All-weather forest fire prevention intelligent monitoring system suitable for complex environment

Abstract

The application belongs to the field of fire prevention monitoring, and particularly relates to an all-weather forest fire prevention intelligent monitoring system suitable for complex environments, which constructs a dynamic monitoring path based on thermal radiation gradient analysis through a collection module, and identifies abnormal areas in real time; an image quality is improved by using a selection enhancement module to adaptively call an enhanced filtering model according to meteorological data; a dual recognition module is used to fuse thermal imaging and visible light features, and a fire situation is accurately distinguished through a multi-level filtering and bidirectional verification mechanism; finally, a fire situation positioning, situation assessment and graded early warning are realized through an early warning module; the application solves the problems of low early fire situation recognition sensitivity and high false alarm rate in complex environments, and realizes all-weather and high-precision forest fire situation monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing monitoring, and in particular relates to an all-weather intelligent forest fire monitoring system suitable for complex environments. Background Technology

[0002] Traditional forest fire monitoring mainly relies on manual observation, monospectral camera monitoring, and drone patrols, which have several technical limitations. First, manual observation is limited by high labor costs, limited field of view, and is easily affected by day-night cycles and inclement weather, making it difficult to achieve continuous 24-hour monitoring, resulting in a significant delay in fire detection. Second, monospectral cameras are only effective during the day when there is sufficient light. At night, or in rain, snow, or dense fog, the image quality deteriorates significantly, making it impossible to accurately identify key fire characteristics such as smoke and open flames. In addition, drone patrols are limited by battery life and flight weather conditions, making it difficult to achieve large-scale, all-weather real-time monitoring. This makes the problem of "difficulty in early detection" of forest fires particularly prominent, especially at night and in adverse weather conditions, when the best time to extinguish fires is easily missed, leading to the spread of fires and causing ecological and economic losses. While some thermal imaging monitoring equipment has been introduced into existing technologies, the problem of false alarms caused by interference sources such as animal heat sources has not been effectively solved. At the same time, there is a lack of systematic solutions for interference from rain and snow images and low recognition accuracy in dense fog scenes, which cannot meet the actual needs of all-weather, high-precision fire monitoring in complex forest environments. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an all-weather intelligent forest fire monitoring system suitable for complex environments. This system constructs a dynamic monitoring path based on thermal radiation gradient analysis through a data acquisition module, enabling real-time identification of abnormal areas. An enhancement module adaptively calls enhancement filtering models based on meteorological data to improve image quality. A dual-identification module fuses thermal imaging and visible light features, accurately identifying fire conditions through multi-level filtering and a two-way verification mechanism. Finally, an early warning module enables fire location, situation assessment, and tiered early warning. This invention solves the problems of low sensitivity and high false alarm rate in early fire identification under complex environments, achieving all-weather, high-precision forest fire monitoring and early warning.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An all-weather intelligent forest fire monitoring system suitable for complex environments includes: a data acquisition module and a selection and enhancement module;

[0006] The acquisition module determines an initial monitoring path based on the target monitoring area and monitors along the initial monitoring path based on preset initial discrimination indicators. It acquires images containing at least some initial discrimination indicator anomalies and their corresponding initial timestamps. Simultaneously, it adjusts the initial monitoring path according to the gradient change direction of the initial discrimination indicators in the images containing at least some initial discrimination indicator anomalies, and acquires an image sequence containing at least some initial discrimination indicator anomalies as an initial fire suspected sub-area image sequence.

[0007] The selection enhancement module is used to determine the scene type and scene monitoring difficulty coefficient of the target monitoring area based on the meteorological data collected in real time in the target monitoring area and the scene discrimination evaluation model. Based on the scene type and scene monitoring difficulty coefficient, it calls the pre-trained local enhancement filtering model library and combines it with the initial timestamp to synchronously enhance the image sequence of the initial suspected fire sub-area to obtain the enhanced image sequence of the initial suspected fire sub-area.

[0008] Specifically, the all-weather intelligent forest fire monitoring system also includes a dual identification module and an early warning module;

[0009] The dual recognition module is used to perform layered analysis and judgment of the fire situation in the suspected fire area based on the enhanced initial fire suspected sub-area image sequence and the preset dual recognition model. When it is determined to be a fire, the location of the corresponding suspected area is determined. At the same time, the severity and stage of the corresponding fire are obtained based on the temperature change status and real-time area size of the suspected area in the initial fire suspected sub-area image sequence at the corresponding location over a continuous time.

[0010] The early warning module is used to obtain fire prevention strategy information based on the severity and stage of the fire, combined with a preset early warning level and a fire prevention strategy information database, through a deep matching indexing algorithm.

[0011] Specifically, the acquisition module includes a region segmentation unit and an initial path unit;

[0012] The region segmentation unit is used to perform preliminary region segmentation based on the geographical features of the target monitoring area, the deployment and distribution of monitoring equipment, and the monitoring range characteristics, combined with a grid partitioning algorithm, to obtain an initial monitoring sub-region sequence.

[0013] The initial path unit is used to obtain the initial monitoring path based on the frequency and severity of historical fires in each initial monitoring sub-region, the geographical monitoring difficulty of the sub-region, the density and importance of attachments in the sub-region combined with the path algorithm, and the cost loss function constructed based on the density, importance and monitoring coverage of attachments in the sub-region.

[0014] Specifically, the acquisition module also includes an initial discrimination unit and a path adjustment unit;

[0015] The initial discrimination unit is used to construct an initial discrimination index based on the monitoring images collected along the initial monitoring path and the thermal radiation gradient change value. When the thermal radiation gradient value of some areas in the monitoring image at any time point does not meet the preset standard scene thermal radiation gradient distribution map, the sub-region corresponding to the current monitoring image is taken as the initial suspected fire sub-region, and the timestamp of the first abnormal monitoring image and the initial location point corresponding to the abnormal initial discrimination index are recorded. The standard scene thermal radiation gradient distribution map is constructed by combining the radiation gradient values ​​collected at different time points under different weather types in the target monitoring area under the condition that there is no fire, and a statistical analysis algorithm. The weather type includes at least one of sunny, cloudy, rainy, foggy, snowy, daytime and nighttime.

[0016] The path adjustment unit is used to adjust the direction of the initial monitoring path in real time based on the edge structure features of the monitoring image where some areas have abnormal initial discrimination indicators, as well as the thermal radiation gradient extension direction of the sub-regions corresponding to the abnormal initial discrimination indicators, and the rate and state of change of the thermal radiation gradient under the corresponding extension direction.

[0017] Specifically, the data acquisition module also includes a suspicious sub-region identification unit and a region adjustment unit;

[0018] The suspicious sub-region determination unit is used to determine the area size and corresponding edge features of the corresponding initial fire suspicious sub-region based on the edge structure features of the sub-regions corresponding to the initial discrimination index anomaly in the image sequence of the initial fire suspicious sub-regions collected corresponding to the adjusted initial monitoring path, the rate of change of thermal radiation gradient under the extension direction of the sub-regions, and the thermal radiation gradient distribution map of the standard scene.

[0019] The region adjustment unit compares the area size and corresponding edge features of the initially suspected fire sub-region determined in real time with the initial monitoring sub-region. If the initially suspected fire sub-region is larger than the initial monitoring sub-region, the unit adjusts the initial monitoring sub-regions covered by the initially suspected fire sub-region that have adjacent edges by using the proportion of the initially suspected fire sub-region in each initial monitoring sub-region, thereby obtaining an adjusted monitoring sub-region sequence. The initially suspected fire sub-region is obtained by combining the corresponding edge structure features and pixel positions of the initially suspected fire sub-region in the image with a region integration algorithm.

[0020] Specifically, the selected enhancement module includes an enhanced processing unit;

[0021] The scene discrimination unit is used to obtain the scene type label of the target demand monitoring area by combining the meteorological data synchronously collected within the target demand monitoring area for the current preset time length with the preset scene parameter library and matching algorithm.

[0022] The difficulty assessment unit is used to obtain the scene monitoring difficulty coefficient by combining the scene type label of the target required monitoring area with the correlation influence matrix of the meteorological data collected in real time under the corresponding scene on the collected data through the scene discrimination assessment model.

[0023] The scenario discrimination and evaluation model is constructed by combining the correlation influence matrix built on the meteorological data collected in real time under the corresponding scenario and the degree of influence of the meteorological data under each scenario on the target collection index, and trained with expert experience algorithm. It is used to judge the degree of influence of meteorological data on the target collection data under the current scenario in real time.

[0024] Specifically, the enhancement module also includes an enhancement processing unit;

[0025] The enhancement processing unit is used to perform real-time enhancement processing on the initial suspected fire sub-region image sequence by combining the scene monitoring difficulty coefficient with the accuracy of the hierarchical analysis and judgment corresponding to the dual recognition module and the accuracy of the fire severity, and by combining the preset local enhancement filtering model library with the matching algorithm and the preset matching threshold and the timestamp of the first abnormal monitoring image, and calling at least one local enhancement filtering model that meets the matching threshold to obtain the enhanced initial suspected fire sub-region image sequence.

[0026] The real-time enhancement process includes: adjusting the kernel parameters and enhancement magnitude of the local enhancement filter model based on the magnitude and direction of the change in thermal radiation gradient of the initially suspected fire sub-region.

[0027] Specifically, the dual identification module includes a filtering unit; the dual identification model includes an abnormal region filtering layer.

[0028] The filtering unit is used to perform temperature analysis on the thermal radiation gradient of each pixel in the initial suspected fire sub-region based on the enhanced image of the initial suspected fire sub-region combined with a preset fire temperature threshold and an abnormal region filtering layer, and to mark all pixels with temperature values ​​higher than the fire temperature threshold as abnormal pixels to obtain an abnormal pixel set.

[0029] Based on the region growing algorithm combined with the abnormal pixel set, the spatially adjacent abnormal pixels are clustered to obtain multiple connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target.

[0030] Based on multiple connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target, static contour features of the suspected heat source sub-targets are extracted; the static contour features include at least area and aspect ratio features.

[0031] The extracted static contour features are matched with the thermal imaging contour parameters of a preset fire heat source feature library. If the similarity between the static contour features of the current suspected heat source sub-target and any thermal imaging contour parameter in the fire heat source feature library is less than a first set standard, it is determined to be a non-fire heat source and excluded. If it is greater than the standard, dynamic discrimination is performed, specifically:

[0032] Based on the pixel position changes of the corresponding suspected heat source targets in the image sequence of multiple consecutive initial suspected fire sub-regions, the motion speed, motion direction and expansion of the suspected heat source targets are calculated.

[0033] The movement speed and direction of the suspected heat source sub-targets obtained from the analysis are matched with the typical fire expansion patterns stored in the fire heat source feature database. If the corresponding matching degree is less than the second set standard and the expansion of the suspected heat source sub-target remains unchanged, it is determined to be a non-fire heat source and is excluded. At the same time, suspected heat source sub-targets that exceed the second set standard are marked as high-confidence suspicious sub-targets, and the pixel coordinate set of the corresponding high-confidence suspicious sub-targets is output.

[0034] Specifically, the dual recognition module also includes a dual verification unit; the dual recognition model also includes a thermal imaging recognition layer and a visible light recognition layer;

[0035] The dual verification unit is used to extract temperature features along the thermal radiation gradient direction corresponding to the high-confidence suspicious sub-targets in continuous time through the thermal imaging recognition layer, and obtain the temperature gradient change rate and regional expansion continuity in the corresponding direction. When the temperature gradient change rate in the corresponding direction meets the fire temperature threshold and the matching degree between the regional expansion continuity and the fire expansion parameter distribution map corresponding to the fire heat source feature library is greater than the preset matching threshold, it is preliminarily determined that there is a fire in the corresponding high-confidence suspicious sub-targets, and the accuracy of the preliminary determination is output.

[0036] When it is initially determined that there is a fire in the corresponding high-confidence suspicious sub-target, the smoke grayscale feature detection of the high-confidence suspicious sub-target with fire is performed by combining the accuracy of the initial determination with the visible light recognition layer. When smoke is detected, the smoke contour and offset trajectory are extracted to obtain the smoke contour offset trajectory.

[0037] The smoke profile offset trajectory and the wind direction data, temperature gradient direction and regional expansion continuity in the meteorological data are used to determine the directional consistency. If the corresponding directional determinations are consistent, it is determined that there is a fire.

[0038] Specifically, the dual identification module also includes a fire location unit and a dynamic evaluation unit, and the dual identification model also includes a dynamic regional integration layer and an expert evaluation layer;

[0039] The fire location unit is used to obtain the location point and real-time burned area of ​​the corresponding fire by combining the position of the high-confidence suspicious sub-target with fire in the initial monitoring sub-region sequence with the corresponding temperature gradient change rate, smoke profile offset trajectory and regional expansion continuity through a dynamic regional integration layer.

[0040] The dynamic assessment unit is used to obtain the severity and stage of the fire based on the real-time burned area, the rate of change of temperature gradient in the corresponding direction, the rate of regional expansion, the smoke concentration, and the expert assessment layer.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention addresses the shortcomings of existing technologies by constructing a dynamic monitoring system based on changes in thermal radiation gradients, enabling early and accurate identification of forest fire hazards. It employs intelligent path adjustment and adaptive image enhancement to effectively overcome monitoring blind spots under different meteorological conditions. Through heat source feature analysis, animal interference elimination, and dual fire identification, it significantly improves identification accuracy. Finally, by combining multi-dimensional parameter fusion analysis, it achieves precise assessment of fire location, spread trend, and hazard level, providing efficient and reliable intelligent decision support for forest fire prevention and comprehensively enhancing the accuracy and timeliness of forest fire control. Attached Figure Description

[0043] Figure 1 This is a block diagram of the all-weather intelligent forest fire monitoring system applicable to complex environments according to the present invention;

[0044] Figure 2 This is a flowchart of the corresponding units of the all-weather intelligent forest fire monitoring system applicable to complex environments according to the present invention. Detailed Implementation

[0045] Please see Figure 1 The present invention provides an embodiment of an all-weather intelligent forest fire monitoring system suitable for complex environments, comprising: a data acquisition module, a selection enhancement module, a dual identification module, and an early warning module;

[0046] The acquisition module determines an initial monitoring path based on the target monitoring area and monitors along the initial monitoring path based on preset initial discrimination indicators. It acquires images containing at least some initial discrimination indicator anomalies and their corresponding initial timestamps. Simultaneously, it adjusts the initial monitoring path according to the gradient change direction of the initial discrimination indicators in the images containing at least some initial discrimination indicator anomalies, and acquires an image sequence containing at least some initial discrimination indicator anomalies as an initial fire suspected sub-area image sequence.

[0047] The selection enhancement module is used to determine the scene type and scene monitoring difficulty coefficient of the target monitoring area based on the meteorological data collected in real time in the target monitoring area and the scene discrimination evaluation model. Based on the scene type and scene monitoring difficulty coefficient, it calls the pre-trained local enhancement filtering model library and combines it with the initial timestamp to synchronously enhance the image sequence of the initial suspected fire sub-area to obtain the enhanced image sequence of the initial suspected fire sub-area.

[0048] The dual recognition module is used to perform layered analysis and judgment of the fire situation in the suspected fire area based on the enhanced initial fire suspected sub-area image sequence and the preset dual recognition model. When it is determined to be a fire, the location of the corresponding suspected area is determined. At the same time, the severity and stage of the corresponding fire are obtained based on the temperature change status and real-time area size of the suspected area in the initial fire suspected sub-area image sequence at the corresponding location over a continuous time.

[0049] The early warning module is used to obtain fire prevention strategy information based on the severity and stage of the fire, combined with a preset early warning level and a fire prevention strategy information database, through a deep matching indexing algorithm.

[0050] In the current process of forest fire prevention data collection, there are several technical problems that are difficult to solve simultaneously. These include insufficient targeted path planning due to differences in the frequency of fire occurrence, conservation value, and geographical monitoring difficulty in different monitoring sub-regions; misjudgment of fire anomalies caused by natural heat radiation fluctuations under different weather and time conditions; easy neglect or misjudgment of weak heat radiation anomalies in the early stages of a fire; inability of fixed monitoring paths to keep up with the direction of heat radiation gradient extension when a fire is dynamically spreading; and incomplete monitoring due to mismatch between the initial monitoring sub-region division and the actual range of suspected fire-prone areas. These issues, in turn, affect the reliability of subsequent fire identification and early warning.

[0051] Please see Figure 2 This is a flowchart of a unit in an all-weather intelligent forest fire monitoring system suitable for complex environments. It should be further noted that the data acquisition module in this embodiment includes a region segmentation unit, an initial path unit, an initial discrimination unit, a path adjustment unit, a suspicious sub-region determination unit, and a region adjustment unit.

[0052] The region segmentation unit is used to perform preliminary region segmentation based on the geographical features of the target monitoring area, the deployment and distribution of monitoring equipment, and the monitoring range characteristics, combined with a grid partitioning algorithm, to obtain an initial monitoring sub-region sequence.

[0053] It should be further explained that the geographical features mentioned in this embodiment include the altitude, topographic relief, slope aspect, and slope of the monitoring area, used to identify areas prone to fire and with high risk of fire spread; the deployment distribution of the monitoring equipment includes the geographical coordinates of each monitoring node and the model of the monitoring equipment to which it belongs, used to determine the actual monitoring capability and coverage of each device; the monitoring range features include the maximum effective monitoring radius, horizontal and vertical viewing angle of each monitoring device, used to accurately calculate the actual monitorable geographical range within its field of view; by combining the above features, the grid partitioning algorithm can generate an optimized regional partitioning scheme that matches the geographical risk distribution and equipment monitoring capabilities, ensuring that each initial monitoring sub-region is both effectively covered and has similar fire risk monitoring value.

[0054] The initial path unit is used to obtain the initial monitoring path based on the frequency and severity of historical fires in each initial monitoring sub-region, the geographical monitoring difficulty of the sub-region, the density and importance of attachments in the sub-region combined with the path algorithm, and the cost loss function constructed based on the density, importance and monitoring coverage of attachments in the sub-region.

[0055] It should be further explained that the process of obtaining the initial monitoring path in this embodiment includes:

[0056] Based on the historical fire database of each initial monitoring sub-region, the frequency and severity of fires are extracted; the terrain relief and accessibility parameters of the sub-region are analyzed through digital elevation model, and the geographical monitoring difficulty coefficient is calculated.

[0057] Remote sensing image recognition technology is used to obtain the distribution data of attachment types in sub-regions. Combined with a preset importance level classification library and expert evaluation algorithm, the density of attachments and the degree of protection are obtained. Finally, the real-time monitoring coverage of each sub-region is calculated based on the monitoring angle and distance parameters of the equipment. It should be further noted that the importance level classification library in this embodiment is constructed by those skilled in the art based on the attachment types in the corresponding regions and the importance and corresponding damage value of the manually marked attachments to be protected.

[0058] Based on the multi-dimensional feature data of each sub-region, a cost value calculation matrix is ​​constructed, in which the frequency and severity of historical fires are combined into a fire risk factor, the geographical monitoring difficulty is used as a basic cost factor, the density and importance of attachments are combined into a protection value factor, and the monitoring coverage rate is used as an efficiency correction factor.

[0059] Based on fire risk factor, basic cost factor, protection value factor and efficiency correction factor, the comprehensive cost value of each initial monitoring sub-region is obtained through a linear weighted algorithm.

[0060] Based on the cost value calculation matrix of the generated initial monitoring sub-regions, the sub-regions are sorted according to their comprehensive cost value. Key sub-regions with comprehensive cost values ​​exceeding a preset cost threshold are given the highest priority. Sub-regions with comprehensive cost values ​​in the middle range are sorted using a weighted round-robin strategy. Regular sub-regions with low comprehensive cost values ​​are arranged using a basic monitoring sequence, generating a sub-region access order table classified by monitoring urgency. It should be further noted that the monitoring urgency is constructed from the comprehensive cost value of the corresponding initial monitoring sub-region.

[0061] Based on the obtained sub-region access order list, an improved genetic algorithm is used to perform path optimization calculation. The improved genetic algorithm constructs a fitness function with the optimization objectives of minimizing the total monitoring time and maximizing the risk coverage rate. The optimal path is solved iteratively through population initialization, selection operation, crossover operation and mutation operation, with constraints such as device movement path length, gimbal turning angle limit and device stability requirements.

[0062] It should be further noted that, in this embodiment, the total monitoring time represents the estimated total time required for the candidate path to complete the monitoring tasks for all sub-regions in the sub-region access sequence table. It should also be noted that, in this embodiment, the calculation of the total monitoring time incorporates the movement time consumed by the device's movement path length, the turning time consumed by the change in the pan-tilt-zoom angle, and the preset dwell time at each monitoring point.

[0063] It should be further explained that, in this embodiment, maximizing risk coverage means the sum of the risk weight coefficients of all sub-regions successfully visited by the candidate path. The risk weight coefficient of each sub-region is pre-calculated based on factors such as its historical fire frequency, severity, and the importance of attached structures.

[0064] It should be further explained that one implementation process of using an improved genetic algorithm for path optimization calculation in this embodiment includes:

[0065] Step 1: Based on the monitoring point sequence determined by the sub-region access sequence list, an initial population is generated through a heuristic initialization strategy. Each individual represents a monitoring path scheme and uses real number encoding to represent the dwell time, rotation speed and scanning sequence parameters of the pan-tilt unit at each monitoring point.

[0066] Step 2: Based on the path scheme of each individual in the population, evaluate its performance through a multi-objective fitness function. This function considers two optimization objectives at the same time: minimizing the total monitoring time and maximizing the risk coverage. It also handles constraints such as the length of the device's movement path, the gimbal turning angle limit, and the device stability requirements through a constraint handling mechanism.

[0067] Step 3: Based on the fitness assessment results, a selection strategy combining improved roulette wheel selection and elite retention is adopted to prioritize the retention of individuals with high fitness values ​​while ensuring population diversity, thereby obtaining a set of parent individuals for reproduction.

[0068] Step 4: Based on the selected set of parent individuals, gene exchange is performed using an improved sequential crossover operator based on path sequences. By maintaining effective path segments while exchanging the access order of monitoring points, offspring individuals with new characteristics are generated.

[0069] Step 5: Based on the offspring individuals generated by crossover, a combination strategy of multiple mutation operators is adopted, including mutation operations such as path segment inversion, monitoring point replacement, and parameter adaptive adjustment. By introducing random perturbation, the population diversity is enhanced and premature convergence is avoided.

[0070] Step 6: Based on the new generation of individuals generated through selection, crossover, and mutation operations, update the population through intergenerational competition strategy, retaining excellent individuals and eliminating individuals with poor fitness to obtain an optimized new population.

[0071] Step 7: Based on the optimal fitness value and number of generations of the current population, determine whether the termination condition is met through a convergence detection algorithm. The iteration terminates when the maximum number of generations is reached or the optimal solution is stable; otherwise, return to step 2 to continue the evolution process.

[0072] Step 8: Based on the individuals in the final population, the path scheme with the highest fitness is selected as the optimal solution through the elite individual extraction algorithm, and it is converted into a control command sequence containing the coordinates of the monitoring point, the dwell time, the gimbal rotation speed and the optical zoom parameters. Finally, a configuration file of the standardized initial monitoring path is generated and output to the equipment control system.

[0073] The initial discrimination unit is used to construct an initial discrimination index based on the monitoring images collected along the initial monitoring path and the thermal radiation gradient change value. When the thermal radiation gradient value of some areas in the monitoring image at any time point does not meet the preset standard scenario thermal radiation gradient distribution map, the sub-area corresponding to the current monitoring image is taken as the initial suspected fire sub-area, and the timestamp of the first abnormal monitoring image and the initial location point corresponding to the abnormal initial discrimination index are recorded. The standard scenario thermal radiation gradient distribution map is constructed by combining the radiation gradient values ​​collected at different time points under different weather types in the target required monitoring area with a statistical analysis algorithm under the condition that there is no fire. It should be further noted that the weather type in this embodiment includes at least one of sunny, cloudy, rainy, foggy, snowy, daytime and nighttime. It should be further noted that the standard scenario thermal radiation gradient distribution map in this embodiment is used to provide a thermal environment benchmark reference template for the monitoring area under the condition of no fire. It establishes the normal fluctuation range of thermal radiation at each time and space location under different weather conditions through historical data, and provides a dynamic threshold basis for the initial anomaly discrimination.

[0074] It should be further explained that the process of obtaining the initial suspected fire sub-region in this embodiment includes:

[0075] Based on the historical fire-free periods in the target monitoring area, raw thermal radiation data under different weather types are collected at preset time intervals. The data are then processed using the Kriging spatial interpolation algorithm to obtain the standard thermal radiation gradient distribution matrix for the corresponding weather type and time point. Combined with the graph database, a standard scene thermal radiation gradient distribution map database is constructed.

[0076] Based on the real-time monitoring images and current time on the initial monitoring path, the real-time thermal radiation raw data is processed by the Sobel operator gradient algorithm to obtain the real-time thermal radiation gradient distribution matrix. At the same time, the current weather type is extracted and the standard thermal radiation gradient distribution matrix matched in the database is called.

[0077] Based on the statistical standard deviation of historical fire-free samples, an allowable deviation threshold for thermal radiation gradient was set, and initial discrimination indicators for different weather types were constructed.

[0078] Based on the initial discrimination index under different weather types and combined with the current weather type, the initial discrimination index is verified at each coordinate point in the real-time thermal radiation gradient distribution matrix. The difference between the real-time initial discrimination index value and the standard initial discrimination index value is determined to exceed the deviation threshold, thereby obtaining the abnormal thermal radiation pixels.

[0079] Based on the abnormal pixels of thermal radiation, adjacent abnormal pixels are aggregated by the 8-neighborhood connectivity analysis algorithm to obtain a continuous sub-region, namely the initial suspected fire sub-region.

[0080] Based on the initial suspected fire sub-region, the acquisition time of the monitoring image corresponding to the first appearance of the initial suspected fire sub-region is extracted as the first abnormal monitoring image timestamp, and its minimum bounding rectangle coordinate range is determined as the initial location point, so as to obtain the initial suspected fire sub-region with associated timestamp and initial location point.

[0081] The path adjustment unit is used to adjust the direction of the initial monitoring path in real time based on the edge structure features of the monitoring image where some regions have abnormal initial discrimination indicators, the extension direction of the thermal radiation gradient of the sub-regions corresponding to the initial discrimination indicators, and the rate and state of change of the thermal radiation gradient under the corresponding extension direction. It should be further noted that the state of change here refers to whether the rate of change of the thermal radiation gradient under the corresponding extension direction is increasing or decreasing, determined by calculating the sign of the first derivative of the rate of change of the thermal radiation gradient: a positive value indicates an increase, and a negative value indicates a decrease. Based on whether the state of change is increasing or decreasing, the direction and step size of the monitoring path are dynamically adjusted. When the rate of change increases, the path direction deflects towards the gradient extension direction and the scanning step size is reduced; when the rate of change decreases, the path direction deflection weight is reduced and the step size is increased. Combining the continuous temporal characteristics of the state of change, the abnormal boundary is predicted and an arc-shaped scanning path is generated to accurately delineate the edge structure.

[0082] It should be further explained that the specific process of adjusting the initial monitoring path direction in real time in this embodiment includes:

[0083] Based on real-time acquired images of suspected fire sub-regions and corresponding thermal radiation gradient data, edge structure features of abnormal regions are extracted using edge detection algorithms to obtain contour information and morphological features of abnormal regions. At the same time, thermal radiation gradient data is processed using gradient direction analysis algorithms to obtain the extension direction of thermal radiation gradient in abnormal regions.

[0084] Based on the obtained thermal radiation gradient extension direction, the unit vector of the main extension direction is calculated by the direction vector statistical algorithm to obtain the direction vector characterizing the abnormal diffusion direction. At the same time, the rate of change of thermal radiation gradient in the corresponding direction is calculated by time-series gradient analysis to obtain the gradient change rate parameter.

[0085] Based on the obtained direction vector and gradient change rate parameters, a gimbal adjustment scheme is generated through a path planning algorithm. When the gradient change rate exceeds the normal fluctuation range, the gimbal is controlled to perform priority scanning along the direction vector. When there are multiple abnormal areas, the scanning priority is determined according to the magnitude of the gradient change rate of each area. It should be further noted that in this embodiment, the normal fluctuation range is the historical statistical normal fluctuation range of the thermal radiation gradient change rate of the target monitoring area under the corresponding weather type and time point.

[0086] Based on the generated gimbal adjustment scheme, the horizontal rotational angular velocity, pitch angle change, and focal length adjustment parameters of the gimbal are calculated through motion control algorithms. The gimbal movement speed is dynamically adjusted according to the gradient change rate, and the scanning speed is increased when the gradient change rate is large.

[0087] Based on the calculated gimbal control parameters, the monitoring path direction is updated in real time through the device control interface, enabling the gimbal to scan along the thermal radiation gradient extension direction; at the same time, the path adjustment time and adjustment parameters are recorded, and an updated monitoring path configuration file is generated.

[0088] The suspected sub-region determination unit is used to determine the area size and corresponding edge features of the corresponding initial suspected sub-region based on the edge structure features of the sub-regions corresponding to the initial discrimination index anomalies in the image sequence of the initial suspected sub-regions collected corresponding to the adjusted initial monitoring path, the rate of change of thermal radiation gradient under the extension direction of the sub-regions, and the thermal radiation gradient distribution map of the standard scene.

[0089] It should be further explained that the process of obtaining the area size of the initial suspected fire sub-region in this embodiment includes:

[0090] Based on the image sequence of the initial suspected fire sub-region acquired by the adjusted initial monitoring path, the spatial alignment of each frame image is performed by a time-series image registration algorithm to obtain the image sequence after position calibration.

[0091] Based on the image sequence after position calibration, the edge structure features of the abnormal region are extracted by a multi-frame feature fusion algorithm, the edge feature map of each frame is obtained by an edge detection operator, and an enhanced edge feature map is generated by a time-series integration method.

[0092] Based on the thermal radiation gradient data in the enhanced edge feature map, the extension direction of the abnormal region is calculated by the gradient direction statistical analysis algorithm, and the rate of change of thermal radiation gradient in the extension direction is obtained to obtain the gradient change rate distribution map.

[0093] Based on the obtained gradient change rate distribution map, the baseline value of the corresponding location is obtained by querying the standard scene thermal radiation gradient distribution map database. The significance of the abnormal region is verified by the difference calculation algorithm, and the verified abnormal region mask is obtained.

[0094] Based on the obtained enhanced edge feature map and the obtained abnormal region mask, the binary image features of the abnormal region mask are corrected by morphological optimization algorithm, the edge connection algorithm is used to fill the edge discontinuities in the mask, and the noise filtering algorithm is used to eliminate isolated edge points in the mask, so as to obtain the optimized continuous mask edge features.

[0095] Based on the optimized continuous mask edge features, the boundary of the closed abnormal region is obtained through the edge contour tracking algorithm. The number of pixels occupied by the abnormal region is calculated by the pixel point statistics method. At the same time, the number of pixels is converted into the actual physical area value by combining the region integration algorithm to obtain the accurate area size of the initial suspected fire sub-region.

[0096] The region adjustment unit compares the area size and corresponding edge features of the initially suspected fire sub-region determined in real time with the initial monitoring sub-region. If the initially suspected fire sub-region is larger than the initial monitoring sub-region, the unit uses the proportion of the initially suspected fire sub-region in each initial monitoring sub-region to adjust the initial monitoring sub-regions covered by the initially suspected fire sub-region that have adjacent edges, thereby obtaining the adjusted monitoring sub-region sequence.

[0097] It should be further explained that the adjustment process of the monitoring sub-region in this embodiment includes:

[0098] Based on the spatial coordinate range of the initial suspected fire sub-region and the initial monitoring sub-region determined in real time, the overlap area ratio of each initial monitoring sub-region and the initial suspected fire sub-region is calculated by the spatial geometric relationship analysis algorithm to obtain the set of overlap coefficients of each sub-region;

[0099] Based on the spatial distribution characteristics of the initial monitoring sub-regions, a sub-region adjacency relationship model is established using the Delaunay triangulation construction algorithm. All initial monitoring sub-regions that are spatially adjacent to the initial suspected fire sub-regions are identified, and a set of adjacent sub-regions is obtained.

[0100] Based on the obtained set of overlap coefficients and the obtained set of adjacent sub-regions, an adjusted weight value for each sub-region is calculated using a weighted fusion algorithm. The overlap coefficient weight accounts for a% (usually set to 60%-80% to highlight the actual coverage impact of suspected areas), and the adjacency relationship weight accounts for b% (usually set to 20%-40% to consider spatial correlation), thus obtaining the dynamic weight distribution of the sub-regions. The weight calculation formula is as follows:

[0101] W i =a%×OverlapRatio i +b%×NeighborCount i ;

[0102] OverlapRatio i NeighborCount is the overlap coefficient of sub-region i. i W is the normalized value of the number of adjacent subregions between subregion i and the initial suspected fire subregion; i This represents the adjusted weight value for sub-region i;

[0103] Based on the dynamic weight distribution of sub-regions, the boundaries of the initial monitoring sub-regions are adjusted using the Thiessen polygon reconstruction algorithm. The focus is on optimizing the boundaries of sub-regions with weight values ​​higher than the preset weight threshold, so that the adjusted sub-region boundaries better match the actual contours of the initial suspected fire sub-regions, and thus obtain the optimized sub-region boundary sequence.

[0104] Based on the optimized sub-region boundary sequence, the monitoring priority of each sub-region is recalculated through the path coverage optimization algorithm to generate a new monitoring sub-region sequence, ensuring that the adjusted sequence can fully cover the initial suspected fire sub-regions while maintaining monitoring efficiency.

[0105] Forest fire prevention in this embodiment is a long-term monitoring process. If a uniform enhancement method is used for all the large number of images collected, there are two shortcomings: First, under complex and changeable weather conditions, such as foggy days, rainy days, and low-light nights, fixed enhancement algorithms are difficult to overcome image quality degradation problems and cannot effectively suppress scene-specific interference, such as fog obstruction, rain and snow noise, and motion blur, resulting in poor enhancement effects and difficulty in extracting key fire features. Second, in low-complexity scenes with good weather conditions and high image quality, using high-intensity enhancement processing will waste computing resources and reduce the overall efficiency of the system. Existing processing flows usually only focus on the content processing of the image itself and fail to combine the actual weather scene and monitoring difficulty information at the time of image acquisition, thus lacking adaptive adjustment capabilities.

[0106] It should be further noted that the selection enhancement module in this embodiment includes a scene discrimination unit, a difficulty assessment unit, and an enhancement processing unit;

[0107] The scene discrimination unit is used to obtain the scene type label of the target demand monitoring area by combining the meteorological data synchronously collected within the target demand monitoring area for the current preset time length with the preset scene parameter library and matching algorithm.

[0108] It should be further explained that the meteorological data in this embodiment includes at least: real-time wind speed in the target monitoring area, which affects the smoke diffusion status and the degree of dynamic interference in the image; relative humidity, which affects the vegetation thermal radiation characteristics and water vapor interference in the image; light intensity, which affects the brightness and contrast of the visible light image; precipitation status, which can distinguish between no precipitation, light rain, moderate rain, etc., and is associated with raindrop obstruction interference in the image; atmospheric visibility, which reflects the impact of fog, haze, etc. on image clarity; cloud cover, which affects the amount of solar radiation received by the ground surface and thus affects the regional thermal radiation benchmark; and surface temperature, which can help correct the thermal radiation gradient discrimination benchmark. Each meteorological data is accompanied by a timestamp synchronized with the monitoring image acquisition time to ensure data timeliness and relevance.

[0109] It should be further explained that the process of obtaining the scene type label of the target demand monitoring area in this embodiment includes:

[0110] Based on the typical scenario classification of forest fire prevention monitoring, it covers basic scenarios such as sunny daytime, sunny nighttime, cloudy daytime, cloudy nighttime, rainy daytime, foggy daytime, and snowy daytime, as well as composite scenarios such as foggy daytime, low light daytime, rainy daytime, and strong wind. It collects historical meteorological data of the target monitoring area for the past 3 years and monitoring image feature data of the corresponding time period. The historical meteorological data includes parameters such as wind speed, humidity, and light intensity under each scenario, and the monitoring image feature data includes the average image brightness, contrast, and noise type.

[0111] Statistical analysis methods were used to determine the normal fluctuation range of each meteorological parameter under each scenario, and at the same time, the correlation mapping relationship between each scenario and the corresponding image features was established, such as the low contrast of foggy images and the local highlights of rainy images.

[0112] The association items of scene type, meteorological parameter range and image feature are stored in a structured format to build a preset scene parameter library. The scene parameter library also reserves a parameter update interface to support dynamic optimization of parameter range based on subsequent monitoring data.

[0113] Based on distributed weather stations deployed in the target demand monitoring area, at least one station is set up within a 5-kilometer radius, with denser deployment in key areas. Combined with the micro meteorological sensors carried by the monitoring equipment, the micro meteorological sensors can simultaneously collect micro meteorological data around the equipment and collect the above meteorological data in real time.

[0114] Outliers are removed through data cleaning algorithms. Outliers refer to data that exceeds the reasonable physical range due to sensor malfunction. Data normalization is used to convert meteorological parameters of different dimensions into standardized parameters of the same dimension. Based on the acquisition timestamp of the monitoring images, meteorological data with a deviation from the image acquisition time not exceeding a preset duration are selected to achieve time alignment between meteorological data and monitoring images.

[0115] Based on the preprocessed real-time meteorological data, a preset scenario parameter library is invoked. An interval matching algorithm is used to compare each real-time meteorological parameter with the meteorological parameter range corresponding to each scenario in the scenario parameter library. The number of meteorological parameter items falling within the scenario parameter range is counted, and a preliminary set of candidate scenario types with a percentage exceeding a preset ratio is selected. If the candidate scenario type set is empty, a supplementary data collection mechanism is triggered, extending the meteorological data collection time to a preset time, and re-executing preprocessing and initial matching.

[0116] Based on real-time acquired monitoring images, preliminary image features are extracted through image preprocessing algorithms. These preliminary image features include average brightness, contrast, edge sharpness, and texture features such as whether raindrops or fog droplets are present. The image feature association items corresponding to candidate scene types in the preset scene parameter library are called, and the cosine similarity algorithm is used to compare the similarity between the real-time image features and the associated image features in the scene parameter library. Candidate scene types with similarity exceeding the preset similarity threshold are selected.

[0117] Based on the advanced matching results, if there is only one candidate scene type, it is directly used as the scene type label for the target demand monitoring area; if there are multiple candidate scene types, the comprehensive matching score of each candidate scene type is calculated. The candidate scene type with the highest comprehensive matching score is selected as the scene type label according to the meteorological parameter matching ratio and the image feature similarity ratio, and the confidence of the label is calculated simultaneously. The confidence is determined based on the ratio of the difference between the highest and second highest comprehensive matching scores.

[0118] If the confidence level is lower than the preset confidence threshold, the synchronous meteorological data from nearby meteorological stations will be called for supplementary verification. The scene type of the current area will be deduced by combining the scene types of the nearby areas, and finally the scene type label and the corresponding confidence level evaluation result will be output.

[0119] The difficulty assessment unit is used to obtain the scene monitoring difficulty coefficient by combining the scene type label of the target required monitoring area with the correlation influence matrix of the meteorological data collected in real time under the corresponding scene on the collected data through the scene discrimination assessment model.

[0120] The scenario discrimination and evaluation model is constructed by combining the correlation influence matrix built on the meteorological data collected in real time under the corresponding scenario and the degree of influence of the meteorological data under each scenario on the target collection index, and trained with expert experience algorithm. It is used to judge the degree of influence of meteorological data on the target collection data under the current scenario in real time.

[0121] It should be further explained that the process of obtaining the scene monitoring difficulty coefficient in this embodiment includes:

[0122] Based on the characteristics of forest fire monitoring scenarios in the target monitoring areas, this study analyzes key meteorological parameters affecting the collected indicators from historical monitoring data to determine that the row dimension of the correlation influence matrix consists of real-time meteorological parameters, including at least wind speed, relative humidity, light intensity, precipitation status, atmospheric visibility, cloud cover, and surface temperature. By analyzing the core quality indicators of the data collected by monitoring equipment, the column dimension of the correlation influence matrix consists of target collected indicator items, including at least thermal radiation data accuracy, visible light image clarity, infrared image signal-to-noise ratio, smoke feature recognition accuracy, and temperature gradient calculation accuracy. Based on the scenario type classification results, a corresponding correlation influence matrix is ​​constructed for each scenario to ensure that the correlation influence matrix adapts to the differences in the correlation between meteorological parameters and collected indicators under different scenarios, thus obtaining the basic framework of the scenario-specific correlation influence matrix.

[0123] It should be further explained that the construction process of the correlation influence matrix in this embodiment includes:

[0124] Based on the historical monitoring dataset of the target demand monitoring area for the past 5 years, this dataset contains the quality detection results of different meteorological parameter values ​​and corresponding collection indicators under each scenario. The corresponding data pairs of meteorological parameter values ​​and collection indicator quality are extracted through data filtering algorithms, and invalid data pairs caused by equipment failure or extreme abnormal weather are excluded.

[0125] The variation range of the quality of the collected indicators was calculated by statistical analysis methods under different value ranges of the same meteorological parameter, and the degree of impact was initially classified into four levels: no impact, slight impact, moderate impact, and severe impact.

[0126] Meanwhile, based on the changes in meteorological data and indicator quality collected under different value ranges of the same meteorological parameter, the initial correlation influence matrix is ​​obtained through the minimum membership degree algorithm;

[0127] Based on real-time monitoring data of the target demand monitoring area, by periodically extracting data pairs corresponding to newly generated meteorological parameters and collected indicators, and using statistical analysis methods consistent with the basic quantification process, the quantitative value of the impact degree corresponding to the new data is calculated.

[0128] The new quantified value is compared with the existing quantified value in the correlation influence matrix through the difference comparison algorithm. If the difference exceeds the preset reasonable range, the matrix calibration mechanism is triggered to adjust the influence degree quantified value at the corresponding position in the matrix and obtain the adjusted initial correlation influence matrix.

[0129] The enhancement processing unit is used to perform real-time enhancement processing on the initial suspected fire sub-region image sequence by combining the scene monitoring difficulty coefficient with the accuracy of the hierarchical analysis and judgment corresponding to the dual recognition module and the accuracy of the fire severity, and by combining the preset local enhancement filtering model library with the matching algorithm and the preset matching threshold and the timestamp of the first abnormal monitoring image, and calling at least one local enhancement filtering model that meets the matching threshold to obtain the enhanced initial suspected fire sub-region image sequence.

[0130] The real-time enhancement processing includes: adjusting the kernel parameters and enhancement amplitude of the local enhancement filter model based on the magnitude and direction of the change in thermal radiation gradient of the initial suspected fire sub-region, so as to reduce the demand for computing resources and improve the monitoring efficiency of the initial suspected fire sub-region.

[0131] It should be further explained that the process of calling at least one local enhancement filtering model that meets the matching threshold to perform real-time enhancement processing on the initial suspected fire sub-region image sequence in this embodiment includes:

[0132] Based on typical forest fire monitoring scenario types and difficulty levels, this study collects enhancement processing cases from historical fire monitoring under different scenarios and difficulties. It extracts suitable filter model types from these cases, including bilateral filtering, guided filtering, and nonlocal mean filtering, along with their corresponding parameter configuration ranges. These models are categorized and stored in a four-level structure: scenario type, difficulty level, model type, and parameter range, thus constructing a pre-defined local enhancement filter model library. Each model in this library is associated with image features it excels at processing. For example, the bilateral filtering model is suitable for image edge preservation by combining spatial proximity and pixel value similarity to smooth the image while... It can effectively preserve the edge structure of the thermal radiation area; the guided filter model is adapted to the need for detail enhancement. It utilizes the characteristics of the guided image to enhance subtle features such as smoke outlines and thermal radiation gradients, while suppressing background noise. In the local enhancement filter model library, each filter model is associated with the type of image features it is good at processing, and comes with an enhancement effect score of the model in historical application cases. This score is calculated based on multiple objective indicators, mainly including thermal radiation detail preservation, noise suppression effect, edge sharpness improvement rate, and stability performance under different meteorological scenarios, thus providing a quantitative basis for subsequent adaptive selection of the optimal model for different fire characteristics and scenario difficulty.

[0133] Based on the scene monitoring difficulty coefficient, the difficulty level of the current scene is determined. If the coefficient falls into the low difficulty range, it corresponds to a low difficulty level.

[0134] Based on the hierarchical analysis and discrimination results of the dual recognition module, the recognition accuracy value of the initial suspected fire sub-region is extracted, such as the target contour recognition accuracy.

[0135] Based on the characteristics of thermal radiation gradient changes in suspected fire sub-regions, the severity level of the fire is determined. Specifically, this is achieved by calculating the average rate of change of the thermal radiation gradient within the sub-region per unit time and comparing it to a preset threshold range for quantitative assessment: if the rate of change is below the lower threshold, it is classified as a minor fire, indicating low heat release energy and possibly in a smoldering or initial stage; if the rate of change is between the upper and lower thresholds, it is classified as a moderate fire, indicating that heat release is stabilizing and the fire is in a developing stage; if the rate of change exceeds the upper threshold, it is classified as a severe fire, indicating a sharp increase in heat release and the fire may have entered a stage of intense combustion. This classification provides important information on the fire situation for subsequent adaptive selection of image enhancement filtering models.

[0136] The difficulty level, recognition accuracy value, and severity level are converted into a model matching condition vector with a unified dimension through a feature mapping algorithm. Each dimension corresponds to a quantitative indicator of difficulty adaptability, accuracy improvement capability, and severity adaptability in the model library.

[0137] Based on the generated model matching condition vector, the local enhanced filtering model library is called, and a multi-dimensional similarity calculation algorithm is used to calculate the comprehensive similarity between each model in the local enhanced filtering model library and the model matching condition vector.

[0138] The overall similarity is compared with a preset matching threshold to select a set of candidate models whose overall similarity exceeds the matching threshold. If the set of candidate models is empty, the matching threshold is lowered, and the selection is repeated by decreasing the threshold by a preset step size until at least one candidate model is obtained.

[0139] If the number of candidate models exceeds the preset limit, they will be sorted from high to low based on their overall similarity, and the preset number of models at the top will be selected as the final candidate models.

[0140] Based on the timestamp of the first anomaly monitoring image, extract the seasonal and time period information corresponding to the timestamp, such as winter night and summer day;

[0141] The historical application records of candidate models are retrieved, and the average score of the enhancement effect of the model under the same season and time period is calculated. If the average score exceeds the preset timeliness score threshold, the model is determined to be suitable for the current time scenario and the model is retained.

[0142] If the mean score is lower than the timeliness score threshold, the model is removed and a second-highest similarity model is selected from the local enhanced filtering model library to ensure that the final model has time scene adaptability.

[0143] Based on the number of candidate models and their respective strengths in enhancement, such as Model A being good at noise suppression and Model B being good at detail enhancement, if there is only one candidate model, then the model is directly called to perform frame-by-frame enhancement processing on the initial image sequence of the suspected fire sub-region. During processing, the parameters of the corresponding candidate model are adjusted in real time based on the corresponding real-time image resolution, the rate and direction of change of thermal radiation gradient, and the target processing results required.

[0144] If multiple candidate models exist, the model combination method is determined through a collaborative strategy. Specifically, the complementarity score of each model in handling similar scenarios in historical cases is calculated, such as the complementarity between noise suppression and detail enhancement. The two models with the highest complementarity scores are selected to form a collaborative enhancement combination. The first model prioritizes processing image noise interference, and the second model enhances thermal radiation detail features on this basis. During collaborative enhancement, a dynamic weight allocation algorithm is used to adjust the weights based on the noise ratio and detail blur of the current image to control the intensity of the two models and avoid over-enhancement or feature conflict.

[0145] Based on the enhanced image sequence, three core indicators—thermal radiation gradient detail retention rate, noise suppression rate, and fire feature recognition rate—are extracted using an effectiveness evaluation algorithm. These indicators are compared with the accuracy thresholds required by the dual recognition module. For example, the detail retention rate must meet the feature extraction requirements of the hierarchical analysis. If all indicators meet the standards, the enhancement process is complete. If any indicator fails to meet the standards, the reasons are analyzed. For instance, insufficient noise suppression rate indicates that the noise processing parameters of the current model are inappropriate. The model parameter adjustment rules are invoked; for example, to address insufficient noise suppression, the smoothing parameter of the filtering model is increased, and the enhancement process is repeated. If the standards are still not met after adjusting for a preset number of times, the next model in the candidate model set is used, and the enhancement and verification process is repeated until an initial image sequence of the suspected fire sub-region that meets the accuracy requirements is obtained.

[0146] It should be further explained that the process of adjusting the kernel parameters and enhancement magnitude of the local enhancement filter model in this embodiment includes:

[0147] Based on the initial image sequence of suspected fire sub-regions, all pixels in the sub-regions are scanned frame by frame. The gradient detection algorithm is used to extract the magnitude of the thermal radiation gradient change value of each pixel. At the same time, the directional information corresponding to each gradient change value is recorded by the angle statistics algorithm. By associating with the thermal radiation gradient benchmark data of the sub-region during historical fire-free periods, the distribution of effective thermal radiation gradient change values ​​and the corresponding directional set in the sub-region are obtained.

[0148] The direction set is classified according to spatial distribution characteristics by clustering algorithm to determine the main direction of gradient change, that is, the main direction and secondary direction with the highest frequency of occurrence;

[0149] Based on the magnitude of the extracted effective thermal radiation gradient change value, it is divided into three levels: slight change, moderate change, and significant change according to the classification rules. The classification is based on the critical difference data between the gradient change in the early stage of historical fires and natural fluctuations.

[0150] Based on the preset gradient change level-kernel parameter range mapping relationship in the local enhancement filter model library, the corresponding kernel parameter range is called. Slight changes correspond to a smaller kernel parameter range, which is suitable for low detail enhancement needs. Moderate changes correspond to a medium kernel parameter range, which is suitable for regular detail enhancement needs. Significant changes correspond to a larger kernel parameter range, which is suitable for high detail enhancement needs. The initial values ​​of the kernel parameters are obtained.

[0151] By statistically analyzing the percentage of pixels at different gradient change levels within a sub-region, if the percentage of pixels at a certain level exceeds half of the total number of pixels in the sub-region, the median value of the kernel parameter range corresponding to that level is used as the initial value of the kernel parameter, ensuring that the initial value matches the mainstream gradient characteristics of the sub-region.

[0152] This embodiment achieves adaptive enhancement and precise allocation of computing resources in forest fire monitoring image processing through the coordinated operation of the enhanced filtering module. This module first accurately identifies the scene type based on real-time meteorological data, then assesses the monitoring difficulty coefficient under the current environment, and finally dynamically calls the local enhanced filtering model that best matches the scene difficulty, fire identification accuracy, and severity. This process effectively overcomes the inherent defects of traditional single enhancement algorithms, such as poor performance in processing complex meteorological scene images and resource waste in simple scenes. By adjusting model parameters and enhancement amplitude in real time according to the characteristics of thermal radiation gradient changes, the system can specifically suppress environmental noise interference such as fog, rain, and snow, significantly improving the thermal radiation detail retention rate and the identification accuracy of fire features. This not only provides high-quality image data for the subsequent dual identification module, fundamentally improving the accuracy and reliability of early fire identification, but also greatly enhances the system's adaptability to changing natural environments, significantly reducing the risk of missed and false alarms. This provides a more accurate and reliable data foundation for forest fire prevention decision-making, and the overall monitoring efficiency and effectiveness are simultaneously optimized.

[0153] The dual identification module in this embodiment includes a screening unit, a dual verification unit, a fire location unit, and a dynamic evaluation unit; the dual identification model includes an abnormal area screening layer, a thermal imaging identification layer, a visible light identification layer, a dynamic area integration layer, and an expert evaluation layer.

[0154] The filtering unit is used to perform temperature analysis on the thermal radiation gradient of each pixel in the initial suspected fire sub-region based on the enhanced image of the initial suspected fire sub-region combined with a preset fire temperature threshold and an abnormal region filtering layer, and to mark all pixels with temperature values ​​higher than the fire temperature threshold as abnormal pixels to obtain an abnormal pixel set.

[0155] It should be further explained that the process of obtaining the abnormal pixel set in this embodiment includes:

[0156] Based on the enhanced initial images of suspected fire sub-regions, the thermal radiation gradient value of each pixel is converted into a temperature value using a radiometric calibration algorithm. Sensor calibration parameters are then used to eliminate the influence of equipment differences, resulting in a standardized temperature data matrix.

[0157] Based on current environmental meteorological data and historical fire feature database, a basic fire temperature threshold adapted to the current environment is calculated through a temperature compensation model. Differential thresholds are set for different regions by combining vegetation type distribution data to obtain dynamically adjusted fire temperature thresholds.

[0158] Based on a standardized temperature data matrix, the temperature value of each pixel is compared with the dynamic fire temperature threshold at the corresponding location using a pixel-by-pixel scanning algorithm combined with parallel processing technology to obtain temperature analysis results.

[0159] Based on the temperature analysis results, pixels whose temperature values ​​exceed the dynamic fire temperature threshold are marked as abnormal pixels by threshold judgment logic. At the same time, the coordinate position and the temperature value exceeding the limit of each abnormal pixel are recorded to generate the original abnormal pixel set.

[0160] Based on the original set of abnormal pixels, isolated abnormal pixels are removed by spatial continuity analysis algorithm, morphological opening operation is used to eliminate scattered noise, and abnormal pixels with spatial clustering characteristics are retained to obtain a denoised set of abnormal pixels.

[0161] Based on the abnormal pixel detection results of multiple consecutive frames of images, an abnormal pixel caused by transient interference is eliminated by a temporal consistency verification algorithm, and only abnormal pixels that continuously appear in multiple frames of images are retained to obtain a verified abnormal pixel set.

[0162] Based on a validated set of anomalous pixels, spatially adjacent anomalous pixels are clustered into continuous regions using a region growing algorithm. The area, shape, and temperature distribution characteristics of each continuous region are calculated, and a set of anomalous regions with these characteristics is output.

[0163] Based on the region growing algorithm combined with the abnormal pixel set, the spatially adjacent abnormal pixels are clustered to obtain several connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target.

[0164] Based on several connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target, the static contour features of the suspected heat source sub-targets are extracted; the static contour features include at least area and aspect ratio features.

[0165] The extracted static contour features are matched with the thermal imaging contour parameters of a preset fire heat source feature library. If the similarity between the static contour features of the current suspected heat source sub-target and any thermal imaging contour parameter in the fire heat source feature library is less than a first set standard, it is determined to be a non-fire heat source and excluded. If it is greater than the standard, dynamic discrimination is performed, specifically:

[0166] Based on the pixel position changes of the corresponding suspected heat source sub-targets in a series of consecutive initial suspected fire sub-region image sequences, the movement speed, movement direction, and expansion of the suspected heat source sub-targets are calculated. It should be further noted that the expansion of the suspected heat source sub-targets in this embodiment is mainly used to distinguish between real fires and interfering targets such as animal heat sources. Real fires usually exhibit the characteristic of continuous expansion of heat source sub-targets, while animal heat sources maintain a relatively stable area or exhibit irregular movement. The expansion of the suspected heat source sub-targets is obtained by analyzing the contour changes of the suspected heat source sub-targets in a series of consecutive initial suspected fire sub-region image sequences. Specifically, the contour of the sub-target in each frame image is extracted and its area is calculated. Then, the area expansion rate per unit time is calculated through temporal area change analysis. At the same time, the continuity of expansion is judged by combining the trend of contour shape change, thereby quantitatively evaluating the expansion characteristics of the heat source sub-target.

[0167] The movement speed and direction of the suspected heat source sub-targets obtained from the analysis are matched with the typical fire expansion patterns stored in the fire heat source feature database. If the corresponding matching degree is less than the second set standard and the expansion of the suspected heat source sub-target remains unchanged, it is determined to be a non-fire heat source and is excluded. At the same time, suspected heat source sub-targets that exceed the second set standard are marked as high-confidence suspicious sub-targets, and the pixel coordinate set of the corresponding high-confidence suspicious sub-targets is output.

[0168] The dual verification unit is used to extract temperature features along the thermal radiation gradient direction corresponding to the high-confidence suspicious sub-targets in continuous time through the thermal imaging recognition layer, and obtain the temperature gradient change rate and regional expansion continuity in the corresponding direction. When the temperature gradient change rate in the corresponding direction meets the fire temperature threshold and the matching degree between the regional expansion continuity and the fire expansion parameter distribution map corresponding to the fire heat source feature library is greater than the preset matching threshold, it is initially determined that there is a fire in the corresponding high-confidence suspicious sub-target, and the accuracy of the initial determination is output. When it is initially determined that there is a fire in the corresponding high-confidence suspicious sub-target, the visible light recognition layer is used to detect the smoke grayscale features of the high-confidence suspicious sub-target with fire by combining the accuracy of the initial determination. When smoke is detected, the smoke contour and offset trajectory are extracted to obtain the smoke contour offset trajectory.

[0169] When it is initially determined that there is a fire in the corresponding high-confidence suspicious sub-target, the visible light recognition layer is used to perform smoke grayscale feature detection on the high-confidence suspicious sub-target with fire based on the accuracy of the initial determination. When smoke is detected, the smoke contour and offset trajectory are extracted to obtain the smoke contour offset trajectory.

[0170] It should be further explained that the process of obtaining the smoke contour offset trajectory in this embodiment includes:

[0171] Based on the preliminary accuracy level and the coordinates of the thermal radiation center of the high-confidence suspicious sub-target, differentiated smoke detection parameters are set, and the detection area range with the thermal radiation center coordinates as the focus is determined. The accuracy level is divided into three levels: high, medium, and low, based on a preset interval. The high accuracy level uses conventional detection parameters, the medium accuracy level expands the pixel sampling range to cover a larger area around the suspicious target, and the low accuracy level enhances detail capture by improving the accuracy of grayscale feature extraction and increasing the sampling point density per unit area. The detection area range is a circular area with the thermal radiation center coordinates as the center and a preset length as the radius.

[0172] Based on historical monitoring data, common scene types, smoke types, and fog types in the target monitoring area, a multi-scene smoke and fog grayscale-temperature feature benchmark library is constructed. The historical monitoring data includes image data and smoke and fog sample data of different scene types. The benchmark library stores the grayscale feature benchmarks and temperature feature benchmarks of the samples according to scene type. The grayscale feature benchmarks include grayscale value range, grayscale uniformity, and edge blurring, while the temperature feature benchmarks include regional temperature mean, temperature gradient change, and ambient temperature difference.

[0173] Using the smoke detection parameters and detection area range, multi-scale grayscale sampling is performed within the detection area in the enhanced visible light image. The multi-scale grayscale sampling involves collecting pixel grayscale values ​​within circular windows of different radii to adapt to the spatial distribution characteristics of smoke of different concentrations. The mean grayscale value, grayscale variance, and edge grayscale jump value of the region are calculated. The edge grayscale jump value is obtained by calculating the maximum value of the grayscale difference between adjacent pixels and combined with the smoke grayscale benchmark of the current scene for screening to obtain potential smoke areas and their coordinate range in the visible light image.

[0174] The coordinate range of the potential smoke area is mapped to the synchronously acquired thermal imaging image using a coordinate mapping algorithm. The temperature features of the corresponding thermal imaging sub-region are extracted. The temperature features include the mean temperature of the region, the difference between the mean temperature of the region and the real-time ambient temperature, the temperature gradient distribution, and the proportion of high-temperature pixels. These features are then compared with the smoke temperature benchmark and fog temperature benchmark of the current scene. If the mean temperature, temperature gradient distribution, and proportion of high-temperature pixels all conform to the smoke temperature benchmark and have significant differences from the fog temperature benchmark, then the region is determined to have smoke temperature features, and the corresponding coordinates and temperature feature data are obtained.

[0175] Based on the grayscale feature data of the potential smoke area, the temperature feature data of the smoke temperature feature area, real-time environmental parameters and feature benchmark library, a joint criterion algorithm is used to distinguish smoke from fog.

[0176] The joint criterion algorithm first determines whether the grayscale feature data falls within the grayscale reference range of smoke and whether the temperature feature data conforms to the temperature reference of smoke. If so, it is determined to be a suspected smoke area. If the grayscale feature data is close to the grayscale reference of fog and the temperature feature data conforms to the temperature reference of fog, it is determined to be a fog area and is removed. When feature mismatch occurs, it is verified in combination with real-time environmental parameters. In high humidity environments, it is preferentially determined to be fog. Finally, the area that meets the dual features of smoke is retained as a valid suspected smoke area, and its coordinate correlation data in visible light and thermal imaging images is obtained.

[0177] Based on the coordinate association data of the effective suspected smoke area, contour extraction and matching are performed on multiple consecutive frames of visible light images and thermal imaging images. The smoke contour is extracted using the Canny edge detection algorithm with adaptive threshold. The contour irregularity and the contour perimeter to area ratio are calculated. The centroid offset trajectory of the same smoke contour in consecutive frames is tracked by the contour matching algorithm. The trajectory is calibrated using the offset direction of the temperature center in the thermal imaging image. Smoke contour data, calibrated contour offset trajectory and smoke drift direction are obtained.

[0178] Based on the temperature and contour change trends of the smoke region in multiple consecutive frames, temporal consistency verification is performed. The temporal consistency verification analyzes whether the temperature shows a stable or slowly increasing trend and whether the contour shows an expanding trend, and checks whether the two trends are synchronized. If the trends are synchronized and continue to reach a preset frame threshold, it is determined that there is a real fire, and the final fire determination result, smoke contour data, drift direction, and corrected determination accuracy are output.

[0179] The smoke profile offset trajectory and the wind direction data, temperature gradient direction and regional expansion continuity in the meteorological data are used to determine the directional consistency. If the corresponding directional determinations are consistent, it is determined that there is a fire.

[0180] The fire location unit is used to obtain the location point and real-time burned area of ​​the corresponding fire by combining the position of the high-confidence suspicious sub-target with fire in the initial monitoring sub-region sequence with the corresponding temperature gradient change rate, smoke profile offset trajectory and regional expansion continuity through a dynamic regional integration layer.

[0181] It should be further explained that the process of obtaining the location of the corresponding fire and the real-time burned area in this embodiment includes:

[0182] Based on the location of high-confidence suspicious sub-targets indicating fire in the initial monitoring sub-region sequence, the rate of change of temperature gradient in the corresponding direction, and the current scene type, the spatial distribution model of the fire area in the corresponding scene is called from the multi-scene feature benchmark library. Combined with the initial starting coordinates of the smoke contour offset trajectory, a minimum bounded region containing the core area of ​​the sub-target and the initial smoke contour is delineated as the initial dynamic integration region. Among them, a differentiated boundary delineation strategy is adopted for different scenes: for foggy scenes, since the fog temperature is close to the ambient temperature, the boundary of the initial region is expanded outward by a preset proportion; for sunny scenes, based on the significant difference between the smoke and background temperatures, the boundary is delineated according to 1.2 times the range of the initial smoke contour, thereby obtaining the spatial coordinate range of the initial dynamic integration region.

[0183] Based on the spatial coordinate range of the initial region of dynamic integration, the temperature gradient change rate data, smoke contour offset trajectory data, and region expansion continuity data from different sensors are synchronized in time using a timestamp alignment algorithm; the feature data are unified to the same coordinate system using a spatial coordinate mapping algorithm; based on the synchronized multi-source feature data, a weight allocation model is used to dynamically allocate integration weights to each sub-region within the region. The allocation criteria include: higher temperature gradient change rate, higher weight; higher consistency between smoke contour offset trajectory and temperature gradient direction, higher weight; and more stable region expansion continuity, higher weight. At the same time, the weight ratio of temperature features is increased for nighttime scenes, and the weight ratio of temperature and smoke features is balanced for cloudy scenes, thereby obtaining the dynamic integration weight value of each sub-region.

[0184] Based on the initial dynamic integration region, the dynamic integration weight values ​​of each sub-region, and the synchronized multi-source feature data, a sliding window integration algorithm is used to perform integration operations on each pixel within the region: the sliding window traverses each pixel within the region, and the weighted sum of the temperature gradient change rate, the proportion of smoke contour pixels, and the region expansion rate within the window is calculated as the integration value of that pixel; an extreme value detection algorithm is used to extract the pixel with the highest integration value as the core candidate point of the fire; combined with the integration results of multiple consecutive frames, the position change of the candidate point is tracked, and if its position deviation is less than a preset range within a preset number of consecutive frames, the point is determined as the fire location point, and its precise coordinates are obtained;

[0185] Based on the coordinates of the fire location, the distribution of integral values ​​within the dynamic integration area, and the time-series data of the smoke contour offset trajectory, a threshold segmentation algorithm is used to classify areas with integral values ​​higher than the fire determination threshold as candidate fire areas. A contour extraction algorithm is used to extract the boundary contour of the candidate fire area, and the boundary is calibrated by combining the expansion trajectory of the smoke contour in consecutive frames: when the expansion direction of the smoke contour is consistent with the diffusion direction of the temperature gradient, the intersection contour of the two is used as the final fire boundary; when the directions are inconsistent, the contour corresponding to the diffusion direction of the temperature gradient is used as the reference to correct the smoke contour. The number of pixels in the calibrated boundary contour is counted by a pixel counting algorithm, and the real-time fire area is calculated based on the pixel size conversion relationship of the monitored image.

[0186] Based on the rate of change of the real-time burned area, the stability of the fire location, and the current scene type, a feature verification model is used for secondary verification: for high humidity scenes, it verifies whether the temperature of the burned area is continuously higher than the ambient temperature by a preset range; for strong wind scenes, it verifies whether the smoke outline offset trajectory is consistent with the wind direction; if the verification passes, the coordinates of the fire location, the real-time burned area, and the confidence level calculated based on the stability of the temperature gradient rate of change and the continuity of the smoke outline expansion are output; if the verification fails, it returns to redefining the initial region of dynamic integration and re-executes the integration calculation process.

[0187] The dynamic assessment unit is used to obtain the severity and stage of the fire based on the real-time burned area, the rate of change of temperature gradient in the corresponding direction, the rate of regional expansion, the smoke concentration, and the expert assessment layer.

[0188] It should be further explained that the process of obtaining the severity and stage of the fire in this embodiment includes:

[0189] Based on the current monitoring scenario type and multi-source data collected by the dynamic evaluation unit, a scenario-based data filtering algorithm is used to preprocess the data. The multi-source data includes temperature data, temperature gradient change rate data, area expansion rate data, and smoke concentration data. The preprocessing includes: for foggy scenarios, performing temperature threshold filtering, and removing records with temperature values ​​lower than the preset range of the ambient temperature based on the characteristic that fog temperature is close to the ambient temperature; for nighttime scenarios, performing signal enhancement, amplifying the temperature gradient change rate and area expansion rate data to compensate for the weakness of visible light data; for windy scenarios, performing wind direction correction, introducing real-time wind direction data, and eliminating the interference of wind on the direction of area expansion rate through vector operations; finally, using a data normalization algorithm, mapping all data to the same preset numerical range to achieve uniformity of magnitude and obtain effective evaluation data after scenario-based preprocessing.

[0190] Based on the aforementioned effective evaluation data and a multi-scene feature benchmark library, a scenario-based differentiation algorithm is used to verify suspected smoke areas. The multi-scene feature benchmark library is derived from historical monitoring data and stores characteristic standard values ​​of smoke and fog under different scenarios. The verification includes: for foggy scenarios, calculating the difference between the average temperature of the suspected area and the average ambient temperature, and comparing it with the range of smoke temperature difference in the multi-scene feature benchmark library; if they match, it is determined to be valid smoke. For sunny scenarios, calculating the gray-scale jump value at the edge of the suspected area and comparing it with the range of gray-scale jump values ​​at the edge of the smoke in the multi-scene feature benchmark library; if they match, it is determined to be valid smoke. For nighttime scenarios, analyzing whether there is a temperature decrease relationship from the center to the surrounding areas in the suspected area; if so, it is determined to be valid smoke. Through the above verification, effective smoke feature data after removing fog interference is obtained, including smoke concentration, outline, and dispersion direction data.

[0191] Based on the effective assessment data and the effective smoke feature data, four core assessment indicators are calculated using an indicator quantification algorithm. Indicator quantification includes: calculating the actual real-time burned area based on the number of pixels in the candidate burned area and pixel scale parameters; calculating the temperature gradient change rate based on the temperature gradient difference and time interval between the same sub-regions in consecutive frames; calculating the area expansion rate based on the difference and time interval between the actual burned areas in consecutive frames; and calculating the smoke concentration quantification value based on the average grayscale value of smoke area pixels and the concentration-grayscale mapping relationship in a multi-scene feature benchmark library, or based on the proportion of smoke pixels. Furthermore, an indicator calibration algorithm is used to correct the above indicators in conjunction with the current scene type. For example, in a windy weather scenario, the wind direction influence coefficient constructed from wind force level is used to correct the area expansion rate, thereby obtaining a standardized set of core fire assessment indicators.

[0192] Based on the fire stage-indicator weight mapping relationship preset by the expert evaluation layer, the weight is dynamically assigned to the standardized core fire evaluation indicators through a weight adaptation algorithm. The fire stage-indicator weight mapping relationship is set by expert experience based on the historical fire evolution pattern and is used to specify the weight value of each core evaluation indicator under different fire stages. The weight calculation algorithm is used to multiply the standardized quantitative value of each indicator by the weight value of its corresponding stage and then sum them to obtain the weighted comprehensive score of fire evaluation.

[0193] Based on the comprehensive fire assessment score and the preset judgment criteria of the expert assessment layer, the severity and stage of the fire are determined by a comprehensive judgment algorithm. The judgment criteria include a score threshold and a stage characteristic trend. The comprehensive judgment includes: comparing the comprehensive fire assessment score with the preset score threshold to determine the severity level of the fire as mild, moderate, severe, or extremely severe; extracting the core indicator change trend of multiple consecutive frames and matching it with the preset characteristic trend of each stage to determine whether the fire is in the initial, development, vigorous, or decaying stage; verifying the evaluation results of consecutive frames through a verification algorithm. If the judgment results are consistent within a preset number of consecutive frames, the final state is confirmed; otherwise, the indicator calculation and weight adaptation steps are re-executed.

[0194] Based on the current scenario type, the final judgment result is modified according to the scenario. The modification includes: for high humidity scenarios, analyzing the fluctuation of smoke concentration, and if the fluctuation is too large, the severity level is downgraded; for windy scenarios, comparing the area expansion rate value after wind direction correction with the preset development stage rate range, and if the rate is too low, adjusting the stage judgment to the initial combustion stage; finally, through the result integration algorithm, the complete evaluation result consisting of the modified fire severity, stage, core evaluation index set and effective smoke characteristic data is output.

[0195] It should be further explained that the process of obtaining fire prevention strategy information in this embodiment includes:

[0196] Based on the current monitoring scenario type, fire severity, fire stage, and previously acquired effective smoke feature data, a scenario-based calibration algorithm is used to call the temperature profile difference data of smoke and fog in a multi-scenario feature benchmark library, eliminate the warning level deviation caused by fog interference, and obtain the calibrated target warning level.

[0197] Based on historical forest fire prevention and control cases and expert experience data, a fire prevention strategy information database is constructed in three dimensions according to scenario type, warning level, and fire stage using a structured modeling algorithm. Under each strategy entry in the fire prevention strategy information database, appropriate smoke and temperature characteristic conditions are marked. The strategy entries are then bound to the characteristic conditions using an association algorithm to obtain a hierarchical structured fire prevention strategy information database.

[0198] Based on the calibrated target warning level, current scenario type, and fire stage, an index building algorithm is used to construct a multi-dimensional index in the fire prevention strategy information database, including scenario dimension, warning level dimension, fire stage dimension, smoke feature dimension, and temperature feature dimension. A deep matching algorithm is used to compare the smoke and temperature features of the current fire with the feature conditions marked on the strategy entries in the fire prevention strategy information database, and the candidate strategy set with the highest similarity is selected.

[0199] Based on the current scenario type and the set of candidate strategies, the feasibility of each candidate strategy is checked by using an adaptation verification algorithm, combined with the smoke drift direction and temperature gradient change rate in the current scenario. Candidate strategies with insufficient scenario adaptability are eliminated, and the strategy with the best adaptability is retained as the target fire prevention strategy candidate.

[0200] Based on real-time updated data on changes in smoke dispersion direction and temperature gradient change rate, a strategy adjustment algorithm is used to fine-tune the response measures in the candidate fire prevention strategies. The final adapted fire prevention strategy information is then transmitted to the user terminal to obtain executable fire prevention strategy information.

[0201] This implementation first utilizes radiometric calibration and dynamic temperature compensation techniques to calculate suitable fire temperature thresholds for each pixel under different vegetation and meteorological environments. This step fundamentally eliminates the benchmark drift problem caused by equipment differences and environmental fluctuations, providing a stable and reliable data foundation for subsequent analysis. Subsequently, through efficient parallel pixel scanning and spatial continuity analysis, abnormal heat source sub-targets are quickly located. Furthermore, by innovatively combining static contour matching and dynamic characteristic analysis, real fires are effectively distinguished from common interference sources such as animals and hot exhaust gases. This design significantly reduces the false alarm phenomenon commonly found in traditional single temperature threshold methods. After confirming a suspected heat source, the system enters a dual verification phase, which is crucial for improving the reliability of the judgment. The thermal imaging recognition layer not only analyzes absolute temperature but also focuses on extracting deeper features such as the rate of change and directionality of temperature gradients, as well as the continuity of regional expansion, enabling it to accurately identify fire sources. Early, subtle signs of a fire are more sensitive; the visible light recognition layer simultaneously extracts multi-scale smoke features, constructs a comprehensive smoke-fog feature benchmark library and applies a joint criterion algorithm, achieving accurate smoke identification and effective elimination of fog interference at the multi-source data level; finally, the consistency of smoke dispersion trajectory, temperature gradient direction and meteorological wind direction data is fused and judged, and the existence of a fire is jointly corroborated by multiple independent evidence chains, making the preliminary judgment result both highly accurate and highly confident; for confirmed fires, the system's dynamic region integration layer, by fusing spatiotemporally synchronized multi-source feature data (temperature gradient, smoke outline, expansion rate) and introducing a weight allocation model based on physical laws (such as assigning high weights to high-temperature areas and expansion consistency areas), achieves accurate positioning of the fire core and dynamic and accurate calculation of the burned area, with an accuracy far exceeding that of traditional simple outline delineation methods. Based on this, the dynamic assessment unit relies on the fire stage-indicator weight mapping relationship constructed by the expert knowledge base to perform weighted comprehensive evaluation of the standardized core indicators (burned area, rate of change, expansion rate, and smoke concentration). It also introduces a real-time correction mechanism based on scenario type (such as downgrading the level in high humidity and correcting the expansion rate in windy weather). This makes the assessment conclusions on the severity and development stage of the fire not only quantitative and objective, but also highly scenario-specific, truly reflecting the actual situation of the fire. Finally, based on the above accurate situational awareness results, the system drives intelligent decision generation. The fire prevention strategy information database adopts a three-dimensional structured design of scenario-level-stage, and through deep matching and adaptation verification algorithms, it selects and fine-tunes the most suitable executable strategies for the current fire characteristics from massive historical cases and expert experience, ensuring the scientific and practical nature of the disposal recommendations.

[0202] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. An all-weather intelligent forest fire monitoring system suitable for complex environments, characterized in that: include: Acquisition module, selection enhancement module, dual recognition module; The acquisition module determines an initial monitoring path based on the target monitoring area and monitors along the initial monitoring path based on preset initial discrimination indicators. It acquires images containing at least some initial discrimination indicator anomalies and their corresponding initial timestamps. Simultaneously, it adjusts the initial monitoring path according to the gradient change direction of the initial discrimination indicators in the images containing at least some initial discrimination indicator anomalies, and acquires an image sequence containing at least some initial discrimination indicator anomalies as an initial fire suspected sub-area image sequence. The selection enhancement module is used to determine the scene type and scene monitoring difficulty coefficient of the target monitoring area based on the meteorological data collected in real time in the target monitoring area and the scene discrimination evaluation model. Based on the scene type and scene monitoring difficulty coefficient, it calls the pre-trained local enhancement filtering model library and combines it with the initial timestamp to synchronously enhance the image sequence of the initial suspected fire sub-area to obtain the enhanced image sequence of the initial suspected fire sub-area. The dual recognition module is used to perform layered analysis and judgment of the fire situation in the suspected fire area based on the enhanced initial fire suspected sub-area image sequence and the preset dual recognition model. When it is determined to be a fire, the location of the corresponding suspected area is determined. At the same time, the severity and stage of the corresponding fire are obtained based on the temperature change status and real-time area size of the suspected area in the initial fire suspected sub-area image sequence at the corresponding location over a continuous time. The dual recognition module includes a filtering unit and a dual verification unit; the dual recognition model includes an abnormal area filtering layer and a visible light recognition layer; the filtering unit is used to perform temperature analysis on the thermal radiation gradient of each pixel in the initial suspected fire sub-region based on the enhanced image of the initial suspected fire sub-region and a preset fire temperature threshold and the abnormal area filtering layer, mark all pixels with temperature values ​​higher than the fire temperature threshold as abnormal pixels, obtain an abnormal pixel set, and output the pixel coordinate set of the corresponding high-confidence suspected sub-targets; The dual verification unit is used to detect smoke grayscale features of the high-confidence suspicious sub-target when it is initially determined that there is a fire. It uses the visible light recognition layer in combination with the accuracy of the initial determination to detect smoke grayscale features of the high-confidence suspicious sub-target. When smoke is detected, the smoke contour and offset trajectory are extracted to obtain the smoke contour offset trajectory. The smoke profile offset trajectory and the wind direction data, temperature gradient direction and regional expansion continuity in the meteorological data are used to determine the directional consistency. If the corresponding directional determinations are consistent, it is determined that there is a fire.

2. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 1, characterized in that, The all-weather intelligent forest fire monitoring system also includes an early warning module; The early warning module is used to obtain fire prevention strategy information based on the severity and stage of the fire, combined with a preset early warning level and a fire prevention strategy information database, through a deep matching indexing algorithm.

3. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 2, characterized in that, The acquisition module includes a region segmentation unit and an initial path unit; The region segmentation unit is used to perform preliminary region segmentation based on the geographical features of the target monitoring area, the deployment and distribution of monitoring equipment, and the monitoring range characteristics, combined with a grid partitioning algorithm, to obtain an initial monitoring sub-region sequence. The initial path unit is used to obtain the initial monitoring path based on the frequency and severity of historical fires in each initial monitoring sub-region, the geographical monitoring difficulty of the sub-region, the density and importance of attachments in the sub-region combined with the path algorithm, and the cost loss function constructed based on the density, importance and monitoring coverage of attachments in the sub-region.

4. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 3, characterized in that, The acquisition module also includes an initial discrimination unit and a path adjustment unit; The initial discrimination unit is used to construct an initial discrimination index based on the monitoring images collected along the initial monitoring path and the thermal radiation gradient change value. When the thermal radiation gradient value of some areas in the monitoring image at any time point does not meet the preset standard scene thermal radiation gradient distribution map, the sub-region corresponding to the current monitoring image is taken as the initial suspected fire sub-region, and the timestamp of the first abnormal monitoring image and the initial location point corresponding to the abnormal initial discrimination index are recorded. The standard scenario thermal radiation gradient distribution map is constructed by combining the radiation gradient values ​​collected at different time points under different weather types in the target monitoring area under the condition that there is no fire, and the results with statistical analysis algorithms. The path adjustment unit is used to adjust the direction of the initial monitoring path in real time based on the edge structure features of the monitoring image where some areas have abnormal initial discrimination indicators, as well as the thermal radiation gradient extension direction of the sub-regions corresponding to the abnormal initial discrimination indicators, and the rate and state of change of the thermal radiation gradient under the corresponding extension direction.

5. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 4, characterized in that, The acquisition module also includes a suspicious sub-region determination unit and a region adjustment unit; The suspicious sub-region determination unit is used to determine the area size and corresponding edge features of the corresponding initial fire suspicious sub-region based on the edge structure features of the sub-regions corresponding to the initial discrimination index anomaly in the image sequence of the initial fire suspicious sub-regions collected corresponding to the adjusted initial monitoring path, the rate of change of thermal radiation gradient under the extension direction of the sub-regions, and the thermal radiation gradient distribution map of the standard scene. The region adjustment unit compares the area size and corresponding edge features of the initially suspected fire sub-region determined in real time with the initial monitoring sub-region. If the initially suspected fire sub-region is larger than the initial monitoring sub-region, the unit uses the proportion of the initially suspected fire sub-region in each initial monitoring sub-region to adjust the initial monitoring sub-regions covered by the initially suspected fire sub-region that have adjacent edges, thereby obtaining the adjusted monitoring sub-region sequence.

6. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 5, characterized in that, The selection enhancement module includes a scene discrimination unit and a difficulty assessment unit; The scene discrimination unit is used to obtain the scene type label of the target demand monitoring area by combining the meteorological data synchronously collected within the target demand monitoring area for the current preset time length with the preset scene parameter library and matching algorithm. The difficulty assessment unit is used to obtain the scene monitoring difficulty coefficient by combining the scene type label of the target required monitoring area with the correlation influence matrix of the meteorological data collected in real time under the corresponding scene on the collected data through the scene discrimination assessment model. The scenario discrimination and evaluation model is constructed by combining the correlation influence matrix built on the meteorological data collected in real time under the corresponding scenario and the degree of influence of the meteorological data under each scenario on the target collection index, and trained with expert experience algorithm. It is used to judge the degree of influence of meteorological data on the target collection data under the current scenario in real time.

7. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 6, characterized in that, The selection enhancement module also includes an enhancement processing unit; The enhancement processing unit is used to perform real-time enhancement processing on the initial suspected fire sub-region image sequence by combining the scene monitoring difficulty coefficient with the accuracy of the hierarchical analysis and judgment corresponding to the dual recognition module and the accuracy of the fire severity, and by combining the preset local enhancement filtering model library with the matching algorithm and the preset matching threshold and the timestamp of the first abnormal monitoring image, and calling at least one local enhancement filtering model that meets the matching threshold to obtain the enhanced initial suspected fire sub-region image sequence. The real-time enhancement process includes: adjusting the kernel parameters and enhancement magnitude of the local enhancement filter model based on the magnitude and direction of the change in thermal radiation gradient of the initially suspected fire sub-region.

8. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 7, characterized in that, The filtering unit is also used for: Based on the region growing algorithm combined with the abnormal pixel set, the spatially adjacent abnormal pixels are clustered to obtain multiple connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target. Based on multiple connected suspected heat source sub-targets and the initial coordinate set of each suspected heat source sub-target, the static contour features of the suspected heat source sub-targets are extracted. The static contour features include at least area and aspect ratio features; The extracted static contour features are matched with the thermal imaging contour parameters of a preset fire heat source feature library. If the similarity between the static contour features of the current suspected heat source sub-target and any thermal imaging contour parameter in the fire heat source feature library is less than a first set standard, it is determined to be a non-fire heat source and excluded. If it is greater than the standard, dynamic discrimination is performed, specifically: Based on the pixel position changes of the corresponding suspected heat source targets in the image sequence of multiple consecutive initial suspected fire sub-regions, the motion speed, motion direction and expansion of the suspected heat source targets are calculated. The movement speed and direction of the suspected heat source sub-targets obtained from the analysis are matched with the typical fire expansion patterns stored in the fire heat source feature database. If the corresponding matching degree is less than the second set standard and the expansion of the suspected heat source sub-target remains unchanged, it is determined to be a non-fire heat source and is excluded. At the same time, suspected heat source sub-targets that exceed the second set standard are marked as high-confidence suspicious sub-targets.

9. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 8, characterized in that, The dual recognition model also includes a thermal imaging recognition layer; The dual verification unit is also used to extract temperature features along the thermal radiation gradient direction corresponding to the high-confidence suspicious sub-targets in continuous time through the thermal imaging recognition layer, and obtain the temperature gradient change rate and regional expansion continuity in the corresponding direction. When the temperature gradient change rate in the corresponding direction meets the fire temperature threshold and the matching degree between the regional expansion continuity and the fire expansion parameter distribution map corresponding to the fire heat source feature library is greater than the preset matching threshold, it is preliminarily determined that there is a fire in the corresponding high-confidence suspicious sub-targets, and the accuracy of the preliminary determination is output.

10. The all-weather intelligent forest fire monitoring system suitable for complex environments as described in claim 9, characterized in that, The dual identification module also includes a fire location unit and a dynamic evaluation unit, and the dual identification model also includes a dynamic regional integration layer and an expert evaluation layer. The fire location unit is used to obtain the location point and real-time burned area of ​​the corresponding fire by combining the position of the high-confidence suspicious sub-target with fire in the initial monitoring sub-region sequence with the corresponding temperature gradient change rate, smoke profile offset trajectory and regional expansion continuity through a dynamic regional integration layer. The dynamic assessment unit is used to obtain the severity and stage of the fire based on the real-time burned area, the rate of change of temperature gradient in the corresponding direction, the rate of regional expansion, the smoke concentration, and the expert assessment layer.

Images