Forest fire point recognition method based on multispectral remote sensing and ai edge computing

Abstract

The present application belongs to the technical field of image recognition, and particularly relates to a forest fire point recognition method based on multispectral remote sensing and AI edge computing, steps of which comprise: performing spatial local anomaly analysis on the multispectral remote sensing image of the target forest to obtain potential abnormal pixels and abnormal spectral features; performing spectral continuous response verification on the potential abnormal pixels to obtain candidate fire point pixels, and performing intensity calibration on the candidate fire point pixels to obtain the continuous response confidence of the candidate fire point pixels; performing spatial topological relationship reconstruction on the candidate fire point pixels to obtain a fire point spatial distribution map; performing synergistic enhancement on the node connection strength of the fire point spatial distribution map to obtain a strengthened distribution map; performing edge contour vectorization on the fire field region of the candidate fire point pixels to obtain the contour geometric data of the fire field region; and performing multi-dimensional evidence fusion decision on the fire field region to obtain the fire point event of the target forest. The present application can improve the accuracy, completeness and real-time efficiency of forest fire point recognition.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to a forest fire identification method based on multispectral remote sensing and AI edge computing. Background Technology

[0002] In the field of forest fire prevention, early and accurate fire point identification is a prerequisite for fire warning, efficient response, and disaster prevention, and is crucial for reducing fire losses and ensuring the safety of forest areas. Currently, multispectral remote sensing technology, with its advantages of wide-area, all-weather, and non-contact monitoring, has become the mainstream application in forest fire point identification. However, existing technologies still have significant shortcomings in practical applications.

[0003] First, in the core abnormal pixel detection stage, the traditional processing method of single-dimensional radiation feature analysis is still used. This method neither analyzes and quantifies local spatial anomalies at multiple scales, nor accurately captures the radiation residual intensity characteristics corresponding to the fire point. It also lacks in-depth analysis of the full-band spectral characteristics of abnormal pixels, making it difficult to effectively distinguish between real spectral anomalies caused by fire points and false anomalies caused by differences in terrain and vegetation. This has become the core problem of the low accuracy of fire point identification.

[0004] Secondly, existing technologies also have multiple shortcomings in the subsequent processing of fire point pixels. For example, the spectral verification stage lacks a multi-scale continuous evaluation system, making it impossible to accurately calibrate the response characteristics of fire point pixels; the spatial topology reconstruction does not combine confidence levels with collaborative enhancement of node connection strength, resulting in insufficient topological integrity of the spatial distribution of fire points; fire contour extraction lacks refined vectorization processing and geometric feature verification; and multi-dimensional evidence fusion decision-making does not integrate key information such as contour geometric data and confidence gradient fields. These problems, layered upon each other, ultimately lead to low accuracy in fire boundary identification and insufficient reliability in fire event determination. The overall efficiency and accuracy of the identification process are difficult to meet the actual technical requirements of real-time forest fire monitoring. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a forest fire spot identification method based on multispectral remote sensing and AI edge computing. By combining multispectral remote sensing and AI edge computing with multi-scale analysis, spectral verification, topology enhancement, contour vectorization and multi-dimensional fusion decision-making, the accuracy, completeness and real-time efficiency of forest fire spot identification can be improved.

[0006] To achieve the above objectives, the technical solution adopted by this invention is a forest fire point identification method based on multispectral remote sensing and AI edge computing, comprising the following steps: S1. Perform spatial local anomaly analysis on the multispectral remote sensing image of the target forest to obtain the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image. S2. Based on the abnormal spectral characteristics, the spectral continuous response of potential abnormal pixels is verified to obtain candidate fire point pixels of potential abnormal pixels, and the intensity of candidate fire point pixels is calibrated to obtain the continuous response confidence of candidate fire point pixels. S3. Reconstruct the spatial topological relationship of the candidate fire point pixels to obtain the fire point spatial distribution map of the candidate fire point pixels; S4. Based on the continuous response confidence, the node connection strength of the fire point spatial distribution map is enhanced collaboratively to obtain the enhanced distribution map of candidate fire point pixels. S5. Based on the enhanced distribution map, the edge contour vectorization of the fire area of ​​the candidate fire point pixels is performed to obtain the contour geometric data of the fire area. S6. Based on the contour geometric data, perform multi-dimensional evidence fusion decision-making on the fire area to obtain the fire point events of the target forest.

[0007] Preferably, in step S1, obtaining the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image includes: Multi-scale sliding window traversal is performed on the multispectral remote sensing image of the target forest to obtain the local neighborhood radiative statistics of the multispectral remote sensing image. Spatial autocorrelation weighted fusion is performed on the central pixel of the local neighborhood radiation statistics to obtain the expected radiation value of the central pixel; Based on the expected radiation value, the deviation of the actual observed radiation value of the central pixel is measured to obtain the radiation residual intensity of the central pixel; Based on the radiation residual intensity, adaptive threshold segmentation is performed on multispectral remote sensing images to obtain potential anomalous pixels in the multispectral remote sensing images. Full-band spectral characteristic analysis was performed on potential anomalous pixels to obtain the anomalous spectral features of multispectral remote sensing images.

[0008] Preferably, in step S2, obtaining the continuous response confidence of the candidate fire point pixel includes: The anomalous spectral features are divided into multi-scale bands to obtain coarse-scale band sequences and fine-scale band sequences of the anomalous spectral features. Based on Planck's radiation law, the radiation increment of coarse-scale band sequences is verified, and the macroscopic continuity index of coarse-scale band sequences is obtained. Differential continuity verification was performed on the fine-scale band sequence to obtain the micro-continuity index of the fine-scale band sequence; A synergistic evaluation of the results of macro-continuity indicators and micro-continuity indicators was conducted to obtain multi-scale synergistic scores for potential anomalous pixels. Based on multi-scale collaborative scoring, adaptive threshold screening is performed on potential anomalous pixels to obtain candidate fire point pixels of potential anomalous pixels. The candidate fire point pixels are subjected to nonlinear normalization mapping to obtain the continuous response confidence of the candidate fire point pixels.

[0009] Preferably, the process for obtaining the multi-scale collaborative score of potential anomalous pixels is as follows: We performed an analysis of the inter-scale differences between macro-continuity indicators and micro-continuity indicators to obtain the absolute differences between them. By merging macro-level and micro-level continuous indicators in a balanced manner, a geometric harmonic quantity of macro-level and micro-level continuous indicators is obtained. By coupling the absolute difference quantity and the geometric harmonic quantity, the intrinsic coherence index of potential anomalous pixels is obtained. The formula for calculating the intrinsic coherence index is as follows: ; In the formula, Let be the intrinsic coherence index of the i-th potential anomalous pixel. It is a natural exponential function. The preset collaborative penalty coefficient, For the i-th potential anomalous pixel, the macroscopic continuity index is... For the i-th potential anomalous pixel, the micro-continuity index is... This is a preset zero-prevention constant; By performing range normalization mapping on the intrinsic co-indices, multi-scale co-scores of potential anomalous pixels are obtained.

[0010] Preferably, in step S3, obtaining the spatial distribution map of candidate fire point pixels includes: The spatial distance of the candidate fire point pixels is obtained by measuring the nearest neighbor distance; Based on spatial distance, local density peak detection is performed on candidate fire point pixels to obtain the spatial aggregation center of candidate fire point pixels; By imposing radiation intensity similarity constraints on spatial clustering centers, the coupling relationships of spatial clustering centers are obtained; Based on the coupling relationship, spatial attribution determination is performed on candidate fire point pixels to obtain the regional attribution identifier of the candidate fire point pixels; Based on the region affiliation identifier, connectivity analysis is performed on the candidate fire point pixels to obtain the connectivity components of the candidate fire point pixels. Based on connectivity components, spatial topology regularization is performed on candidate fire point pixels to obtain a fire point spatial distribution map of candidate fire point pixels.

[0011] Preferably, in step S4, obtaining the enhancement distribution map of the candidate fire point pixels includes: Based on the continuous response confidence, the difference perception of adjacent nodes in the fire point spatial distribution map is performed to obtain the confidence gradient of adjacent nodes; Based on the confidence gradient, the connection strength of adjacent nodes is dynamically modulated to obtain the adaptive connection strength of adjacent nodes. The adaptive connection strength is regionally balanced to obtain the balanced connection strength of adjacent nodes. Based on the equalization of connection strength, confidence-aggregated subgraphs are extracted from the spatial distribution map of fire points to obtain the core subgraph of the spatial distribution map of fire points. Spatial topological diffusion verification is performed on the core subgraph, and structural fusion is performed on neighboring nodes that meet the diffusion conditions to obtain the enhanced distribution map of candidate fire point pixels.

[0012] Preferably, the process of obtaining the adaptive connection strength of adjacent nodes is as follows: The initial connection relationship of adjacent nodes is resolved to obtain the spatial connection strength of adjacent nodes; Based on the confidence gradient, the collaborative confidence of adjacent nodes is fused to obtain the joint confidence of adjacent nodes; Based on the confidence gradient, the difference between adjacent nodes is quantized to obtain the gradient difference intensity between adjacent nodes; Based on joint confidence and gradient difference strength, nonlinear competitive modulation of spatial connectivity strength is performed to obtain adaptive connectivity strength between neighboring nodes. The formula for calculating adaptive connectivity strength is as follows: ; In the formula, For adaptive connection strength, For spatial connection strength, It is the hyperbolic tangent function. The sensitivity parameters for the preset joint confidence enhancement path, For joint confidence level, It is an exponential function. The sensitivity parameters for the preset gradient decay gating path, For gradient difference intensity, It is the numerical stability constant. The preset gradient influence decision threshold parameter, u is the index identifier of the neighboring node, and v is another index identifier of the neighboring node that is directly connected to node u.

[0013] Preferably, in step S5, obtaining the outline geometric data of the fire area includes: Topological morphology tracing is performed on the reinforcement distribution map to obtain the skeleton nodes of the reinforcement distribution map; Based on the skeleton nodes, closed contour detection is performed on the reinforcement distribution map to obtain the closed boundary loop of the reinforcement distribution map; Based on the closed boundary loop, the loop geometric saliency is extracted from the fire area of ​​the candidate fire point pixels to obtain the main contour loop of the fire area. Adaptive step-size curvature-aware resampling is performed on the main contour loop to obtain the feature point sequence of the main contour loop; Based on the feature point sequence, piecewise spline fitting is performed on the main contour loop to obtain a smooth vector contour of the fire area; Topological verification is performed on the smooth vector profile to obtain the outline geometry data of the fire area.

[0014] Preferably, in step S6, obtaining the fire event of the target forest includes: The contour geometry data is divided into internal spatial grids to obtain the isoparametric sampling grid of the contour geometry data; Based on continuous response confidence and isoparametric sampling grid, spatial autocorrelation analysis is performed on contour geometric data to obtain the consistency coefficient of contour geometric data. The consistency coefficient is subjected to threshold boundary neighborhood extraction to obtain the outer neighborhood band of the contour geometric data; Based on the confidence level of the continuous response, the gradient direction of the surrounding neighborhood band is deduced to obtain the confidence gradient field of the neighborhood band. Based on the confidence gradient field, the conformity of the contour geometric data is evaluated to obtain the gradient conformity rate of the contour geometric data. Based on the gradient matching rate, the contour geometry data is dynamically optimized segment by segment to obtain the optimized fire zone boundary of the target forest. A closure check is performed on the optimized fire area boundary to obtain the fire events of the target forest.

[0015] Preferably, the process of obtaining the gradient matching rate of the contour geometry data is as follows: The closed boundary of the contour geometry data is discretized by arc length parameterization to obtain a discrete boundary point sequence of the closed boundary; Based on the confidence gradient field, gradient vector field correlation is performed on the discrete boundary point sequence to obtain the gradient direction vector of the discrete boundary point. Local surface normal analysis is performed on the discrete boundary point sequence to obtain the unit normal vector of the discrete boundary point sequence; Based on the gradient direction vector and the unit normal vector, the directional consistency of the discrete boundary point sequence is determined, and the boundary points to be optimized in the discrete boundary point sequence are obtained. Based on the gradient direction vector, the position of the boundary point to be optimized is optimized to obtain the optimized boundary point of the discrete boundary point sequence; Based on the confidence gradient field, the closed contour is reconstructed by optimizing the boundary points, and the geometric alignment of the reconstructed boundary points is fused to obtain the comprehensive gradient matching rate of the contour geometric data.

[0016] The present invention has the following beneficial effects: This invention achieves spatial local anomaly analysis of target forest remote sensing images through multi-scale sliding window traversal. It accurately extracts potential anomalous pixels and anomalous spectral features by combining full-band spectral characteristic analysis. Then, it completes multi-scale co-correlation score evaluation by verifying the continuous spectral response at both coarse and fine scales and calculating the intrinsic co-correlation index, thus achieving accurate screening of candidate fire point pixels. At the same time, it obtains continuous response confidence by intensity calibration of candidate fire point pixels, which greatly improves the accuracy and effectiveness of fire point pixel identification and provides a quantitative confidence reference for the fire point screening process, providing a high-quality data foundation for subsequent fire point spatial analysis.

[0017] This invention relies on continuous response confidence to enhance the node connection strength of the fire point spatial distribution map and vectorize the edge contour of the fire area, accurately acquiring the geometric data of the fire contour. Then, through isoparametric sampling grid division and confidence gradient field inference, it completes multi-dimensional evidence fusion decision-making, realizing dynamic optimization and closure verification of the fire boundary. This not only optimizes the integrity of the fire point spatial topology and improves the refinement of the fire contour extraction, but also realizes multi-dimensional data support for fire point event determination, significantly improving the overall efficiency and accuracy of forest fire point identification, and accurately outputting relevant information about fire point events. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, the forest fire point identification method based on multispectral remote sensing and AI edge computing includes the following steps: S1. Perform spatial local anomaly analysis on the multispectral remote sensing image of the target forest to obtain the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image. S2. Based on the abnormal spectral characteristics, the spectral continuous response of potential abnormal pixels is verified to obtain candidate fire point pixels of potential abnormal pixels, and the intensity of candidate fire point pixels is calibrated to obtain the continuous response confidence of candidate fire point pixels. S3. Reconstruct the spatial topological relationship of the candidate fire point pixels to obtain the fire point spatial distribution map of the candidate fire point pixels; S4. Based on the continuous response confidence, the node connection strength of the fire point spatial distribution map is enhanced collaboratively to obtain the enhanced distribution map of candidate fire point pixels. S5. Based on the enhanced distribution map, the edge contour vectorization of the fire area of ​​the candidate fire point pixels is performed to obtain the contour geometric data of the fire area. S6. Based on the contour geometric data, perform multi-dimensional evidence fusion decision-making on the fire area to obtain the fire point events of the target forest.

[0020] In S1, the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image are obtained, including: Multi-scale (e.g., 3×3, 5×5, 7×7) sliding window traversal is performed on the multispectral remote sensing image of the target forest to obtain the local neighborhood radiative statistics of the multispectral remote sensing image. Spatial autocorrelation weighted fusion is performed on the central pixel of the local neighborhood radiation statistics to obtain the expected radiation value of the central pixel; Based on the expected radiation value, the deviation of the actual observed radiation value of the central pixel is measured to obtain the radiation residual intensity of the central pixel; Based on the radiation residual intensity, adaptive threshold segmentation is performed on multispectral remote sensing images to obtain potential anomalous pixels in the multispectral remote sensing images. Full-band spectral characteristic analysis was performed on potential anomalous pixels to obtain the anomalous spectral features of multispectral remote sensing images.

[0021] Forest multispectral remote sensing images can be acquired through a collaborative approach using space-based, airborne, and ground-based platforms equipped with multispectral remote sensing sensors. At the space-based level, low-orbit remote sensing satellites and forestry-specific observation satellites enable periodic, full-coverage image acquisition across large forest areas. At the airborne level, forest fire prevention drones and aerial monitoring aircraft / helicopters conduct flexible patrols or fixed-point, high-resolution real-time image acquisition of key forest areas and high-fire-risk zones. At the ground level, fixed multispectral imaging equipment is deployed at key locations such as forest lookout towers, intelligent monitoring base stations, and mountain access checkpoints to achieve continuous, all-weather image acquisition of core forest areas and key passageways. Image data acquired from various platforms can be transmitted back in real-time via satellite data transmission, wireless communication, and forestry dedicated networks, providing multi-dimensional, multi-resolution raw remote sensing imagery for fire point identification. After obtaining the multispectral remote sensing images, basic preprocessing such as radiometric correction, geometric correction, and atmospheric correction can be performed.

[0022] A multi-scale sliding window traversal operation was performed on the multispectral remote sensing image of the target forest. Rectangular windows of different sizes were used to sequentially translate and cover the image pixel by pixel. The image area covered by each translated window is the local neighborhood of the corresponding central pixel. A comprehensive summary and statistical analysis was conducted on the radiometric values ​​of all pixels in each local neighborhood. The statistical content included the mean, median, mode of the radiometric values, as well as relevant data reflecting the dispersion of the values. The various radiometric statistical results of each local neighborhood were systematically recorded, and the corresponding local neighborhood radiometric statistical data were matched for each central pixel in the image. Finally, the results were integrated to form the local neighborhood radiometric statistics of the entire multispectral remote sensing image.

[0023] Using the central pixel of each local neighborhood as the core, pixels at different spatial locations around the central pixel are selected as spatially associated pixels. Weight values ​​are assigned based on the spatial distance between the associated pixels and the central pixel; the closer the associated pixels are, the higher their weight values ​​are assigned. Combining the radiation-related data from the local neighborhood radiation statistics, the radiation values ​​of each associated pixel are multiplied by their corresponding weight values ​​and summed. The summation result is then normalized. This method determines a corresponding radiation reference value for each central pixel, which is the expected radiation value of the central pixel.

[0024] The radiometric values ​​originally recorded for each central pixel in a multispectral remote sensing image are extracted as the actual observed radiometric values. The difference between the actual observed radiometric value of each central pixel and the expected radiometric value obtained by spatial autocorrelation weighted fusion is calculated. The absolute value of the difference is converted and combined with the data reflecting the dispersion of radiometric values ​​in the local neighborhood radiometric statistics corresponding to the central pixel. The converted absolute value is then standardized and quantitatively characterized to achieve a specific measurement of the degree of deviation between the actual observed radiometric value and the expected radiometric value of each central pixel. The quantitative characterization result obtained after measurement is the radiometric residual intensity of the central pixel.

[0025] The numerical distribution characteristics of the radiometric residual intensity of all central pixels in the entire multispectral remote sensing image are analyzed. Based on the distribution range and concentration of the values, a fixed segmentation threshold is determined. The radiometric residual intensity of each central pixel in the image is compared with the segmentation threshold one by one. Central pixels with radiometric residual intensity values ​​exceeding the segmentation threshold are marked one by one. All marked pixels are screened and integrated as a whole. The integrated pixel set is the potential anomalous pixels of the multispectral remote sensing image.

[0026] For each potential anomalous pixel identified through screening, its corresponding radiometric response values ​​across all spectral bands of the multispectral remote sensing image are comprehensively extracted. The radiometric response values ​​of individual potential anomalous pixels across different spectral bands are systematically analyzed to determine the interrelationships and variation patterns among the radiometric response values ​​in each band. This clarifies the overall distribution characteristics of the band radiometric response values ​​of individual potential anomalous pixels. The band radiometric response characteristics of all potential anomalous pixels are then summarized and organized, extracting the common spectral response characteristics of all potential anomalous pixels while recording the unique spectral response characteristics of some potential anomalous pixels. These common and unique spectral response characteristics are then systematically integrated to form the overall feature set, which constitutes the anomalous spectral characteristics of the multispectral remote sensing image.

[0027] Multi-scale sliding window traversal allows for comprehensive and detailed acquisition of radiometric statistics for each local neighborhood of the image, laying a solid data foundation for subsequent anomaly analysis. Combining spatial autocorrelation-weighted fusion with spatial distance weights ensures that the expected radiometric value of the central pixel closely matches the actual radiometric distribution characteristics of the image. The measurement of radiometric residual intensity accurately characterizes the deviation between actual and expected radiometric values. Determining thresholds and screening potential anomaly pixels based on the numerical characteristics of radiometric residual intensity guarantees the accuracy of potential anomaly pixel extraction. Full-band spectral characteristic analysis of potential anomaly pixels fully extracts their common and unique spectral response features. The resulting anomaly spectral features accurately reflect the spectral essence of potential anomaly pixels, providing precise and effective feature evidence for subsequent continuous spectral response verification.

[0028] In S2, the continuous response confidence of the candidate fire point pixels is obtained, including: The anomalous spectral features are divided into multi-scale bands to obtain coarse-scale band sequences and fine-scale band sequences of the anomalous spectral features. Based on Planck's radiation law, the radiation increment of coarse-scale band sequences is verified, and the macroscopic continuity index of coarse-scale band sequences is obtained. Differential continuity verification was performed on the fine-scale band sequence to obtain the micro-continuity index of the fine-scale band sequence; A synergistic evaluation of macroscopic and microscopic continuity indicators yields multi-scale synergistic scores for potential anomalous pixels, including: We performed an analysis of the inter-scale differences between macro-continuity indicators and micro-continuity indicators to obtain the absolute differences between them. By merging macro-level and micro-level continuous indicators in a balanced manner, a geometric harmonic quantity of macro-level and micro-level continuous indicators is obtained. By coupling the absolute difference quantity and the geometric harmonic quantity, the intrinsic coherence index of potential anomalous pixels is obtained. The formula for calculating the intrinsic coherence index is as follows: ; In the formula, Let be the intrinsic coherence index of the i-th potential anomalous pixel. It is a natural exponential function. The preset collaborative penalty coefficient, For the i-th potential anomalous pixel, the macroscopic continuity index is... For the i-th potential anomalous pixel, the micro-continuity index is... The preset zero-prevention constant is a very small constant; Range normalization mapping is performed on the intrinsic co-indices to obtain the multi-scale co-scores of potential anomalous pixels; Based on multi-scale collaborative scoring, adaptive threshold screening is performed on potential anomalous pixels to obtain candidate fire point pixels of potential anomalous pixels. The candidate fire point pixels are subjected to nonlinear normalization mapping to obtain the continuous response confidence of the candidate fire point pixels.

[0029] All spectral bands of the anomalous spectral features are divided according to wavelength interval and response resolution. Bands with large wavelength intervals and low response resolution are grouped into one category and arranged in the original band order to form an ordered band combination. Bands with small wavelength intervals and high response resolution are grouped into another category and arranged in the original band order to form an ordered band combination. In this way, coarse-scale band sequences and fine-scale band sequences of anomalous spectral features are obtained.

[0030] Based on the inherent correlation between radiation intensity and spectral bands reflected by Planck's radiation law, the radiation intensity values ​​corresponding to each band in the coarse-scale band sequence are extracted. The radiation intensity values ​​of adjacent bands in the coarse-scale band sequence are compared and analyzed in turn to check whether the changing pattern of radiation intensity values ​​with bands conforms to the relevant characteristics of Planck's radiation law. The radiation changing characteristics that conform to the law are systematically quantified and characterized, and the result of this quantification is used as a macroscopic continuity index of the coarse-scale band sequence.

[0031] The radiation intensity values ​​corresponding to each band in the fine-scale band sequence are extracted. The radiation intensity values ​​of adjacent bands in the fine-scale band sequence are then processed by difference to obtain the radiation intensity variation values ​​between each group of adjacent bands. The overall pattern and characteristics of all radiation intensity variation values ​​are sorted out and analyzed to check whether the radiation intensity variation values ​​show continuous variation characteristics. The continuous variation characteristics are systematically quantified and characterized, and the results of the quantification and characterization are used as the microscopic continuity index of the fine-scale band sequence.

[0032] Extract the specific values ​​of the macro-continuity index and micro-continuity index corresponding to each potential abnormal pixel, perform a difference operation on the two values, and then perform an absolute value transformation on the result. The transformed value is the absolute difference between the macro-continuity index and the micro-continuity index.

[0033] Extract the specific values ​​of the macro-continuity index and micro-continuity index corresponding to each potential abnormal pixel, multiply the two values, and then take the square root of the result. The processed value is the geometric harmonic quantity of the macro-continuity index and the micro-continuity index.

[0034] A comprehensive numerical processing method is carried out by combining the absolute difference and geometric harmonics for each potential anomalous pixel. First, the absolute difference is normalized. Then, the normalized absolute difference is weakened by the natural index decay method. The weakened result is multiplied with the geometric harmonics. The final product is the intrinsic harmonic index of the potential anomalous pixel.

[0035] The overall numerical range of intrinsic co-indices of all potential anomalous pixels in the entire multispectral remote sensing image is analyzed. The intrinsic co-indices of each potential anomalous pixel are uniformly mapped to a fixed numerical range, and the standardization process of all intrinsic co-indices is completed. The standardized values ​​are the multi-scale co-scores of potential anomalous pixels.

[0036] The overall numerical distribution characteristics of the multi-scale collaborative scores of all potential anomalous pixels are analyzed. A fixed screening threshold is determined based on the distribution range and concentration of the values. The multi-scale collaborative score of each potential anomalous pixel is compared with the screening threshold one by one. Potential anomalous pixels whose multi-scale collaborative scores exceed the screening threshold are selected. All selected potential anomalous pixels are integrated, and the integrated pixel set is the candidate fire point pixel of potential anomalous pixels.

[0037] The overall numerical range of the multi-scale collaborative scores corresponding to all candidate fire pixels is analyzed. The multi-scale collaborative scores of each candidate fire pixel are adjusted through a non-linear numerical transformation method. The adjusted values ​​are uniformly mapped to a fixed confidence value range. The value after mapping is the continuous response confidence of the candidate fire pixel.

[0038] The macroscopic continuity index is derived from the coarse-scale band sequence obtained by dividing the anomalous spectral features. First, the anomalous spectral features are divided according to wavelength interval and response resolution to obtain the coarse-scale band sequence. Then, the radiation intensity value of each band in the sequence is extracted based on Planck's radiation law. The variation law of radiation intensity of adjacent bands is compared to see if it conforms to Planck's radiation law. The radiation variation characteristics that conform to the law are systematically quantified. The final quantification result is the macroscopic continuity index.

[0039] The micro-continuity index is derived from the fine-scale band sequence obtained by dividing the anomalous spectral features. Similarly, the fine-scale band sequence is first divided, and then the radiation intensity value of each band in the sequence is extracted. The radiation intensity values ​​of adjacent bands are then processed by difference to obtain the radiation intensity change value. The continuous characteristics of all the change values ​​are sorted out and systematically quantified. The final quantification result is the micro-continuity index.

[0040] The co-penalty coefficient is a pre-set fixed value used to adjust the degree of influence of the difference between macroscopic and microscopic continuous indicators on the final result. The zero-minimum constant is a pre-set minimum fixed value used to avoid the denominator being zero in calculations involving the addition of macroscopic and microscopic continuous indicators.

[0041] This formula weakens the difference between macro and micro continuity indicators through a natural exponential function, and combines the product of the two indicators for comprehensive calculation. Its core function is to quantify the degree of synergistic matching of the continuity features of potential anomalous pixels at both macro and micro scales. The generated intrinsic synergistic index can accurately integrate the key information of the dual-scale indicators, providing core and necessary data support for subsequent range normalization mapping of the intrinsic synergistic index to obtain the multi-scale synergistic score of potential anomalous pixels, ensuring that the multi-scale synergistic score can accurately reflect the dual-scale synergistic characteristics of potential anomalous pixels.

[0042] When the numerical difference between macro-level and micro-level continuous indicators decreases, the result of the natural exponential function will stabilize, and the intrinsic synergistic index will remain stable under the influence of this part. When the product of the two indicators increases, the overall value of the intrinsic synergistic index will increase accordingly, and the influence of the difference between the two is reasonably regulated by the decay effect of the natural exponential function, so that the intrinsic synergistic index can always stably reflect the synergistic characteristics of the two-scale indicators, and is not disturbed by the extreme values ​​of a single indicator, thus always maintaining an accurate representation of the synergistic characteristics of the two scales.

[0043] By dividing the anomalous spectral characteristic bands according to wavelength intervals and response resolution, an ordered sequence of coarse and fine-scale bands can be accurately obtained, laying a precise band foundation for subsequent spectral verification. Macroscopic and microscopic continuity verification and quantification indicators are completed based on Planck's radiation law and difference processing, comprehensively reflecting the continuous response characteristics of the spectrum. The absolute difference and geometric harmonicity are obtained by taking the square root of the product of the absolute difference, and the intrinsic synergy index is obtained through comprehensive numerical processing combined with natural exponential decay. Further standardization yields the multi-scale synergy score, accurately integrating dual-scale indicator characteristics. Candidate fire pixels are selected based on the score, and the continuous response confidence level is obtained through nonlinear numerical transformation. This achieves precise selection of fire pixels and provides a quantitative and reliable confidence reference for subsequent fire point analysis.

[0044] Both macroscopic and microscopic continuity indicators are obtained based on the precise classification of anomalous spectral features. Combined with Planck's radiation law or difference processing and quantitative characterization, this ensures that the indicators accurately reflect the spectral continuity characteristics of different scale bands. Preset parameters such as the synergistic penalty coefficient and the zero-minimum constant are used to reasonably adjust for the influence of differences and avoid zero values ​​in the calculation process, ensuring the stability of subsequent calculations. The formula weakens the difference between the two indicators through natural indices and combines multiplication operations to accurately quantify the dual-scale synergistic matching degree of potential anomalous pixels. The generated intrinsic synergistic index provides core support for multi-scale synergistic scoring, and its stable trend avoids interference from extreme values ​​of a single indicator, consistently and accurately representing the dual-scale synergistic characteristics, providing reliable data for the selection of candidate fire pixels.

[0045] In S3, the spatial distribution map of candidate fire point pixels is obtained, including: The spatial distance of the candidate fire point pixels is obtained by measuring the nearest neighbor distance; Based on spatial distance, local density peak detection is performed on candidate fire point pixels to obtain the spatial aggregation center of candidate fire point pixels; By imposing radiation intensity similarity constraints on spatial clustering centers, the coupling relationships of spatial clustering centers are obtained; Based on the coupling relationship, spatial attribution determination is performed on candidate fire point pixels to obtain the regional attribution identifier of the candidate fire point pixels; Based on the region affiliation identifier, connectivity analysis is performed on the candidate fire point pixels to obtain the connectivity components of the candidate fire point pixels. Based on connectivity components, spatial topology regularization is performed on candidate fire point pixels to obtain a fire point spatial distribution map of candidate fire point pixels.

[0046] The spatial coordinates of each candidate fire pixel in the multispectral remote sensing image are determined. These coordinates are converted from the image row and column position corresponding to the pixel to Cartesian coordinates. Then, taking each candidate fire pixel as a reference, the straight-line distance between it and all other candidate fire pixels is calculated one by one. The distance calculation is directly derived from the Cartesian coordinates of the two pixels. All distance results corresponding to each reference pixel are recorded separately. Finally, a set containing the distance information of each candidate fire pixel to all other candidate fire pixels is formed, which is the spatial distance of the candidate fire pixel.

[0047] For each candidate fire point pixel, a fixed-range neighborhood region is defined. The neighborhood region extends outward from the central pixel by a fixed number of pixels. The total number of other candidate fire point pixels contained in the neighborhood region of each central pixel is counted. This number is the local density of the candidate fire point pixel. After calculating the local density of all candidate fire point pixels, the local density of each pixel is compared with the local density of all its surrounding neighboring pixels one by one. Candidate fire point pixels with local density values ​​higher than all their surrounding neighboring pixels are selected. These selected pixels are the spatial cluster centers of the candidate fire point pixels.

[0048] The radiative intensity value corresponding to each spatial cluster center is extracted. This value comes from the original radiative response data of the candidate fire point pixels in the multispectral remote sensing image. Then, for any two spatial cluster centers, the difference between their radiative intensity values ​​is calculated. A fixed range of radiative intensity similarity is preset, and it is determined whether the calculated difference is within this range. If the difference is within the range, the two spatial cluster centers are determined to have similar radiative intensity. All spatial cluster centers with similarity are paired and the correspondence of each pair of cluster centers is recorded. The correspondence of all the marked centers together constitutes the coupling relationship of the spatial cluster centers.

[0049] Based on the coupling relationship of spatial cluster centers, interconnected spatial cluster centers are divided into clusters. The spatial coverage of each cluster is defined. Taking the cluster as the unit, the spatial distance from each candidate fire point pixel to all spatial cluster centers in the cluster is calculated. The distances from each candidate fire point pixel to each cluster center in the cluster are compared, and the candidate fire point pixel is assigned to the region corresponding to the spatial cluster center with the smallest distance. Each candidate fire point pixel is assigned a unique identifier corresponding to its region. This identifier is the region affiliation identifier of the candidate fire point pixel.

[0050] Based on the different regional attribution identifiers, all candidate fire pixels are divided into multiple independent groups. Candidate fire pixels within the same group have the same regional attribution identifier. Within each group, it is determined whether any two candidate fire pixels have a spatial adjacency relationship. The criterion for adjacency relationship is whether the two pixels share an edge or corner in the image. If an adjacency relationship exists, the two are considered connected. Following the connectivity relationship step by step, all interconnected candidate fire pixels within a group are integrated into an independent group. Each such independent group is the connectivity component of the candidate fire pixel. All connectivity components within each group are fully recorded.

[0051] For each connectivity component, the spatial relationships of all candidate fire point pixels are analyzed, clarifying the relative position of each pixel within the component and its spatial relationships with other pixels, such as adjacency and parallelism. The pixels within the component are arranged and regularized according to the logical order of spatial position to construct the spatial topology of each connectivity component. This structure clearly presents the distribution pattern and association mode of pixels within the component. Subsequently, the topologies of all connectivity components are integrated into the same image spatial framework, and the spatial position, coverage area, and association details of internal pixels corresponding to each topology are labeled. Finally, a fire point spatial distribution map that can completely reflect the spatial distribution pattern of candidate fire point pixels is formed.

[0052] By transforming image rows and columns and deriving straight-line distances, the spatial distance between each candidate fire point pixel and other pixels is accurately obtained, providing precise data support for subsequent spatial analysis. Fixed-range neighborhood statistics and density comparisons are used to screen spatial cluster centers, and the coupling relationship is determined by combining radiation intensity differences to ensure the effective association of similar cluster centers. Region identifiers are assigned according to distance, and connectivity analysis is conducted in groups to obtain connectivity components, clarifying pixel spatial relationships. Through topological regularization and integration, a fire point spatial distribution map is formed that fully reflects the spatial distribution pattern of candidate fire point pixels, clearly presenting the pixel distribution morphology and related details, providing a reliable spatial distribution basis for subsequent fire point identification and related processing.

[0053] In S4, the enhancement distribution map of the candidate fire point pixels is obtained, including: Based on the continuous response confidence, the difference perception of adjacent nodes in the fire point spatial distribution map is performed to obtain the confidence gradient of adjacent nodes; Based on the confidence gradient, the connection strength between neighboring nodes is dynamically modulated to obtain the adaptive connection strength between neighboring nodes, including: The initial connection relationship of adjacent nodes is resolved to obtain the spatial connection strength of adjacent nodes; Based on the confidence gradient, the collaborative confidence of adjacent nodes is fused to obtain the joint confidence of adjacent nodes; Based on the confidence gradient, the difference between adjacent nodes is quantized to obtain the gradient difference intensity between adjacent nodes; Based on joint confidence and gradient difference strength, nonlinear competitive modulation of spatial connectivity strength is performed to obtain adaptive connectivity strength between neighboring nodes. The formula for calculating adaptive connectivity strength is as follows: ; In the formula, For adaptive connection strength, For spatial connection strength, It is the hyperbolic tangent function. The sensitivity parameters for the preset joint confidence enhancement path, For joint confidence level, It is an exponential function. The sensitivity parameters for the preset gradient decay gating path, For gradient difference intensity, It is the numerical stability constant. The preset gradient influence decision threshold parameter, u is the index identifier of the neighboring node, and v is another index identifier of the neighboring node that is directly connected to node u. The adaptive connection strength is regionally balanced to obtain the balanced connection strength of adjacent nodes. Based on the equalization of connection strength, confidence-aggregated subgraphs are extracted from the spatial distribution map of fire points to obtain the core subgraph of the spatial distribution map of fire points. Spatial topological diffusion verification is performed on the core subgraph, and structural fusion is performed on neighboring nodes that meet the diffusion conditions to obtain the enhanced distribution map of candidate fire point pixels.

[0054] The continuous response confidence value corresponding to each node in the spatial distribution map of fire points is determined. The continuous response confidence value of a node comes from the intensity calibration results of the candidate fire point pixels. Each node is identified as a directly connected neighboring node in the spatial topology. The continuous response confidence value of each node is compared with the continuous response confidence value of its neighboring nodes one by one. The confidence value of the current node is subtracted from the confidence value of the neighboring nodes to obtain the confidence difference between each node and its corresponding neighboring nodes. These differences are arranged in the order of connection between nodes, and the ordered set of differences is the confidence gradient of the neighboring nodes.

[0055] By analyzing the spatial relationships of all nodes in the fire point spatial distribution map, it is determined whether there are direct spatial connections between the nodes. For adjacent node pairs with direct connections, the initial connection value is set based on the spatial distance between the nodes in the image and the smoothness of the connection path. Adjacent node pairs that are close in space and have unobstructed connection paths are assigned higher initial connection values, while adjacent node pairs that are far in space or have indirect connections in the connection paths are assigned lower initial connection values. The initial connection value corresponding to each adjacent node pair is the spatial connection strength of the adjacent nodes.

[0056] Extract the continuous response confidence values ​​of each adjacent node, and combine them with the corresponding confidence gradient directions to determine whether the gradient direction points to an increasing confidence trend. If the gradient direction indicates an increasing confidence, assign a higher fusion weight to the high-confidence node; if the gradient direction indicates a decreasing confidence, assign a lower fusion weight to the low-confidence node. Multiply the confidence values ​​of the two nodes by their corresponding weights and sum the results. The sum is the joint confidence of the adjacent nodes.

[0057] Calculate the absolute difference between the confidence scores of consecutive responses of adjacent nodes, record the specific magnitude of the difference, and observe the changes in the confidence gradient between adjacent nodes. Statistically analyze whether the gradient direction changes or whether the gradient value changes abruptly. Combine the absolute difference with the gradient change characteristics to reflect the degree of correlation between the two in a unified numerical form. This representation result is the gradient difference strength between adjacent nodes.

[0058] Based on the spatial connection strength of adjacent nodes, the connection strength is first enhanced according to the joint confidence level. If the joint confidence level is high, the spatial connection strength is increased by a fixed proportion. If the joint confidence level is low, the original spatial connection strength is maintained or slightly adjusted. Then, the connection strength is suppressed according to the gradient difference strength. If the gradient difference strength is large, the connection strength value after enhancement is reduced by a fixed proportion. If the gradient difference strength is small, the enhanced value is not suppressed. Through this competitive balance between enhancement and suppression, the nonlinear numerical adjustment of the spatial connection strength is completed. The final value is the adaptive connection strength of adjacent nodes.

[0059] The fire point spatial distribution map is divided into multiple continuous and non-overlapping regions according to spatial location. Each region contains a certain number of adjacent node pairs. The adaptive connection strength values ​​of all adjacent nodes in each region are counted. The average value of all adaptive connection strengths in the region is calculated. The adaptive connection strength of each adjacent node is compared with the average value of its region. The adaptive connection strengths that are too high above the average value are appropriately lowered, and the adaptive connection strengths that are too low below the average value are appropriately increased, so that the connection strengths of all adjacent nodes in the region are kept within a similar value range. The adjusted connection strength is the balanced connection strength of adjacent nodes.

[0060] A fixed connection strength screening threshold is set, which is determined based on the overall distribution of the balanced connection strength of all adjacent nodes. The relationship between the balanced connection strength of each adjacent node and the screening threshold is compared one by one. Adjacent node pairs whose balanced connection strength exceeds the screening threshold are retained. All nodes involved in these node pairs are integrated, and the spatial topology relationship between the integrated nodes is sorted out to ensure that the nodes form interconnected clusters through high connection strength. The subgraph formed by these interconnected node clusters is the core subgraph of the fire point spatial distribution map.

[0061] Using the boundary nodes of the core subgraph as the diffusion starting point, a fixed spatial diffusion range is set. This range extends outward from the boundary nodes by a fixed number of pixels. Neighboring nodes within the diffusion range that are not included in the core subgraph are identified. The confidence level of the continuous response of these neighboring nodes is checked to see if it meets the preset diffusion conditions. The preset diffusion conditions are that the confidence level of the continuous response of the neighboring nodes is not lower than a fixed proportion of the confidence level of the boundary nodes of the core subgraph. For neighboring nodes that meet this condition, a connection relationship is established between them and the boundary nodes of the core subgraph. These nodes are integrated into the spatial topology of the core subgraph to form a more closely connected and more comprehensive distribution map. This distribution map is the enhanced distribution map of candidate fire point pixels.

[0062] Spatial connectivity strength is derived from the analysis of the initial connectivity relationships between adjacent nodes. This involves analyzing the spatial positional relationships of nodes in the fire point spatial distribution map, clarifying whether direct spatial connectivity exists, and setting initial connectivity values ​​based on the spatial distance between nodes in the image and the smoothness of the connectivity path, thus forming the spatial connectivity strength. Joint confidence is derived from the fusion of the collaborative confidence of adjacent nodes. The continuous response confidence values ​​of each adjacent node are extracted, and corresponding fusion weights are assigned based on the confidence gradient direction. The joint confidence is obtained by multiplying the confidence values ​​by the weights and summing the results. Gradient difference strength is derived from the quantification of differences between adjacent nodes. The absolute difference between the confidence values ​​of adjacent nodes is calculated, and a comprehensive characterization is performed based on gradient direction inflection, numerical abrupt changes, and other change characteristics, forming the gradient difference strength. The sensitivity parameter for the joint confidence enhancement path is a preset fixed value used to adjust the enhancement magnitude of the connectivity strength by the joint confidence. The sensitivity parameter for the gradient decay gated path is a preset fixed value used to adjust the decay force of the connectivity strength by the gradient difference strength. The numerical stability constant is a preset minimum fixed value used to avoid zero denominators during calculation. The gradient influence decision threshold parameter is a preset fixed value used to define the boundary of the influence of gradient difference intensity on connection strength. The index identifier of adjacent nodes is a unique label for each node in the fire point spatial distribution map, and directly connected nodes are assigned corresponding index identifiers for distinction.

[0063] This formula is based on spatial connectivity strength. It processes the joint confidence using a hyperbolic tangent function to strengthen the connectivity between nodes with high confidence. At the same time, it modulates the influence of gradient difference strength through an exponential function to weaken the connectivity between nodes with excessive differences. This achieves dynamic modulation of spatial connectivity strength. The resulting adaptive connectivity strength can accurately reflect the actual degree of correlation between adjacent nodes under confidence synergy and difference characteristics. This provides accurate and realistic basic data for subsequent regional balancing of adjacent node connectivity strength.

[0064] As the joint confidence increases, the result processed by the hyperbolic tangent function shows a stable strengthening trend, driving the adaptive connection strength to adjust towards a direction that aligns with high-confidence associations. Conversely, as the gradient difference intensities increase, the result processed by the exponential function produces a stable suppression effect, preventing the adaptive connection strength from being over-enhanced. Through the synergistic effect of these two mechanisms, the adaptive connection strength consistently matches the actual association characteristics of adjacent nodes, avoiding both excessive enhancement due to high joint confidence and excessive suppression due to large gradient differences, thus maintaining a precise fit to the true association state between nodes.

[0065] By comparing node confidence scores and arranging the differences in an orderly manner, the confidence gradients of adjacent nodes are accurately obtained, providing a reliable basis for subsequent connection strength modulation. Spatial connection strength is set based on spatial distance and connectivity paths. Combined with gradient direction, fusion weights are assigned to obtain joint confidence scores. Gradient difference strength is formed by integrating differences and gradient features. An adaptive connection strength is obtained through competitive balance optimization of enhancement and suppression, ensuring that the connection strength closely matches the actual association characteristics of the nodes. Regional balancing processing maintains the consistency of connection strength across intervals. Clusters of nodes with high connection strength are selected to obtain the core subgraph. Then, through spatial diffusion fusion of neighboring nodes that meet the conditions, a tightly connected and comprehensively covered enhanced distribution map is formed, providing an accurate and complete spatial distribution foundation for subsequent fire scene contour extraction.

[0066] Spatial connectivity strength is set based on the spatial relationship between nodes and the smoothness of the connectivity path. Joint confidence is obtained by summing the confidence gradient direction with corresponding weights. Gradient difference strength is formed by combining the absolute difference of confidence and gradient change characteristics. All three methods are tailored to the actual association characteristics of nodes, ensuring accurate basic data. Preset parameters adjust the enhancement magnitude of joint confidence, the attenuation of gradient difference strength, avoid zero denominators in calculations, and define the boundaries of gradient influence, ensuring a stable and reasonable calculation process. The formula strengthens the association of high-confidence nodes through a hyperbolic tangent function and regulates the influence of differences through an exponential function. The dynamically modulated adaptive connectivity strength accurately reflects the actual association of nodes, providing a reliable foundation for subsequent regional balancing of adjacent node connectivity strength and helping to improve the accuracy of fire point identification processing.

[0067] In S5, the outline geometric data of the fire area is obtained, including: Topological morphology tracing is performed on the reinforcement distribution map to obtain the skeleton nodes of the reinforcement distribution map; Based on the skeleton nodes, closed contour detection is performed on the reinforcement distribution map to obtain the closed boundary loop of the reinforcement distribution map; Based on the closed boundary loop, the loop geometric saliency is extracted from the fire area of ​​the candidate fire point pixels to obtain the main contour loop of the fire area. Adaptive step-size curvature-aware resampling is performed on the main contour loop to obtain the feature point sequence of the main contour loop; Based on the feature point sequence, piecewise spline fitting is performed on the main contour loop to obtain a smooth vector contour of the fire area; Topological verification is performed on the smooth vector profile to obtain the outline geometry data of the fire area.

[0068] Starting from the edge cells of the enhanced distribution map, trace point by point along the spatial connectivity between shared edges or corners. First, mark all edge cells adjacent to non-fire point cells on the periphery of the distribution area. Then, extend from these edge cells into the inner area to find core cells that can connect different edge segments and have connectivity with at least 4 surrounding cells. These core cells should be able to reflect the overall branching and connection pattern of the distribution area. Systematically record all core cells that meet the conditions according to their positions in the distribution area. These recorded core cells are the skeleton nodes of the enhanced distribution map.

[0069] Starting from each skeleton node, the search proceeds clockwise through eight adjacent pixels in each of its eight directions, prioritizing those that share an edge with the current pixel. The planar coordinates of each found pixel are recorded, and the search path is extended along the adjacent pixels. Each extended pixel is checked to ensure it coincides with the coordinates of the starting skeleton node, until the coordinates of the end point of the search path are completely consistent with the coordinates of the starting skeleton node, forming a complete closed path. The same search operation is performed on all skeleton nodes, and the pixels in each closed path are connected sequentially according to the search order. The resulting closed loop structure is the closed boundary loop of the enhanced distribution map.

[0070] The number of candidate fire point pixels enclosed by each closed boundary loop is counted. The perimeter is determined by the total number of pixels in the loop, and the side length of each pixel is calculated as a fixed length. The area is determined by counting the number of complete pixels contained in the loop. The geometric regularity of the loop is analyzed by the ratio of perimeter to area. The closed boundary loop with the most candidate fire point pixels and a perimeter-to-area ratio within a reasonable range is selected. Loops with fewer than 10 enclosed pixels or irregular geometric shapes are excluded. The selected closed boundary loops that meet the conditions are determined as the main outline loops of the fire area.

[0071] First, analyze the curvature change of the main contour loop. By comparing the directional differences between the coordinates of two adjacent pixels in the loop, determine the directional angle. When the directional angle is greater than 45 degrees, it is determined to be a position with a large degree of curvature. In such positions, a small step size is used to select sampling points, selecting one sampling point every 1 pixel and recording its coordinates. When the directional angle is less than or equal to 45 degrees, it is determined to be a position with a small degree of curvature. In such positions, a large step size is used to select sampling points, selecting one sampling point every 3 pixels and recording its coordinates. Arrange all the selected sampling points according to their order in the main contour loop. The resulting ordered set of coordinate points is the feature point sequence of the main contour loop.

[0072] The feature point sequence is segmented based on the coordinate distance and direction changes of adjacent feature points. The straight-line distance and directional angle between two adjacent feature points are compared. When the directional angle between two adjacent points exceeds 15 degrees, the boundary of the segment is defined between these two feature points, dividing the feature point sequence into several continuous segments. For the feature points in each segment, a smooth curve is drawn in a way that passes through all feature points in the segment and the curve has no obvious turning point. This ensures that the curves of adjacent segments have the same direction at the segment boundary and a natural transition. The smooth curves of all segments are connected into a complete closed figure, which is the smooth vector outline of the fire area.

[0073] Check if the starting and ending coordinates of the smooth vector profile are exactly the same. If they are different, adjust the curvature of the curve at the end of the profile to make the ending coordinates coincide with the starting coordinates, ensuring that the profile forms a closed structure. Then compare the coordinates of the profile curves segment by segment to check if there are any self-intersections where the curve coordinates of different segments overlap or intersect. If self-intersections are found, fine-tune the coordinate parameters of the curves of the self-intersecting segments and change the curve direction to eliminate the intersection. After the closure and non-self-intersection checks are passed, record the planar coordinates of all vertices of the smooth vector profile, the total perimeter of the profile, and the area of ​​the region enclosed by the profile. These recorded coordinate information and geometric parameters are the profile geometric data of the fire area.

[0074] By tracing the spatial connectivity of edge pixels in the enhanced distribution map point by point, marking the outer edge pixels and extending inward to find the core pixels, the skeletal nodes reflecting the branching and connection patterns of the distribution area can be accurately located, providing a clear core localization basis for subsequent contour detection. By statistically analyzing the number of candidate fire point pixels enclosed by closed boundary loops, calculating their perimeter and area, and analyzing their ratios, the main contour loops that meet the criteria are selected. This effectively eliminates interfering loops with irregular geometric shapes, ensuring that the main contour loops accurately correspond to the core fire area. Analyzing the curvature changes of the main contour loops and adjusting the sampling step size according to the degree of curvature to select sampling points, the resulting feature point sequence can accurately capture key information at different curvature points of the contour, providing high-quality data support for subsequent curve fitting.

[0075] The feature point sequence is segmented based on coordinate distance and direction changes. A smooth curve is used to fit each segment, ensuring natural transitions between adjacent segments. The resulting smooth vector contour accurately recreates the actual shape of the fire area, avoiding obvious transitions. The smooth vector contour is then checked for closure and non-self-intersection, and vertex coordinates, total perimeter, and enclosed area are recorded. The resulting contour geometric data is complete and accurate, providing reliable geometric information support for subsequent multi-dimensional evidence fusion decision-making in the fire area, thus improving the accuracy and effectiveness of forest fire point identification.

[0076] In S6, the fire event in the target forest is obtained, including: The contour geometry data is divided into internal spatial grids to obtain the isoparametric sampling grid of the contour geometry data; Based on continuous response confidence and isoparametric sampling grid, spatial autocorrelation analysis is performed on contour geometric data to obtain the consistency coefficient of contour geometric data. The consistency coefficient is subjected to threshold boundary neighborhood extraction to obtain the outer neighborhood band of the contour geometric data; Based on the confidence level of the continuous response, the gradient direction of the surrounding neighborhood band is deduced to obtain the confidence gradient field of the neighborhood band. Based on the confidence gradient field, the conformity of the contour geometric data is evaluated to obtain the gradient conformity rate of the contour geometric data. Based on the gradient matching rate, the contour geometry data is dynamically optimized segment by segment to obtain the optimized fire zone boundary of the target forest. A closure check is performed on the optimized fire area boundary to obtain the fire events of the target forest.

[0077] Based on the confidence gradient field, the conformity of the contour geometric data is evaluated to obtain the gradient conformity rate of the contour geometric data, including: The closed boundary of the contour geometry data is discretized by arc length parameterization to obtain a discrete boundary point sequence of the closed boundary; Based on the confidence gradient field, gradient vector field correlation is performed on the discrete boundary point sequence to obtain the gradient direction vector of the discrete boundary point. Local surface normal analysis is performed on the discrete boundary point sequence to obtain the unit normal vector of the discrete boundary point sequence; Based on the gradient direction vector and the unit normal vector, the directional consistency of the discrete boundary point sequence is determined, and the boundary points to be optimized in the discrete boundary point sequence are obtained. Based on the gradient direction vector, the position of the boundary point to be optimized is optimized to obtain the optimized boundary point of the discrete boundary point sequence; Based on the confidence gradient field, the closed contour is reconstructed by optimizing the boundary points, and the geometric alignment of the reconstructed boundary points is fused to obtain the comprehensive gradient matching rate of the contour geometric data.

[0078] Using the closed boundary of the contour geometry data as the range, determine the minimum and maximum planar coordinates of the area enclosed by the boundary. Divide the area into uniform square grids with a fixed row and column spacing, keeping the side length of the grids uniform. The coordinates of the four vertices of each grid are calculated by superimposing the minimum coordinates with the row and column spacing, ensuring that all grids completely cover the interior space of the contour without overlap. Record the vertex coordinates and row and column position information of all grids in a systematic way. The resulting set of regular grids is the isoparametric sampling grid of the contour geometry data.

[0079] Extract the continuous response confidence of all candidate fire point pixels within each isoparametric sampling grid, calculate the average value of the built-in confidence of each grid, and compare the average confidence of each grid with the average confidence of its four adjacent grids (up, down, left, and right) one by one to analyze the similarity of confidence values ​​between adjacent grids. This similarity is represented by a unified numerical form; the higher the similarity, the larger the representation value, and the lower the similarity, the smaller the representation value. The representation value corresponding to each grid is the consistency coefficient of the contour geometric data.

[0080] The consistency coefficient values ​​of all isoparametric sampling grids are analyzed. A fixed extraction threshold is determined based on the concentration range of the values. The consistency coefficient of each grid is compared with the threshold. Grids with consistency coefficient values ​​within a fixed range near the threshold are selected. These grids are connected in space to form a ring-shaped area around the core area of ​​the contour geometric data. This ring-shaped area is the outer neighborhood of the contour geometric data.

[0081] The average confidence score of the continuous response corresponding to each grid in the outer neighborhood band is extracted. Taking each grid as the center, its confidence score is compared with the confidence scores of the four adjacent grids above, below, left, and right to determine the direction of increase or decrease in confidence score. A vector pointing in the direction of change is used to represent the confidence score change trend of the grid. The confidence score change vectors of all grids in the outer neighborhood band are integrated according to their spatial positions to form an overall field that can comprehensively reflect the direction and intensity of confidence score change within the neighborhood band. This field is the confidence score gradient field of the neighborhood band.

[0082] Along the arc length direction of the closed boundary of the contour geometry data, uniform sampling is performed at fixed length intervals. Starting from the starting point of the boundary, a sampling point is selected at fixed intervals, and the planar coordinates of each sampling point are recorded. During the sampling process, it is ensured that the sampling points are evenly distributed on the entire closed boundary until the sampling points cover the entire closed boundary and return to the vicinity of the starting point. All sampling points are arranged in order of their sequence on the boundary, and the resulting ordered set of coordinate points is the discrete boundary point sequence of the closed boundary.

[0083] For each sampling point in the discrete boundary point sequence, determine its spatial location in the confidence gradient field, find the surrounding neighborhood grid to which the location belongs, extract the confidence change vector corresponding to the grid, and directly assign the vector to the corresponding sampling point, so that each discrete boundary point obtains a direction vector that matches the gradient field at its location. This direction vector is the gradient direction vector of the discrete boundary point.

[0084] For each sampling point in the discrete boundary point sequence, a fixed number of adjacent sampling points are selected. A local smooth surface is fitted using the coordinates of these adjacent points. The normal direction of this surface at the current sampling point, which is perpendicular to the tangent direction of the surface, is determined. The vector of this normal direction is normalized so that the length of the vector is uniformly a fixed value. The processed vector is the unit normal vector of the discrete boundary point sequence.

[0085] The gradient direction vector of each sampling point in the discrete boundary point sequence is compared with the unit normal vector. The consistency of direction is determined by observing whether the two vectors tend to be consistent. If the angle between the two vectors exceeds a fixed angle, the direction of the sampling point is determined to be inconsistent. All sampling points with inconsistent directions are marked. The set of marked sampling points is the boundary point to be optimized in the discrete boundary point sequence.

[0086] For each boundary point to be optimized, the coordinate position is adjusted according to its corresponding gradient direction vector. The coordinates of the point to be optimized are moved along the direction of the gradient direction vector. The moving distance is determined according to the distance between the point and the adjacent optimization points. This ensures that the angle between the unit normal vector and the gradient direction vector of the point is within a fixed range after the movement. The adjusted sampling point is the optimized boundary point of the discrete boundary point sequence.

[0087] All optimized boundary points are rearranged according to their order in the original discrete boundary point sequence and connected to form a new closed contour. The degree of fit between the gradient direction vector and the unit normal vector of each optimized boundary point in the new contour is checked one by one. The proportion of points that meet the fit is counted. At the same time, the overall smoothness of the new contour is combined for comprehensive characterization. The resulting quantitative characterization result is the comprehensive gradient fit rate of the contour geometric data.

[0088] Based on the numerical distribution of the comprehensive gradient matching rate, the closed boundary of the contour geometric data is divided into several continuous segments of fixed length. The gradient matching rate of each segment is evaluated one by one. For segments whose matching rate does not meet the fixed standard, the coordinate positions of the optimized boundary points in the segment are adjusted to improve the directional matching degree of all points in the segment until the gradient matching rate of the segment reaches the standard. All adjusted segments are then reconnected and integrated to form a complete closed boundary, which is the optimized fire area boundary of the target forest.

[0089] Check whether the starting point coordinates and ending point coordinates of the optimized fire area boundary are completely coincident. If there is a deviation, fine-tune the coordinates of the boundary points near the ending point to make the ending point coordinates coincide precisely with the starting point coordinates, ensuring that the boundary forms a complete closed structure. At the same time, check whether there are self-intersections or breaks in the boundary. If so, make corrections. After the closure and integrity meet the requirements, record the coordinates of all vertices of the optimized fire area boundary, the total length of the boundary, and the area information of the enclosed area. This information together constitutes the fire event of the target forest.

[0090] By dividing the sampling grid with fixed row and column spacing, complete coverage of the internal space of the contour is ensured without overlap, providing a well-organized data foundation for subsequent spatial analysis. The consistency coefficient, obtained by calculating the average built-in confidence of the grid and comparing the similarity of adjacent grids, accurately characterizes the spatial correlation characteristics of the confidence level. Grids near a selected threshold form an outer neighborhood zone, and the confidence gradient field derived from the confidence difference comprehensively reflects the trend of the built-in confidence of the neighborhood zone. The discrete boundary point sequence obtained by uniform sampling, combined with the comparison and optimization of the gradient direction vector and the unit normal vector, improves the directional fit of the boundary points. The comprehensive gradient matching rate quantitatively characterizes the contour fit and smoothness. Boundary segments are adjusted segment by segment until they meet the standards and undergo closure verification. The final optimized fire area boundary is complete and accurate. The recorded coordinates, length, and area information constitute the fire event, providing a comprehensive and reliable decision-making basis for forest fire identification, effectively improving the accuracy and completeness of fire identification.

[0091] Each step can be streamlined and implemented in a lightweight manner through AI edge computing: S1 performs multi-scale analysis and quantification of spatial local anomalies in multispectral remote sensing images. Relying on lightweight anomaly detection AI models deployed at the edge (such as pruned convolutional autoencoders and lightweight isolated forests), it quickly extracts potential anomalous pixels and anomalous spectral features based on radiometric residual analysis and full-band spectral analysis. S2 performs multi-scale continuity assessment and continuity response verification based on anomalous spectral features. It relies on lightweight spectral classification models deployed at the edge (such as lightweight multilayer perceptron and MobileNet architecture variants) to complete candidate fire point pixel screening and continuity response confidence calibration. S3 reconstructs the spatial topological relationships of candidate fire point pixels, and generates a spatial distribution map of fire points by means of spatial distance constraints, density peak clustering and connectivity verification, relying on an efficient geometric computing library at the edge. S4 collaboratively enhances the node connection strength of the fire point spatial distribution map based on continuous response confidence, and uses edge-end lightweight graph neural network (GNN) to strengthen the spatial distribution structure of candidate fire point pixels; Based on the enhanced spatial distribution map of fire points, S5 performs vectorization processing of the edge contour of the fire area, and completes contour skeleton extraction, spline fitting and geometric feature verification by relying on edge geometric calculation tools, and outputs contour geometric data. Based on contour geometric data and confidence gradient field information, S6 performs multi-dimensional evidence fusion decision-making. It integrates multi-dimensional information through a lightweight edge fusion model (quantized decision tree, lightweight attention network) and finally outputs the fire event of the target forest. The entire process is adapted to edge computing power constraints, realizing low-latency, localized intelligent monitoring of forest fires.

[0092] Example 2: A forest fire detection device based on multispectral remote sensing and AI edge computing, comprising: One or more processors; Memory, used to store one or more computer programs; When one or more programs are executed by one or more processors, the one or more processors execute the method in Example 1.

[0093] Example 3: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method in Example 1.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A forest fire point identification method based on multispectral remote sensing and AI edge computing, characterized by the following steps: include: S1. Perform spatial local anomaly analysis on the multispectral remote sensing image of the target forest to obtain the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image. S2. Based on anomalous spectral features, perform continuous spectral response verification on potential anomalous pixels to obtain candidate fire point pixels. Then, calibrate the intensity of the candidate fire point pixels to obtain their continuous response confidence levels, including: The anomalous spectral features are divided into multi-scale bands to obtain coarse-scale band sequences and fine-scale band sequences of the anomalous spectral features. Based on Planck's radiation law, the radiation increment of coarse-scale band sequences is verified, and the macroscopic continuity index of coarse-scale band sequences is obtained. Differential continuity verification was performed on the fine-scale band sequence to obtain the micro-continuity index of the fine-scale band sequence; A synergistic evaluation of the results of macro-continuity indicators and micro-continuity indicators was conducted to obtain multi-scale synergistic scores for potential anomalous pixels. Based on multi-scale collaborative scoring, adaptive threshold screening is performed on potential anomalous pixels to obtain candidate fire point pixels of potential anomalous pixels. The candidate fire point pixels are nonlinearly normalized to obtain the continuous response confidence of the candidate fire point pixels; S3. Reconstruct the spatial topological relationship of the candidate fire point pixels to obtain the fire point spatial distribution map of the candidate fire point pixels; S4. Based on the continuous response confidence level, the node connectivity strength of the fire point spatial distribution map is enhanced collaboratively to obtain the enhanced distribution map of candidate fire point pixels, including: Based on the continuous response confidence, the difference perception of adjacent nodes in the fire point spatial distribution map is performed to obtain the confidence gradient of adjacent nodes; Based on the confidence gradient, the connection strength of adjacent nodes is dynamically modulated to obtain the adaptive connection strength of adjacent nodes. The adaptive connection strength is regionally balanced to obtain the balanced connection strength of adjacent nodes. Based on the equalization of connection strength, confidence-aggregated subgraphs are extracted from the spatial distribution map of fire points to obtain the core subgraph of the spatial distribution map of fire points. Spatial topological diffusion verification is performed on the core subgraph, and structural fusion is performed on neighboring nodes that meet the diffusion conditions to obtain the enhanced distribution map of candidate fire point pixels. S5. Based on the enhanced distribution map, the edge contour vectorization of the fire area of ​​the candidate fire point pixels is performed to obtain the contour geometric data of the fire area. S6. Based on the contour geometric data, perform multi-dimensional evidence fusion decision-making on the fire area to obtain the fire point events of the target forest.

2. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, In S1, the potential anomalous pixels and anomalous spectral features of the multispectral remote sensing image are obtained, including: Multi-scale sliding window traversal is performed on the multispectral remote sensing image of the target forest to obtain the local neighborhood radiative statistics of the multispectral remote sensing image. Spatial autocorrelation weighted fusion is performed on the central pixel of the local neighborhood radiation statistics to obtain the expected radiation value of the central pixel; Based on the expected radiation value, the deviation of the actual observed radiation value of the central pixel is measured to obtain the radiation residual intensity of the central pixel; Based on the radiation residual intensity, adaptive threshold segmentation is performed on multispectral remote sensing images to obtain potential anomalous pixels in the multispectral remote sensing images. Full-band spectral characteristic analysis was performed on potential anomalous pixels to obtain the anomalous spectral features of multispectral remote sensing images.

3. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, The process of obtaining the multi-scale co-scoring of potential anomalous pixels is as follows: We performed an analysis of the inter-scale differences between macro-continuity indicators and micro-continuity indicators to obtain the absolute differences between them. By merging macro-level and micro-level continuous indicators in a balanced manner, a geometric harmonic quantity of macro-level and micro-level continuous indicators is obtained. By coupling the absolute difference quantity and the geometric harmonic quantity, the intrinsic coherence index of potential anomalous pixels is obtained. The formula for calculating the intrinsic coherence index is as follows: ； In the formula, Let be the intrinsic coherence index of the i-th potential anomalous pixel. It is a natural exponential function. The preset collaborative penalty coefficient, For the i-th potential anomalous pixel, the macroscopic continuity index is... For the i-th potential anomalous pixel, the micro-continuity index is... This is a preset zero-prevention constant; By performing range normalization mapping on the intrinsic co-indices, multi-scale co-scores of potential anomalous pixels are obtained.

4. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, In step S3, the spatial distribution map of candidate fire point pixels is obtained, including: The spatial distance of the candidate fire point pixels is obtained by measuring the nearest neighbor distance; Based on spatial distance, local density peak detection is performed on candidate fire point pixels to obtain the spatial aggregation center of candidate fire point pixels; By imposing radiation intensity similarity constraints on spatial clustering centers, the coupling relationships of spatial clustering centers are obtained; Based on the coupling relationship, spatial attribution determination is performed on candidate fire point pixels to obtain the regional attribution identifier of the candidate fire point pixels; Based on the region affiliation identifier, connectivity analysis is performed on the candidate fire point pixels to obtain the connectivity components of the candidate fire point pixels. Based on connectivity components, spatial topology regularization is performed on candidate fire point pixels to obtain a fire point spatial distribution map of candidate fire point pixels.

5. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, The process of obtaining the adaptive connection strength of adjacent nodes is as follows: The initial connection relationship of adjacent nodes is resolved to obtain the spatial connection strength of adjacent nodes; Based on the confidence gradient, the collaborative confidence of adjacent nodes is fused to obtain the joint confidence of adjacent nodes; Based on the confidence gradient, the difference between adjacent nodes is quantized to obtain the gradient difference intensity between adjacent nodes; Based on joint confidence and gradient difference strength, nonlinear competitive modulation of spatial connectivity strength is performed to obtain adaptive connectivity strength between neighboring nodes. The formula for calculating adaptive connectivity strength is as follows: ； In the formula, For adaptive connection strength, For spatial connection strength, It is the hyperbolic tangent function. The sensitivity parameters for the preset joint confidence enhancement path, For joint confidence level, It is an exponential function. The sensitivity parameters for the preset gradient decay gating path, For gradient difference intensity, It is the numerical stability constant. The preset gradient influence decision threshold parameter, u is the index identifier of the neighboring node, and v is another index identifier of the neighboring node that is directly connected to node u.

6. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, In step S5, the outline geometric data of the fire area is obtained, including: Topological morphology tracing is performed on the reinforcement distribution map to obtain the skeleton nodes of the reinforcement distribution map; Based on the skeleton nodes, closed contour detection is performed on the reinforcement distribution map to obtain the closed boundary loop of the reinforcement distribution map; Based on the closed boundary loop, the loop geometric saliency is extracted from the fire area of ​​the candidate fire point pixels to obtain the main contour loop of the fire area. Adaptive step-size curvature-aware resampling is performed on the main contour loop to obtain the feature point sequence of the main contour loop; Based on the feature point sequence, piecewise spline fitting is performed on the main contour loop to obtain a smooth vector contour of the fire area; Topological verification is performed on the smooth vector profile to obtain the outline geometry data of the fire area.

7. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 1, characterized in that, In S6, obtaining the fire event of the target forest includes: The contour geometry data is divided into internal spatial grids to obtain the isoparametric sampling grid of the contour geometry data; Based on continuous response confidence and isoparametric sampling grid, spatial autocorrelation analysis is performed on contour geometric data to obtain the consistency coefficient of contour geometric data. The consistency coefficient is subjected to threshold boundary neighborhood extraction to obtain the outer neighborhood band of the contour geometric data; Based on the confidence level of the continuous response, the gradient direction of the surrounding neighborhood band is deduced to obtain the confidence gradient field of the neighborhood band. Based on the confidence gradient field, the conformity of the contour geometric data is evaluated to obtain the gradient conformity rate of the contour geometric data. Based on the gradient matching rate, the contour geometry data is dynamically optimized segment by segment to obtain the optimized fire zone boundary of the target forest. A closure check is performed on the optimized fire area boundary to obtain the fire events of the target forest.

8. The forest fire point identification method based on multispectral remote sensing and AI edge computing as described in claim 7, characterized in that, The process of obtaining the gradient matching rate of contour geometry data is as follows: The closed boundary of the contour geometry data is discretized by arc length parameterization to obtain a discrete boundary point sequence of the closed boundary; Based on the confidence gradient field, gradient vector field correlation is performed on the discrete boundary point sequence to obtain the gradient direction vector of the discrete boundary point. Local surface normal analysis is performed on the discrete boundary point sequence to obtain the unit normal vector of the discrete boundary point sequence; Based on the gradient direction vector and the unit normal vector, the directional consistency of the discrete boundary point sequence is determined, and the boundary points to be optimized in the discrete boundary point sequence are obtained. Based on the gradient direction vector, the position of the boundary point to be optimized is optimized to obtain the optimized boundary point of the discrete boundary point sequence; Based on the confidence gradient field, the closed contour is reconstructed by optimizing the boundary points, and the geometric alignment of the reconstructed boundary points is fused to obtain the comprehensive gradient matching rate of the contour geometric data.

Images