first support vector machine model, SVM1, is as follows:
[0059] O = {z1, z2, ..., z} n}
[0060] The decision function of the first support vector machine model, SVM1, is expressed as follows:
[0061]
[0062] Where I represents the input, and in SVM1 the input is (A ′ The input in SVM2 is (C,B′), ′ (x, y) ; n represents the number of samples, O represents the output, in SVM1 O∈{X,Y}, in SVM2 O∈{X,Z}, where X represents the interference class, Y represents the suspected surface fire class, and Z represents the surface fire class; α and e are model parameters, α represents the Lagrange multiplier, representing the contribution of the sample point to the decision surface, and b is the bias term; sign is the sign function, used to determine the sign of the classification result. If f(x, y)>0, then sign(f(x, y))=1, and the input sample is classified as positive.
[0063] If f(x, y) < 0, then sign(f(x, y)) = -1, and the input sample is classified as the negative class.
[0064] The beneficial effects of this invention are: by combining multispectral response features and a two-level support vector machine model, efficient and reliable monitoring and identification of surface fires are achieved, effectively improving the accuracy of surface fire classification and identification. Attached Figure Description
[0065] Figure 1 This is a schematic diagram of the process of the present invention;
[0066] Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation
[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0068] like Figure 1 , Figure 2 The following is an example of a forest-grassland surface fire detection method based on multispectral characteristic response, comprising the following steps:
[0069] Step 1: The mid-wave infrared sensor acquires the infrared response spectrum a in real time and transmits it to the feature extraction module; the ultraviolet sensor acquires the ultraviolet response spectrum b in real time and transmits it to the feature extraction module; the CMOS sensor acquires the image response spectrum c in real time and transmits it to the feature extraction module.
[0070] Step 2: The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the infrared response spectrum a, forming an infrared spectral feature vector set A, and then passes it to the preprocessing module;
[0071] The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the ultraviolet response spectrum b, forming an ultraviolet spectral feature vector set B, which is then passed to the preprocessing module.
[0072] The feature extraction module extracts image feature vector data from the image response spectrum c, forms a time series image feature vector set C, and passes it to the preprocessing module.
[0073] Step 3: The preprocessing module performs standardization and normalization on the infrared spectral feature vector set A, the ultraviolet spectral feature vector set B, and the time series image feature vector set C, respectively, to obtain the infrared spectral normalized vector set A', the ultraviolet spectral normalized vector set B', and the time series image normalized vector set C'. Then, the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' are passed to the first support vector machine model SVM1, and the time series image normalized vector set C' is passed to the second support vector machine model SVM2.
[0074] Step 4: The first support vector machine model SVM1 classifies the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' using a polynomial kernel function, classifying them into interference data X and suspected surface fire data Y, and then sends the suspected surface fire data Y to the second support vector machine model SVM2.
[0075] Step 5: The second support vector machine model SVM2 classifies the normalized vector set C' of the time series image and the suspected surface fire data Y using the radial basis kernel function, classifying them into interference data X and surface fire data Z. If surface fire data Z is not classified, it means that no fire has occurred, and proceed to step 1; otherwise, it means that a fire has occurred, and a fire alarm is triggered.
[0076] In step 2, the time-domain feature vector data of the infrared response spectrum a includes duration T, zero-crossing rate ZCR, energy E, envelope, and variance σ. 2 Kurtosis K, autocorrelation ACF, cross-correlation CCF, and filter complexity;
[0077] The energy E is calculated using the following expression:
[0078] E=∫|A(t| 2 dt
[0079] The variance σ 2 The calculation expression is as follows:
[0080]
[0081] The formula for calculating the kurtosis K is as follows:
[0082]
[0083] The calculation expression for the autocorrelation ACF is as follows:
[0084]
[0085] The calculation expression for the cross-correlation CCF is as follows:
[0086]
[0087] Where A(t) represents the amplitude of the infrared response signal with respect to continuous time t, A i A represents the amplitude of the infrared response signal at the i-th discrete time point. i+τ Let A(i) represent the amplitude of the infrared response signal at the i-th discrete time point, τ represent the average value of the infrared response signal, σ represent the standard deviation of the infrared response signal, and τ represent the time lag; let A(i) represent the amplitude of the infrared response signal at the i-th discrete time point, B(i+τ) represent the amplitude of the ultraviolet response signal at the i-th discrete time point, and N represent the N discrete time points.
[0088] The time-domain feature vector data of the ultraviolet response spectrum b includes duration, zero-crossing rate, energy, envelope, variance, kurtosis, autocorrelation, cross-correlation, and filter complexity.
[0089] In step 2, the frequency domain feature vector data of the infrared response spectrum a includes the zero-frequency component F0 and the non-zero-frequency component F nz Amplitude D, phase φ, power spectral density PSD;
[0090] The feature extraction module converts the time-domain signal A(t) into the frequency-domain signal A(f) using a fast Fourier transform, as shown in the following expression:
[0091]
[0092] The expression for calculating the amplitude D is as follows:
[0093] D = |A(f)|
[0094] The expression for calculating the phase φ is as follows:
[0095] φ = arg(A(f))
[0096] The formula for calculating the power spectral density (PSD) is as follows:
[0097]
[0098] Where A(t) represents the time-domain signal of infrared response spectrum a, and A(f) represents the frequency-domain signal of infrared response spectrum a;
[0099] The frequency domain eigenvector data of the ultraviolet response spectrum b includes the zero-frequency component, non-zero-frequency component, amplitude, phase, and power spectral density.
[0100] In step 2, the image feature vector data includes a color histogram, a texture binary pattern map, a highlight shrinkage and oscillation time series, a highlight target area, and a highlight boundary curvature.
[0101] The infrared response spectrum has a spectral range of 3.0µm-5.0µm, the ultraviolet response spectrum has a spectral range of 180nm-300nm, and the image response spectrum has a spectral range of 350nm-1000nm.
[0102] In step 4, the expression for the polynomial kernel function is as follows:
[0103] K(A′ j ,B′ j )=(γ <A′ j ,B′ j >+r) d
[0104] Where, K(A′) j ,B′ j ) represents the polynomial kernel function, A′ j Let B′ represent the j-th infrared spectral normalized vector sample point. j This represents the j-th normalized vector sample point of the ultraviolet spectrum. <A′ j ,B′ j > represents sample point A′ j and B′ j The inner product, i.e., the dot product; γ represents the kernel function parameter that controls the degree of influence of the inner product; r represents the offset or constant term kernel function parameter; d represents the order of the polynomial, which determines the nonlinearity of the polynomial kernel.
[0105] In step 5, the expression for the radial basis kernel function is as follows:
[0106] G(C′ j ,Y j )=exp(-γ∥C′ j -Y j ∥ 2 )
[0107] Among them, G(C′ j ,Y j ) represents the radial basis function kernel function, C′ j Y represents the j-th time series image normalized vector sample point. j Let $C$ represent the j-th suspected surface fire data sample point, $\overrightarrow{C}$. j -Y j ∥ 2 Represents C′ j and Y j The square of the Euclidean distance between them, exp() represents the natural exponential function.
[0108] In step 4 or step 5, the input expression for the first support vector machine model SVM1 or the second support vector machine model SVM2 is as follows:
[0109] I = {(x1,y1),(x2,y2),…,(x n ,y n )}
[0110] The output expression of the first support vector machine model, SVM1, is as follows:
[0111] O = {z1, z2, ..., z} n}
[0112] The decision function of the first support vector machine model, SVM1, is expressed as follows:
[0113]
[0114] Where I represents the input, and in SVM1 the input is (A ′ The input in SVM2 is (C,B′), ′ (x, y) ; n represents the number of samples, O represents the output, in SVM1 O∈{X,Y}, in SVM2 O∈{X,Z}, where X represents the interference class, Y represents the suspected surface fire class, and Z represents the surface fire class; α and e are model parameters, α represents the Lagrange multiplier, representing the contribution of the sample point to the decision surface, and b is the bias term; sign is the sign function, used to determine the sign of the classification result. If f(x, y)>0, then sign(f(x, y))=1, and the input sample is classified as positive.
[0115] If f(x, y) < 0, then sign(f(x, y)) = -1, and the input sample is classified as the negative class.
[0116] The first support vector machine model SVM1 and the second support vector machine model SVM2 are obtained through the following steps:
[0117] Step 1: The infrared sensor acquires the infrared response spectrum a and transmits it to the feature extraction module; the ultraviolet sensor acquires the ultraviolet response spectrum b and transmits it to the feature extraction module; the CMOS sensor acquires the image response spectrum c and transmits it to the feature extraction module.
[0118] Step 2: The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the infrared response spectrum a, forming an infrared spectral feature vector set A, and then passes it to the preprocessing module.
[0119] The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the ultraviolet response spectrum b, forming an ultraviolet spectral feature vector set B, which is then passed to the preprocessing module.
[0120] The feature extraction module extracts image feature vector data from the image response spectrum c, forms a time series image feature vector set C, and passes it to the preprocessing module.
[0121] Step 3: The preprocessing module performs standardization and normalization on the infrared spectral feature vector set A, the ultraviolet spectral feature vector set B, and the time series image feature vector set C, respectively, to obtain the infrared spectral normalized vector set A', the ultraviolet spectral normalized vector set B', and the time series image normalized vector set C'. Then, the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' are passed to the first support vector machine model SVM1, and the time series image normalized vector set C' is passed to the second support vector machine model SVM2.
[0122] Step 4: The first support vector machine model SVM1 determines the parameters and penalty parameters through a polynomial kernel function, and establishes interference factors for the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B', generating dual-spectral sample sets A” and B with interference factors and without interference factors respectively. The first support vector machine model SVM1 is trained using the dual-spectral sample sets A” and B”, and the optimal parameters of the model are determined by cross-validation, thereby optimizing the first support vector machine model. Then, the dual-spectral signals are classified into interference class X and suspected surface fire class Y, and the suspected surface fire class Y is passed to the second support vector machine model SVM2.
[0123] Step 5: The second support vector machine model SVM2 determines the parameters and penalty parameters through the radial basis kernel function, establishes the feature vector set C” for fire images, no-fire images, and interference images, trains the second support vector machine model SVM2, selects the optimal parameters through cross-validation, and then optimizes the second support vector machine model SVM2, and classifies the suspected surface fire signal into interference class X and surface fire class Z.
[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting forest and grassland surface fires based on multispectral characteristic responses, characterized in that, Includes the following steps: Step 1: The mid-wave infrared sensor acquires the infrared response spectrum 'a' in real time and transmits it to the feature extraction module; The ultraviolet sensor acquires the ultraviolet response spectrum b in real time and transmits it to the feature extraction module; The COMS sensor acquires the image response spectrum c in real time and transmits it to the feature extraction module; Step 2: The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the infrared response spectrum a, forming an infrared spectral feature vector set A, and then passes it to the preprocessing module; The feature extraction module extracts the time-domain feature vector data and frequency-domain feature vector data from the ultraviolet response spectrum b, forming an ultraviolet spectral feature vector set B, which is then passed to the preprocessing module. The feature extraction module extracts image feature vector data from the image response spectrum c, forms a time series image feature vector set C, and passes it to the preprocessing module. Step 3: The preprocessing module performs standardization and normalization on the infrared spectral feature vector set A, the ultraviolet spectral feature vector set B, and the time series image feature vector set C, respectively, to obtain the infrared spectral normalized vector set A', the ultraviolet spectral normalized vector set B', and the time series image normalized vector set C'. Then, the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' are passed to the first support vector machine model SVM1, and the time series image normalized vector set C' is passed to the second support vector machine model SVM2. Step 4: The first support vector machine model SVM1 classifies the infrared spectral normalized vector set A' and the ultraviolet spectral normalized vector set B' using a polynomial kernel function, classifying them into interference data X and suspected surface fire data Y, and then sends the suspected surface fire data Y to the second support vector machine model SVM2. Step 5: The second support vector machine model SVM2 classifies the normalized vector set C' of the time series image and the suspected surface fire data Y using the radial basis kernel function, classifying them into interference data X and surface fire data Z. If surface fire data Z is not classified, it means that no fire has occurred, and proceed to step 1; otherwise, it means that a fire has occurred, and a fire alarm is triggered. In step 4, the expression for the polynomial kernel function is as follows: ; in, Represents the polynomial kernel function. This represents the j-th infrared spectral normalized vector sample point. This represents the j-th normalized vector sample point of the ultraviolet spectrum. Represents sample points and The inner product, i.e., the dot product; γ represents the kernel function parameter that controls the degree of influence of the inner product; r represents the offset or constant term kernel function parameter; d represents the order of the polynomial, which determines the nonlinearity of the polynomial kernel; In step 5, the expression for the radial basis kernel function is as follows: ; in, Represents the radial basis kernel function. This represents the j-th time-series image normalized vector sample point. This represents the j-th suspected surface fire data sample point. express and The square of the Euclidean distance between them, exp() represents the natural exponential function, and γ represents the kernel function parameter that controls the degree of influence of the inner product.
2. The forest and grassland surface fire detection method based on multispectral characteristic response according to claim 1, characterized in that: In step 2, the time-domain feature vector data of the infrared response spectrum a includes duration T, zero-crossing rate ZCR, energy E, envelope, and variance. Kurtosis K, autocorrelation ACF, cross-correlation CCF, and filter complexity; The energy E is calculated using the following expression: ; The variance The calculation expression is as follows: ; The kurtosis The calculation expression is as follows: ; The calculation expression for the autocorrelation ACF is as follows: ; The calculation expression for the cross-correlation CCF is as follows: ; Among them, A This represents the amplitude of the infrared response signal as a function of continuous time t. This represents the amplitude of the infrared response signal at the i-th discrete time point. This represents the amplitude of the infrared response signal per unit time τ after the i-th discrete time point. This represents the average value of the infrared response signal. The standard deviation of the infrared response signal is represented by... Indicates a time lag; This represents the amplitude of the infrared response signal at the i-th discrete time point. Let represent the amplitude of the ultraviolet response signal after the i-th discrete time point, τ per unit time, and N represent the N discrete time points; The time-domain feature vector data of the ultraviolet response spectrum b includes duration, zero-crossing rate, energy, envelope, variance, kurtosis, autocorrelation, cross-correlation, and filter complexity.
3. The forest and grassland surface fire detection method based on multispectral characteristic response according to claim 1, characterized in that: In step 2, the frequency domain eigenvector data of the infrared response spectrum a includes a zero-frequency component. Non-zero frequency components Amplitude Phase Power spectral density ; The feature extraction module uses Fast Fourier Transform to transform the time-domain signal. Convert to frequency domain signal The expression is as follows: ; The amplitude The calculation expression is as follows: ; The phase The calculation expression is as follows: ; The power spectral density The calculation expression is as follows: ; in, This represents the time-domain signal of the infrared response spectrum a. The frequency domain signal representing the infrared response spectrum a is given, and N represents N discrete time points. The frequency domain eigenvector data of the ultraviolet response spectrum b includes the zero-frequency component, non-zero-frequency component, amplitude, phase, and power spectral density.
4. The forest and grassland surface fire detection method based on multispectral characteristic response according to claim 1, characterized in that: In step 2, the image feature vector data includes a color histogram, a texture binary pattern map, a highlight shrinkage and oscillation time series, a highlight target area, and a highlight boundary curvature.
5. The forest and grassland surface fire detection method based on multispectral characteristic response according to claim 1, characterized in that: The infrared response spectrum has a spectral range of 3.0 μm to 5.0 μm, the ultraviolet response spectrum has a spectral range of 180 nm to 300 nm, and the image response spectrum has a spectral range of 350 nm to 1000 nm.
6. The forest and grassland surface fire detection method based on multispectral characteristic response according to claim 1, characterized in that: In step 4 or step 5, the input expression for the first support vector machine model SVM1 or the second support vector machine model SVM2 is as follows: ; The output expression of the first support vector machine model, SVM1, is as follows: ; The decision function of the first support vector machine model, SVM1, is expressed as follows: ; in, This represents the input; in SVM1, the input is... The input in SVM2 is ; n represents the number of samples, O represents the output, O∈{X,Y} in SVM1, O∈{X,Z} in SVM2, where X represents the interference class, Y represents the suspected surface fire class, and Z represents the surface fire class; α and e are model parameters, α represents the Lagrange multiplier, representing the contribution of the sample point to the decision surface, and e is the bias term; sign is the sign function, used to determine the sign of the classification result. If > 0, then sign( If )=1, the input sample is classified as positive. If < 0, then sign( ) = -1, the input sample is classified as the negative class.
Citation Information
Patent Citations
Method for identifying semiconductor nano structure pattern
CN102735183A
Method for removing abnormal spectrum in actual measurement spectrum curve based on support vector machine model
CN104063710A
