Fire monitoring and early warning method and device based on big data analysis

Abstract

The embodiment of the invention discloses a fire monitoring and early warning method and device based on big data analysis, and belongs to the technical field of data analysis. According to the embodiment of the invention, multi-source data such as satellite remote sensing heat source images, meteorological environment parameters, historical fire event records and geographic information raster data are integrated, a space-time associated dynamic monitoring system is constructed, and environmental anomaly symptoms before fire occurrence can be comprehensively captured. Wherein the space-time reference difference of different source data is eliminated through the space-time fusion processing, so that the heat source distribution, the meteorological change and the historical fire mode form multi-dimensional association under a unified geographic framework, the sensitivity and the reliability of early fire point identification are remarkably improved, and the subsequent accurate fire risk prediction is realized.

Description

Technical Field

[0001] The embodiments of the present invention belong to the field of data analysis technology, and specifically relate to a fire monitoring and early warning method and device based on big data analysis. Background Art

[0002] Traditional forest fire monitoring technology mainly relies on a single data source from satellite remote sensing or ground sensors to identify fire points, which has problems such as large monitoring blind spots and high false alarm rates.

[0003] In existing technologies, the simple superposition of meteorological data and heat source images often ignores the differences in temporal and spatial benchmarks, resulting in a time lag of several hours in fire situation analysis. Most fire prediction models use linear regression or empirical formulas, which are difficult to effectively couple dynamic parameters such as terrain slope and vegetation moisture content, and the prediction error increases significantly in scenarios with sudden changes in wind direction.

[0004] Although some existing multi-source data fusion methods integrate visible light and infrared data, they do not solve the spatiotemporal correlation modeling of historical fire patterns and real-time meteorological elements, resulting in a lack of regional adaptability in risk prediction. In addition, the existing early warning system uses a fixed threshold trigger mechanism and cannot dynamically adjust the emergency response scope according to the evolution of the fire. These technical defects make it difficult for existing technologies to ensure the sensitivity and reliability of early fire point identification. Summary of the invention

[0005] The embodiments of the present invention provide a fire monitoring and early warning method and device based on big data analysis, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0006] The embodiment of the present invention provides a fire monitoring and early warning method based on big data analysis, which is applied to a fire monitoring and early warning device. The method includes: Collecting multi-source fire monitoring data, the multi-source fire monitoring data including satellite remote sensing heat source images, meteorological environmental parameters, historical fire event records and geographic information raster data; Performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporally associated fire dynamic data set, wherein the spatiotemporal fusion processing includes timestamp alignment, geocoding mapping, and abnormal data removal; The trained target fire spread prediction model is called to process the fire dynamic data set, and the fire risk index and the spread trend probability distribution of the target area are output; wherein the target fire spread prediction model includes a spatiotemporal convolution layer, a three-dimensional convolution kernel and a graph convolution network node processing module, and the multi-dimensional fire feature matrix of the fire dynamic data set is extracted through the spatiotemporal convolution layer, and is spliced ​​with the fire feature vector set to generate a model input tensor, and the model input tensor is traversed according to the three-dimensional convolution kernel to generate a three-dimensional fire spread field, and the graph convolution network node processing module is used to calculate the fire transmission intensity between each geographic grid unit and the adjacent units based on the three-dimensional fire spread field, and the fire transmission probability matrix and the fire spread prediction path between the nodes are generated, and the fire risk index and the spread trend probability distribution are generated based on the fire transmission probability matrix and the fire spread prediction path; A multi-level warning signal is generated according to the fire risk index and the probability distribution of the spreading trend, and is dynamically updated to the regional emergency response platform.

[0007] The embodiment of the present invention integrates satellite remote sensing heat source images, meteorological environmental parameters, historical fire event records, geographic information raster data and other multi-source data to build a dynamic monitoring system with temporal and spatial correlation, which can comprehensively capture the abnormal environmental signs before the fire occurs. Among them, the temporal and spatial fusion processing eliminates the temporal and spatial reference differences of data from different sources, so that the heat source distribution, meteorological changes and historical fire patterns form a multi-dimensional correlation under a unified geographical framework, significantly improving the sensitivity and reliability of early fire point identification, thereby realizing subsequent accurate fire risk prediction.

[0008] In one implementation, performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporally correlated fire dynamic data set includes: Extracting the fire point pixel coordinates and the radiation intensity value in the satellite remote sensing heat source image, and mapping the fire point pixel coordinates and the radiation intensity value to the geographic coordinate system of the target area based on a geographic coordinate conversion algorithm; The meteorological environment parameters are interpolated piecewise according to the timestamps to generate a continuous meteorological parameter surface matching the geographic coordinate system; Performing spatial cluster analysis on the historical fire event records to generate a probability density map of fire-frequently occurring areas, and superimposing the probability density map on the geographic coordinate system; The fire point pixel coordinates, the continuous meteorological parameter surface and the probability density map are dynamically weighted fused through a spatiotemporal sliding window algorithm to generate the fire situation dynamic data set.

[0009] The embodiment of the present invention adopts geocoding mapping and dynamic weighted fusion technology to effectively solve the problem of spatial resolution mismatch of multi-source data. By accurately mapping the fire point pixels to the geographic coordinate system and combining the meteorological parameter surface interpolation, a continuous dynamic monitoring field is generated. Spatial clustering analysis is superimposed on the probability density of fire-prone areas, so that the data set not only reflects the real-time fire situation, but also dynamically adjusts the monitoring weights in combination with historical laws, enhancing the ability to capture the evolution trend of fire conditions in key areas.

[0010] In one implementation, before calling the trained target fire spread prediction model to process the fire dynamics dataset, the method further includes: Generate a fire characteristic vector set, wherein the fire characteristic vector set includes a real-time fire point spread rate, a surface vegetation flammability index, a wind speed and wind direction correlation factor, and a terrain slope flame retardancy coefficient; A convolutional neural network is used to extract features from the fire dynamic data set to generate a multi-dimensional fire feature matrix; Performing tensor splicing on the multidimensional fire condition feature matrix and the fire condition feature vector set to generate model input data; Among them, the target fire spread prediction model is trained through a generative adversarial network, using historical fire data as real samples to generate simulated fire spread paths, and the model parameters are optimized through a discriminant network.

[0011] The embodiment of the present invention significantly improves the accuracy of fire spread prediction by constructing a multi-dimensional fire feature matrix and using an adversarial generative network training model. Feature tensor splicing combines real-time observation data with expert experience parameters, enabling the model to simultaneously learn physical laws and actual fire evolution patterns. The adversarial training mechanism forces the generator to simulate the distribution of real fire paths, and the feedback optimization of the discriminator enhances the model's generalization ability to complex environmental coupling effects.

[0012] In one implementation, the training process of the target fire spread prediction model includes: Obtaining fire spread trajectory data of a historical time series and a snapshot of environmental parameters at a time point corresponding to the fire spread trajectory data; Creating a three-dimensional fire diffusion field, the three-dimensional fire diffusion field including a spatial diffusion velocity field, a combustible consumption rate field and a thermal radiation attenuation field; A graph convolutional network is used to model the spatiotemporal relationship of the three-dimensional fire diffusion field, and a fire transmission probability matrix between nodes is generated; Optimizing the weight parameters of the fire transmission probability matrix by a reinforcement learning algorithm so as to maximize the similarity between the predicted path output by the target fire spread prediction model and the actual fire spread trajectory; The optimized weight parameters are solidified into the reasoning module of the target fire spread prediction model.

[0013] The construction of the three-dimensional fire diffusion field in the embodiment of the present invention is combined with graph convolutional network modeling to achieve a deep analysis of the fire propagation mechanism. The fire transmission probability matrix is ​​optimized through reinforcement learning, so that the model can adaptively adjust the dynamic relationship between spatial diffusion speed, combustible consumption and thermal radiation attenuation. This modeling method effectively captures the nonlinear effects of complex factors such as terrain undulation and vegetation distribution on fire propagation, greatly improving the spatiotemporal consistency between the predicted path and the actual disaster situation.

[0014] In one implementation, generating a multi-level warning signal according to the fire risk index and the spread trend probability distribution includes: The warning level is divided according to the numerical range of the fire risk index; wherein the first-level warning corresponds to a risk threshold in the first proportional range of the historical peak value, the second-level warning corresponds to a risk threshold in the second proportional range of the historical peak value, and the third-level warning corresponds to a risk threshold in the third proportional range of the historical peak value; Calculating the time evolution curve of the fire wave and range based on the probability distribution of the spreading trend, and determining the effective time window of each warning level according to the time evolution curve; Combining and encoding the warning level and the effective time window to generate a warning instruction set including the coordinates of the geo-fence; The warning instruction set is distributed to the terminal devices within the corresponding geographic fence through the edge computing node, and the emergency response protocol is activated.

[0015] The dynamic generation mechanism of multi-level warning signals in the embodiment of the present invention realizes the refined control of risk levels. The warning level is determined by dividing the historical peak ratio interval, and the effective time window is calculated in combination with the fire evolution curve, so that the warning instruction reflects both the risk intensity and the time dimension information. Geographic fence coding and edge computing node distribution technology ensure that the warning information can accurately reach the terminal equipment in the target area, buying precious time for emergency response.

[0016] In one implementation, the dynamic updating to the regional emergency response platform includes: Real-time monitoring of the rate of change of the fire risk indicator, and triggering a re-evaluation of the warning level when the rate of change of the fire risk indicator exceeds a preset critical value; Adjusting the geo-fence radius and evacuation route planning of the warning instruction set according to the re-evaluation result; Match the adjusted warning instruction set with the emergency resource database to generate an emergency plan including a material dispatch plan and a rescue force deployment diagram; The digital twin system is used to simulate the implementation effects of different emergency plans, and the emergency plan with the smallest quantitative loss is selected as the target response strategy.

[0017] The dynamic update mechanism in the embodiment of the present invention establishes an adaptive adjustment system for early warning levels by monitoring the rate of change of risk indicators in real time. The synergy between emergency resource matching and digital twin simulation technology enables dynamic optimization of evacuation route planning and rescue deployment as the fire situation evolves. This mechanism effectively solves the static defects of traditional emergency plans and significantly reduces the decision-making lag of disaster response through continuous iterative emergency plan comparison and selection.

[0018] In one implementation, the method further includes: An attention mechanism is embedded in the target fire spread prediction model to dynamically adjust the contribution weights of different environmental parameters to fire spread prediction; Extracting meteorological mutation characteristics at the current moment, wherein the meteorological mutation characteristics include the probability of dry thunderstorm occurrence and the air humidity drop rate; When it is detected that the meteorological mutation feature exceeds a preset sensitivity threshold, the activation intensity of the corresponding feature channel in the target fire spread prediction model is enhanced; Based on the enhanced target fire spread prediction model output, the calculation indication of the fire risk index is modified to generate an emergency warning signal under mutation conditions; The step of enhancing the activation intensity of the corresponding feature channel in the target fire spread prediction model includes: Based on the causal relationship diagram between meteorological parameter mutation events and historical fire cases, identify the key trigger factor combination; The sensitivity coefficient of each feature channel to the model output result is calculated through the gradient back propagation algorithm; Applying an exponential weight gain to the feature channels whose sensitivity coefficients exceed a sensitivity threshold; During the model reasoning process of the target fire spread prediction model, the output fluctuation of the gain channel is monitored in real time, and the model self-correction mechanism is started when the fluctuation amplitude exceeds the safety range.

[0019] The prediction model embedded with the attention mechanism in the embodiment of the present invention has the ability to respond to sudden changes in the environment. By adjusting the activation intensity of the feature channel, the decision weight of key meteorological parameters is dynamically improved. Causal correlation analysis and sensitivity coefficient calculation ensure that the model automatically strengthens the relevant feature analysis capabilities under extreme meteorological conditions such as dry thunderstorms and sudden drops in humidity. The self-correction mechanism timely corrects the model parameters by monitoring output fluctuations, ensuring the stability and reliability of the early warning system in a sudden change environment.

[0020] In one implementation, the method further includes: Deploy drone swarms to conduct real-time thermal imaging scans of areas prone to fire risk, and obtain surface temperature distribution maps with sub-meter accuracy; Performing multi-scale feature matching on the surface temperature distribution map and the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points; The target fire spread prediction model is used to simulate the diffusion path of the potential hidden fire point coordinates to generate a hot spot map of the early intervention area; Dispatching firefighting resources to carry out preventive isolation zone construction according to the hot spot map, and updating the preventive isolation zone construction strategy to the regional emergency response platform; The multi-scale feature matching of the surface temperature distribution map with the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points includes: Performing wavelet transform analysis on the surface temperature distribution map to extract frequency domain characteristics of abnormal temperature fluctuations; Creating a hidden fire point detection model based on a residual neural network, and inputting the abnormal temperature fluctuation frequency domain features and the acquired visible light image texture features into the hidden fire point detection model; Migrating the pre-trained fire point recognition knowledge to the hidden fire point detection model through a transfer learning algorithm; The hidden fire point detection model is used to process the temperature anomaly fluctuation frequency domain characteristics, the visible light image texture characteristics and the fire point identification knowledge, and output a probabilistic hidden fire point location confidence map; The hidden fire point location confidence map is spatially intersected and merged with the combustible material distribution layer in a preset geographic information system to determine a target hidden fire point coordinate set.

[0021] The multi-scale feature matching technology of drone thermal imaging and satellite data in the embodiment of the present invention effectively improves the recognition accuracy of potential hidden fire points. The frequency domain features of temperature anomalies are extracted by wavelet transform, and combined with residual network transfer learning, the system can penetrate the surface cover to identify deep hidden fires. The early generation of preventive isolation belt construction strategies transforms traditional post-disaster disposal into pre-emptive prevention and control, significantly reducing the potential risk of fire outbreaks.

[0022] In one implementation, the hidden fire point detection model is used to process the temperature anomaly fluctuation frequency domain features, the visible light image texture features, and the fire point identification knowledge, and output a probabilistic hidden fire point location confidence map, including: The temperature anomaly fluctuation frequency domain feature is input into the first convolutional layer of the residual neural network for local gradient enhancement to generate a high-frequency temperature anomaly feature map; multi-channel spatial pooling is performed on the texture feature of the visible light image to extract vegetation coverage density and surface crack distribution features to generate a texture enhancement feature matrix; the high-frequency temperature anomaly feature map and the texture enhancement feature matrix are channel-concatenated to generate a multi-scale fusion feature vector; The thermal radiation pattern library in the pre-trained fire point identification knowledge is loaded through the transfer learning algorithm, and the typical fire point radiation profile in the thermal radiation pattern library is matched with the multi-scale fusion feature vector for similarity to generate a fire point feature matching coefficient set; the fire point feature matching coefficient set and the multi-scale fusion feature vector are superimposed on the residual block of the residual neural network to perform cross-layer feature fusion to generate a temporally and spatially associated latent fire point enhanced feature tensor; A bidirectional long short-term memory network is used to perform time series modeling on the hidden fire point enhanced feature tensor, capture the temporal correlation between temperature anomalies and texture changes, and generate a dynamic hidden fire point evolution sequence; Performing spatial attention weighting on the dynamic hidden fire point evolution sequence to generate a fire point existence probability value for each geographic unit; performing Gaussian kernel density estimation on the target area according to the fire point existence probability value to generate a hidden fire point location confidence map with a smooth transition; Inputting the hidden fire point location confidence map into the output layer of the residual neural network for normalization processing to generate a normalized probability distribution matrix; performing peak screening on the overlapping probability regions in the normalized probability distribution matrix by a non-maximum suppression algorithm to generate a de-redundant hidden fire point location confidence map; The step of performing spatial intersection and union operations on the hidden fire point location confidence map and the combustible material distribution layer in a preset geographic information system to determine a target hidden fire point coordinate set includes: Performing rasterization processing on the hidden fire point location confidence map to generate a first raster data layer, wherein each raster unit in the first raster data layer contains a corresponding hidden fire point existence probability value; Performing equal-resolution raster conversion on the combustible distribution layer to generate a second raster data layer, wherein each raster cell in the second raster data layer contains a combustible density level value; Using the set hidden fire point existence probability threshold and combustible material density level threshold, the first raster data layer and the second raster data layer are respectively subjected to binary filtering to generate a frequent probability hidden fire point area mask and a frequent combustible material density area mask; Perform spatial overlay analysis on the frequent probability hidden fire point area mask and the frequent combustible material density area mask to extract a set of overlapping grid cells that simultaneously meet the hidden fire point existence probability threshold and the combustible material density level threshold; perform morphological closing operation on the overlapping grid cell set to eliminate discretely distributed isolated grid cells and generate a continuous space area cluster; perform a connected domain labeling algorithm on the continuous space area cluster to identify the circumscribed rectangular bounding box of each independent connected domain, and calculate the weighted centroid coordinates of the hidden fire point existence probability of the grid cells in each circumscribed rectangular bounding box; The weighted centroid coordinates of the existence probability of the hidden fire point are mapped to the geographic coordinate system of the preset geographic information system to generate a candidate hidden fire point coordinate set; based on the combustible distribution gradient direction of each coordinate in the candidate hidden fire point coordinate set, the direction consistency check is performed on adjacent coordinates, and the abnormal coordinate points whose gradient direction does not match the diffusion trend of the hidden fire point are eliminated to generate the target hidden fire point coordinate set; The target hidden fire point coordinate set is spatially matched with the historical hidden fire point verification database, and the newly added hidden fire point coordinates that do not exist in the historical hidden fire point verification database are screened out and updated to the target hidden fire point coordinate set.

[0023] The automated processing flow of hidden fire point detection in the embodiment of the present invention achieves accurate positioning of hidden danger points through multi-stage feature fusion and spatial analysis. Gaussian kernel density estimation generates a smooth confidence map, and the non-maximum suppression algorithm is combined to effectively eliminate false alarm interference. Spatial intersection and superposition analysis of combustible distribution ensures that the target coordinate set meets the dual standards of temperature anomaly and combustible conditions at the same time, greatly improving the spatial directivity of hidden fire point identification and the efficiency of implementing preventive measures.

[0024] An embodiment of the present invention provides a fire monitoring and early warning device, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0025] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flow chart of a fire monitoring and early warning method based on big data analysis provided in an embodiment of the present invention.

[0027] Figure 2 A schematic diagram of the structure of a fire monitoring and early warning device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments in the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present invention.

[0029] The terms "first", "second", etc. in the embodiments of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the embodiments of the present invention can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the embodiments of the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0030] Figure 1 A fire monitoring and early warning method based on big data analysis is shown, which is applied to a fire monitoring and early warning device. The method includes the following steps 110 to 140.

[0031] Step 110: Collect multi-source fire monitoring data, which includes satellite remote sensing heat source images, meteorological environment parameters, historical fire event records and geographic information raster data.

[0032] In the embodiment of the present invention, the fire monitoring and early warning device implements all-weather three-dimensional monitoring data acquisition for the selected target area by deploying a multi-source heterogeneous data acquisition system.

[0033] For example, satellite remote sensing heat source image data comes from infrared sensor equipment carried by geosynchronous orbit meteorological satellites. The equipment obtains surface thermal radiation distribution images of the target area at fixed time intervals. Its spatial resolution meets the needs of medium-scale heat source identification and can effectively capture areas of abnormal surface temperature.

[0034] For example, meteorological environmental parameters are transmitted in real time through a multi-factor meteorological observation station network deployed in the target area, covering core parameters such as atmospheric temperature and humidity, wind speed and direction, and precipitation intensity. Wind speed monitoring uses high-precision three-dimensional dynamic measurement technology to ensure accurate capture of airflow movement characteristics.

[0035] For another example, historical fire event records are extracted from the regional disaster management database, which includes hundreds of fire event files from the past decade. Each record contains multi-dimensional attribute data such as the geographic information of the fire point, disaster duration, and impact range.

[0036] In addition, the geographic information raster data is based on the high-precision digital terrain model released by the national geographic information platform, integrating vegetation cover type distribution maps, infrastructure spatial location data and artificial structure vector layers to construct a multi-level geographic information base framework.

[0037] Step 120: Performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporally correlated fire dynamic data set, wherein the spatiotemporal fusion processing includes timestamp alignment, geocoding mapping, and abnormal data removal.

[0038] In the embodiment of the present invention, the fire monitoring and early warning device starts the spatiotemporal fusion processing core module to perform spatiotemporal consistency optimization processing on the multi-source monitoring data collected in the target area.

[0039] Specifically, the timestamp alignment unit adopts a unified time reference conversion technology to convert heterogeneous time series data such as satellite image acquisition time and meteorological parameter sampling time into a standard time coordinate system, and compensates for the collection time difference of different data sources through a time series interpolation algorithm. For example, for periodically collected meteorological element data, dynamic interpolation calculations are performed at the satellite transit time node.

[0040] Furthermore, the geocoding mapping component applies multiple spatial resolution matching technology to unify the spatial benchmarks of geographic information raster data and satellite heat source images of different resolutions, uses an adaptive spatial aggregation algorithm to achieve seamless conversion of high-resolution geographic data to satellite image scale, and constructs a spatial weight association model between meteorological station observation data and satellite pixels.

[0041] Next, the abnormal data detection unit effectively identifies and eliminates abnormal heat source signals caused by transient equipment failures or environmental interference by designing temperature gradient threshold rules and spatial distribution continuity verification mechanisms. For example, when the temperature value of a pixel is significantly different from that of the surrounding adjacent areas and does not conform to the laws of heat conduction, the data repair program is automatically triggered, and ultimately a dynamic fire data set with complete spatiotemporal dimensions, consistent spatial benchmarks and controllable data quality is output.

[0042] Step 130: calling the trained target fire spread prediction model to process the fire dynamic data set, and outputting the fire risk index and spread trend probability distribution of the target area.

[0043] In this step, the fire risk index is the core quantitative parameter output by the target fire spread prediction model, which is used to characterize the probability level of fire occurrence in different spatial units in the target area at a specific time point. The specific time point is not limited, such as the unified time point after the multi-source data completes the spatiotemporal alignment (such as the satellite transit time 14:00), or the fire evolution timeline node (such as the next 15 minutes, 1 hour, 3 hours). The fire risk index provides a spatially fine-grained risk assessment basis for emergency decision-making. For example, high-risk areas may correspond to areas with dense dry vegetation and strong winds, helping to prioritize the deployment of prevention and control resources.

[0044] For example, the spread trend probability distribution is a prediction result generated based on the calculation of the multidimensional diffusion model, which is used to describe the intensity and possibility of the spread of the fire in different spatial directions in the target area in the next few hours. The spread trend probability distribution generates a probabilistic spatial heat map by analyzing the meteorological evolution laws (such as wind direction changes), terrain dynamic factors (such as valley airflow effects) and combustible distribution characteristics in the fire dynamics data set, combined with the combustion diffusion pattern extracted by the spatiotemporal convolutional neural network. The output form of the spread trend probability distribution can be a multidimensional matrix or a raster layer, and each pixel value represents the probability or intensity level of the fire spreading in that direction. This distribution can assist in predicting the path of the fire head advancement and provide data support for evacuation route planning and barrier zone setting.

[0045] In an embodiment of the present invention, a fire monitoring and early warning device calls a target fire spread prediction model trained with massive historical data, and the target fire spread prediction model adopts a hybrid neural network architecture under a deep learning framework.

[0046] In detail, the input data processing layer constructs the multidimensional data after spatiotemporal fusion into a feature fusion cube: the first dimension integrates the normalized satellite heat source intensity distribution data; the second dimension integrates the wind field dynamic distribution matrix generated by spatial interpolation; the third dimension integrates terrain elevation derivative parameters and slope characteristic data; the fourth dimension associates the distribution of historical fire hotspots with the flammability assessment results of vegetation. The target fire spread prediction model extracts spatial feature patterns through a multi-layer convolutional neural network structure, combines the time series analysis module to capture the evolution law of meteorological elements, and finally generates a dual-channel prediction result at the output layer: the fire risk index uses a standardized numerical range to represent the instantaneous fire occurrence probability level of each spatial unit in the target area, and the spread trend probability distribution calculates the fire spread intensity distribution in multiple directions in the next few hours through a multi-dimensional diffusion model.

[0047] In specific implementation, the target fire spread prediction model may include model architectures such as spatiotemporal convolution layers, three-dimensional convolution kernels, and graph convolution network node processing modules. The multi-dimensional fire feature matrix of the fire dynamic data set can be extracted through the spatiotemporal convolution layer, and the model input tensor can be generated by splicing with the fire feature vector set, and the three-dimensional fire spread field is generated by traversing the model input tensor according to the three-dimensional convolution kernel. Through the graph convolution network node processing module, based on the three-dimensional fire spread field, the fire transmission intensity between each geographic grid unit and the adjacent unit is calculated, and the fire transmission probability matrix and fire spread prediction path between nodes are generated. Based on the fire transmission probability matrix and fire spread prediction path, the fire risk index and the probability distribution of the spread trend are generated.

[0048] For example, the fire dynamics dataset can be input into the spatiotemporal convolution layer of the target fire spread prediction model to extract the spatial distribution pattern of fire points and the trend characteristics of meteorological parameters in the fire dynamics dataset to generate a multidimensional fire feature matrix. The multidimensional fire feature matrix and the fire feature vector set are then spliced ​​in the channel dimension to generate a model input tensor containing real-time fire point spread rate, surface vegetation flammability index, wind speed and direction correlation factor, and terrain slope flame retardant coefficient. Afterwards, the spatial dimension of the model input tensor is traversed through the three-dimensional convolution kernel of the target fire spread prediction model to generate a three-dimensional fire spread field containing a spatial diffusion velocity field, a combustible consumption rate field, and a thermal radiation attenuation field. Based on the spatial diffusion velocity field of the three-dimensional fire diffusion field and the geographic coding mapping relationship of the fire dynamics dataset, a dynamic adjacency matrix of each geographic grid unit in the target fire spread prediction model is constructed. Then, the graph convolutional network node processing module of the target fire spread prediction model is used to combine the dynamic adjacency matrix and the combustible consumption rate field of the three-dimensional fire spread field to calculate the fire transmission probability matrix between each geographic grid unit, and based on the fire transmission probability matrix and the thermal radiation attenuation field of the three-dimensional fire spread field, the set of fire transmission paths that exceed the preset confidence threshold is screened to generate the optimized fire spread prediction path. After that, based on the spatial coverage density of the fire spread prediction path and the transmission intensity of the fire transmission probability matrix, the fire risk accumulation value of each grid unit in the target area is calculated to generate the fire risk index. Among them, let the fire spread prediction path be P, and the number of spatial coverage of path P in each grid unit be N. i (i represents different grid units), the transmission intensity corresponding to each grid unit in the fire transmission probability matrix is ​​T i First, calculate the spatial coverage density D of each grid cell i , D i =N i / Total number of grid cells. Then, for each grid cell, the cumulative fire risk value R i The calculation method is: R i =D i ×T i Finally, the fire risk accumulation value R of each grid cell is i Standardization is performed (for example, the maximum risk accumulation value is R_max and the minimum is R min , then the standardized fire risk index F i =(R i -R min ) / (R_max-R min )), through the above calculation, the fire risk index of each grid cell in the target area can be generated.

[0049] Then, Monte Carlo simulation sampling is performed on the predicted path of fire spread. The spatial diffusion velocity field and thermal radiation attenuation field of the three-dimensional fire spread field are combined to statistically analyze the probability distribution characteristics of the fire coverage in different time windows to generate the probability distribution of the spread trend. Finally, the fire risk index and the probability distribution of the spread trend are spatially superimposed and analyzed. The invalid data in the geographical obstacle area are eliminated by combining the combustible consumption rate field of the three-dimensional fire spread field, and the corrected fire risk index and the probability distribution of the spread trend are output. In the actual application process, the terrain dynamic influence factor can be optimized according to the typical landform characteristics of the target area, thereby significantly improving the prediction accuracy of the special airflow in the valley area on the spread of fire. The terrain dynamic influence factor is a parameter that quantifies the intensity of the terrain effect on the spread of fire, mainly reflecting the acceleration / blocking effect of landform characteristics such as slope, slope aspect, and elevation change rate on the fire. In the optimization process, the slope (Slope) and terrain uplift index (TRI) of the target area can be extracted based on the digital elevation model (DEM), and the basic factor value can be calculated through the physical model (for example, the formula is: terrain dynamic factor = 0.6×tan(Slope)+0.4×TRI 0.5 ); machine learning is used to optimize weight coefficients for typical landforms (such as valleys and ridges) - for example, the random forest algorithm is used to analyze the correlation between historical fire spread paths and terrain parameters. When it is detected that the fire spread speed along the valley is 1.8 times that of the ridge, the slope weight of the valley area is increased from 0.5 to 0.7; finally, a dynamic terrain factor matrix adapted to the regional landform is generated with a spatial resolution of 30 meters, which is used for real-time terrain correction of the fire spread prediction model.

[0050] Step 140: Generate a multi-level warning signal based on the fire risk index and the spread trend probability distribution, and dynamically update it to the regional emergency response platform.

[0051] In this step, the fire monitoring and early warning device further implements a multi-level early warning signal generation and dynamic update mechanism. For example, the early warning level classification module establishes a composite judgment rule based on the threshold interval of the fire risk index and the spatial distribution characteristics of the spread trend. When the target sub-area meets the risk level threshold and the spread direction concentration conditions at the same time, the yellow, orange, and red level early warning signals are triggered in sequence.

[0052] For another example, the warning information generation module encodes the judgment results into a standard emergency data format and transmits them in real time to the core dispatching system of the regional emergency response platform through a dedicated communication protocol. The platform's three-dimensional visualization interface immediately renders the warning area as a semi-transparent thermal layer, and simultaneously associates the spatial distribution information of surrounding emergency rescue resources.

[0053] Next, the dynamic update component continuously receives incremental analysis results regularly pushed by the fire monitoring and early warning device. When it detects that the risk index of a specific sub-region shows a significant upward trend during the continuous monitoring period, it automatically triggers the early warning level upgrade protocol and sends enhanced positioning early warning instructions to patrol personnel in the target area through the emergency broadcast system. The data feedback loop also transmits on-site observation information back to the fire monitoring and early warning device, forming a closed-loop optimization mechanism to continuously improve the spatiotemporal prediction accuracy of the early warning system.

[0054] In an embodiment of the present invention, the multi-level warning signal is a dynamic emergency response instruction generated by the fire monitoring and early warning device, which is triggered in a hierarchical manner according to the threshold of the fire risk index and the spatial characteristics of the spread trend. Typical grading rules include: yellow warning (local risk exceeds the limit but the spread trend is dispersed), orange warning (multiple sub-regional risks are superimposed and the spread direction is concentrated), and red warning (high-risk areas are contiguous and the fire spreads explosively). The warning signal is transmitted to the emergency platform in real time through a standardized coding protocol, driving the thermal rendering of the three-dimensional visualization system, the priority adjustment of the resource scheduling module, and the directional warning broadcast of the broadcast system. Its dynamic update mechanism can automatically upgrade or downgrade the warning level based on the incremental data analysis results, and form a closed-loop optimization with on-site feedback to ensure the spatiotemporal synchronization of response measures and fire evolution.

[0055] In an optional embodiment, the step 120 of performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporal associated fire dynamic data set includes: Step 121: extract the fire point pixel coordinates and the radiation intensity value in the satellite remote sensing heat source image, and map the fire point pixel coordinates and the radiation intensity value to the geographic coordinate system of the target area based on a geographic coordinate conversion algorithm.

[0056] In an embodiment of the present invention, the fire monitoring and early warning device performs an extraction operation of the fire point pixel coordinates and the radiation intensity value on the received satellite remote sensing heat source image. In the specific implementation process, the infrared radiation intensity of each pixel in the satellite image is separated from the background surface temperature through a threshold segmentation algorithm, and the abnormal fire point pixel set exceeding the preset temperature threshold is identified.

[0057] Subsequently, the fire monitoring and early warning device calls the geographic coordinate conversion algorithm to convert the row and column coordinates of the fire point pixels into the geographic coordinate system uniformly adopted by the target area. For example, for the target area using UTM projection, the fire monitoring and early warning device converts the center coordinates of each fire point pixel into the easting coordinates and northing coordinates in the UTM coordinate system through the affine transformation matrix based on the projection parameters in the satellite image metadata, and at the same time associates the radiation intensity value with the corresponding coordinate point to form a fire point feature data set with geographic spatial reference.

[0058] Step 122: performing segmented interpolation on the meteorological environment parameters according to the timestamps to generate a continuous meteorological parameter surface matching the geographic coordinate system.

[0059] In an embodiment of the present invention, the fire monitoring and early warning device performs time dimension interpolation and spatial surface reconstruction on discrete meteorological parameters uploaded by the meteorological observation station network. Specifically, the fire monitoring and early warning device segments the temperature, humidity, wind speed and direction parameters recorded by each station according to minute-level timestamps, and uses a cubic spline interpolation algorithm to generate a continuous time series for missing data of the same meteorological element between adjacent timestamps.

[0060] Next, based on the spatial grid division rules of the target area's geographic coordinate system, the fire monitoring and early warning device uses the Kriging spatial interpolation method to convert discrete site data into a continuous meteorological parameter surface covering the entire area. For example, for wind speed parameters, the fire monitoring and early warning device calculates the horizontal and vertical components of the wind speed vector in the grid unit based on the three-dimensional dynamic measurement data of each site, generates a wind field distribution matrix consistent with the spatial resolution of the geographic coordinate system, and realizes the continuous expression of meteorological parameters in the spatial dimension.

[0061] Step 123: Perform spatial cluster analysis on the historical fire event records to generate a probability density map of fire-frequently occurring areas, and overlay the probability density map onto the geographic coordinate system.

[0062] In step 123, the fire monitoring and early warning device performs spatial clustering analysis and probability density calculation on the historical fire event records. The fire monitoring and early warning device first converts the geographical coordinates of the historical fire points into the unified geographical coordinate system of the target area, and uses a density-based spatial clustering algorithm to identify the spatial distribution pattern of the fire-prone areas.

[0063] For example, the DBSCAN clustering algorithm was used to scan hundreds of fire coordinates in the past ten years, identifying three fire hotspots with high vegetation coverage and complex terrain. Subsequently, the fire monitoring and early warning device constructed a fire probability density map based on the kernel density estimation algorithm, converted the historical fire frequency of each geographic grid unit into a probability density value, and generated a probability distribution grid layer aligned with the geographic coordinate system space, which intuitively reflects the statistical laws of fire occurrence in different regions.

[0064] Step 124: Dynamically weighted fusion of the fire point pixel coordinates, the continuous meteorological parameter surface and the probability density map is performed through a spatiotemporal sliding window algorithm to generate the fire dynamic data set.

[0065] In step 124, the fire monitoring and early warning device uses a spatiotemporal sliding window algorithm to achieve dynamic fusion of multi-source data. The fire monitoring and early warning device defines a sliding window with a time window length of 30 minutes and a spatial window length of 1 kilometer, and aggregates the fire point coordinates, meteorological parameter surfaces, and fire probability density maps in each spatiotemporal window.

[0066] For example, for the geographic grid covered by the current window, the fire monitoring and early warning device calculates the mean heat source intensity based on the radiation intensity of the fire point, extracts the dominant wind direction based on the wind speed surface data in the window, and superimposes the historical fire probability density value of the grid. Through the dynamic weight allocation mechanism, the weight of the fire point data is set to 0.5, the weight of the meteorological parameter is 0.3, and the weight of the historical probability is 0.2. After weighted fusion, the dynamic fire feature vector of the window is generated. The fire monitoring and early warning device advances the window in time series, and finally constructs a spatiotemporal correlation fire dynamic data set covering the entire domain.

[0067] In an alternative embodiment, before calling the trained target fire spread prediction model to process the fire dynamics dataset in step 130, the method further includes: Step 210: Generate a fire condition feature vector set, which includes a real-time fire point spread rate, a surface vegetation flammability index, a wind speed and direction correlation factor, and a terrain slope flame retardancy coefficient.

[0068] In step 210, the real-time fire spread rate is obtained by calculating the spatial displacement of the fire coordinates in adjacent time windows, for example, the spread rate is calculated based on the centroid movement distance and time interval of the same fire cluster in two satellite images. The surface vegetation flammability index is based on the distribution of vegetation types in the geographic information grid data, combined with real-time humidity data, and assigns high index values ​​to flammable vegetation such as coniferous forests and shrubs. Among them, the acquisition of the surface vegetation flammability index can be obtained based on the fusion of multi-source data. For example, the distribution of vegetation types in the target area is first inverted through satellite multispectral remote sensing data (such as the NDVI index of Landsat-8 to distinguish coniferous forests, broad-leaved forests, shrubs and other land types), and the basic flammability benchmark value is assigned to each type of vegetation in combination with the ground survey database (such as coniferous forest 0.8, shrubs 0.7, grassland 0.6, scale 0-1); then, the real-time relative humidity data of the meteorological observation station is connected, and the benchmark value is dynamically corrected using a piecewise linear function (such as flammability increases by 20% when humidity is less than 30%, and decreases by 35% when humidity is greater than 70%); finally, a rasterized flammability index map is generated through a spatial interpolation algorithm. In addition, the dead branches and fallen leaves coverage data obtained by drone inspections (LiDAR point cloud density inversion) can be superimposed for secondary calibration. The wind speed and direction correlation factor quantifies the effect of wind on fire by calculating the cosine value of the angle between the dominant wind direction and the terrain slope. The terrain slope flame retardant coefficient is extracted based on the digital elevation model, and a low flame retardant coefficient is assigned to areas with a slope greater than 15 degrees, which represents the inhibitory effect of steep slope terrain on the spread of fire.

[0069] Step 220: Use a convolutional neural network to extract features from the fire dynamics data set to generate a multi-dimensional fire feature matrix.

[0070] In step 220, the fire monitoring and early warning device converts the fire dynamic data set into a three-dimensional grid structure, in which the channel dimension includes parameters such as heat source intensity, wind speed, and humidity. For example, the tensor size of the input convolutional neural network is 512×512×6, corresponding to the number of rows, columns, and six feature channels of the geographic grid. The size of the first-layer convolution kernel is set to 3×3, and the spatial correlation features of the local area are extracted by sliding scanning, such as identifying the superposition pattern of high temperature areas and high wind speed areas. The deep network compresses the spatial dimension through pooling operations, and finally outputs a multi-dimensional fire feature matrix of 256×256×32, in which each feature map encodes the spatial distribution law of fire at different scales.

[0071] Step 230: Perform tensor concatenation on the multi-dimensional fire condition feature matrix and the fire condition feature vector set to generate model input data.

[0072] In step 230, the fire monitoring and early warning device performs a tensor concatenation operation to integrate multidimensional features. Specifically, the multidimensional fire feature matrix generated by the convolutional neural network is dimensionally aligned with the set of fire feature vectors created in step 210. For example, the multidimensional fire feature matrix is ​​converted into a 16384-dimensional vector through a flattening operation, and is concatenated with the fire feature vector containing four scalar features such as diffusion rate and flammability index into a 16388-dimensional input vector. The vector is then reconstructed into a four-dimensional tensor structure to adapt to the input layer requirements of the target fire spread prediction model, ensuring the coordinated processing of spatiotemporal features and physical features in the model.

[0073] In one example, the target fire spread prediction model is trained through a generative adversarial network, using historical fire data as real samples to generate simulated fire spread paths, and the model parameters are optimized through a discriminant network. Based on this, the training process of the target fire spread prediction model may include: Step 310: Obtain fire spread trajectory data of a historical time series and a snapshot of environmental parameters at a time point corresponding to the fire spread trajectory data.

[0074] In an embodiment of the present invention, the fire monitoring and early warning device extracts 20 major fire cases in the past five years from the disaster management database, and each case contains fire boundary vector data at intervals of minutes. For example, the diffusion trajectory of a forest fire is represented by 15 continuous fire line polygons formed within 12 hours. The fire monitoring and early warning device simultaneously extracts satellite heat source images, meteorological observation data and terrain data when the fire occurs, constructs a time series of environmental parameter snapshots, and forms a multi-dimensional correlation data set between the fire diffusion process and the external environment.

[0075] Step 320: Create a three-dimensional fire spread field, which includes a spatial diffusion velocity field, a combustible consumption rate field, and a thermal radiation attenuation field.

[0076] In the embodiment of the present invention, the spatial diffusion velocity field is calculated by the rate of change of the fire boundary, and the hourly fire line advance distance is converted into a velocity scalar value of a 500-meter grid unit. The combustible consumption rate field quantifies the burning rate of different vegetation types based on the difference in remote sensing data of vegetation coverage before and after the fire. The thermal radiation attenuation field uses infrared satellite data to invert the temperature gradient at the edge of the fire and establish a physical model of the attenuation of thermal radiation intensity with distance. The three diffusion fields together constitute a three-dimensional data structure that describes the spatiotemporal evolution of the fire.

[0077] Step 330: Use a graph convolutional network to model the spatiotemporal relationship of the three-dimensional fire diffusion field and generate a fire transmission probability matrix between nodes.

[0078] In an embodiment of the present invention, the fire monitoring and early warning device uses a graph convolutional network to model the fire transmission relationship, abstracts the geographic grid units of the target area into graph structure nodes, and initializes the connection weights between nodes to the inverse of the spatial distance. The graph convolutional network first learns the spatial propagation law of node features (such as combustible type, slope value), and then simulates the fire transmission process between nodes through time stepping. For example, the network recognizes that when the wind speed of the upstream node is greater than 5m / s and the slope difference with the downstream node is less than 10 degrees, the probability of fire transmission increases to above 0.8, accurately reflecting the coupled effect of terrain and wind field on fire spread.

[0079] Step 340: Optimizing the weight parameters of the fire spread probability matrix through a reinforcement learning algorithm so as to maximize the similarity between the predicted path output by the target fire spread prediction model and the actual fire spread trajectory.

[0080] In the embodiment of the present invention, the fire monitoring and early warning device defines the reward function as the Jaccard similarity coefficient between the predicted fire boundary and the actual trajectory, and adjusts the weight parameters of the graph convolutional network through the policy gradient algorithm. For example, when the northeastern diffusion path of the fire predicted by the model deviates from the actual fire expansion direction by more than 15 degrees, the algorithm automatically reduces the connection weight of the adjacent nodes in this direction, and after multiple rounds of iterations, the spatial overlap rate of the predicted path and the historical data is increased to more than 92%.

[0081] Step 350: Solidify the optimized weight parameters into the inference module of the target fire spread prediction model.

[0082] In the embodiment of the present invention, the fire monitoring and early warning device imports the weight matrix obtained by reinforcement learning training into the inference engine of the target fire spread prediction model, and saves the feature normalization parameters and graph structure configuration information. For example, the feature scaling factor of the terrain slope flame retardant coefficient is set to 0.15 to ensure the consistency of the input data and training data distribution during field deployment. The solidified model has the ability to process fire dynamic data in real time and output prediction results, supporting the business needs of second-level response time.

[0083] In a preferred embodiment, the step 140 of generating a multi-level warning signal according to the fire risk index and the spread trend probability distribution includes: Step 141: Divide the warning level according to the numerical range of the fire risk index; wherein, the first-level warning corresponds to a risk threshold in the first proportional interval of the historical peak, the second-level warning corresponds to a risk threshold in the second proportional interval of the historical peak, and the third-level warning corresponds to a risk threshold in the third proportional interval of the historical peak.

[0084] In an embodiment of the present invention, the fire monitoring and early warning device performs an early warning level classification operation according to the numerical range of the fire risk index. Specifically, the fire monitoring and early warning device calls the peak value data of the risk index in the past five years stored in the historical fire database, and determines the early warning level threshold by percentile statistics.

[0085] For example, the first-level warning corresponds to the interval where the risk index value exceeds 80% of the historical peak, the second-level warning corresponds to the interval between 50%-80% of the historical peak, and the third-level warning corresponds to the interval between 30%-50% of the historical peak. In the eastern forest area of ​​the target area, the fire monitoring and early warning device detected that the current fire risk index reached 0.78, exceeding the 80% threshold of the historical peak of 0.72 in the area, and automatically triggered the first-level warning signal. The warning level division mechanism has a built-in regional differentiation correction module to adjust the threshold interval according to the differences in vegetation types. For example, the first-level warning threshold in the coniferous forest area is lowered to 70% of the historical peak due to the dense characteristics of combustibles.

[0086] Step 142: Calculate the time evolution curve of the fire wave and range based on the probability distribution of the spreading trend, and determine the effective time window of each warning level according to the time evolution curve.

[0087] In an embodiment of the present invention, the fire monitoring and early warning device calculates the time evolution curve of the fire wave based on the spread trend probability distribution. In detail, the fire monitoring and early warning device extracts the time series of the fire propagation intensity in six directions from the spread trend probability distribution grid data, and applies the integral algorithm to generate a curve cluster of the front position of the fire wave changing with time.

[0088] For example, the fire wave in the southwest of the target area advances at a speed of 1.2 kilometers per hour in the first two hours, and then accelerates to 1.8 kilometers per hour due to the terrain uplift, generating an exponential rising curve. The fire monitoring and early warning device determines the effective time window of each warning level through curve inflection point detection. When the slope of the fire wave curve in the northeast direction exceeds 0.5 kilometers per hour, it is determined that the third-level warning window in this direction is shortened to 45 minutes, triggering the adjustment of the time-sensitive warning parameters.

[0089] Step 143: Combine and encode the warning level and the effective time window to generate a warning instruction set including the geographic fence coordinates.

[0090] In an embodiment of the present invention, the fire monitoring and early warning device performs a combined encoding operation of the early warning instruction set. The fire monitoring and early warning device performs a matrix combination of the early warning level divided in step 141 and the effective time window calculated in step 142 to generate an instruction data packet containing three-dimensional geographic fence coordinates. For example, the first-level early warning area is encoded as a polygonal space object with an east coordinate range of [452300, 452900], a north coordinate range of [3321500, 3322100], and an elevation range of [800, 1200] meters in the UTM coordinate system, and the timestamp information of the effective time window 14:30-16:00 is superimposed. The encoding process adopts the GeoJSON standard format, and the encryption check code is embedded to ensure the integrity of the instruction transmission. Each instruction unit contains a warning level code, a geographic fence vertex coordinate set, and a time validity identifier.

[0091] Step 144: Distribute the warning instruction set to the terminal devices within the corresponding geographic fence through the edge computing node, and activate the emergency response protocol.

[0092] In the embodiment of the present invention, the fire monitoring and early warning device implements the distribution and protocol activation of the early warning instruction set through the edge computing node. In detail, the fire monitoring and early warning device divides the instruction set generated in step 143 into transmission domains according to the geographic fence range, and pushes the corresponding data to the 32 edge computing nodes deployed in the target area through the 5G private network.

[0093] For example, the edge node numbered E15 receives three first-level warning fence instructions within its coverage area and immediately starts the protocol activation process: sending a red warning signal to 18 intelligent roadside units in the area, waking up the drone base station within a 5-kilometer range to perform aerial monitoring tasks, and sending positioning warning information to the patrol personnel's handheld terminal through the LoRaWAN protocol. The edge node simultaneously activates the power management module in the emergency response protocol and increases the pressure of the surrounding intelligent fire hydrants to the preset emergency response value.

[0094] In the next step, the dynamic update in step 140 to the regional emergency response platform includes: Step 145: monitor the value change rate of the fire risk indicator in real time, and trigger a re-evaluation of the warning level when the value change rate exceeds a preset critical value.

[0095] In step 145, the fire monitoring and early warning device monitors the numerical change rate of the fire risk index in real time. The numerical change rate refers to the increase or decrease of the fire risk index per unit time, which is achieved by calculating the risk value difference of adjacent monitoring cycles. For example, during specific monitoring, the risk value sequence is extracted with a sliding time window, and the linear regression algorithm is used to fit the change trend line, and its slope is the current rate (unit: risk value / minute); for example, when the absolute value of the slope is detected to be ≥0.02 / min (that is, the risk value changes ≥0.1 every 5 minutes), the rate overlimit alarm is triggered. In detail, the fire monitoring and early warning device establishes a minute-level data sampling channel to calculate the differential change rate of the risk index in adjacent time windows. For example, in the hilly area in the north of the target area, the fire monitoring and early warning device detects that the fire risk index jumps from 0.54 to 0.68 within 10 minutes, and the change rate reaches 0.014 / minute, which exceeds the preset critical value of 0.01 / minute. At this time, the fire monitoring and early warning device triggers the warning level re-evaluation instruction and starts the dynamic update process. The monitoring module has a built-in abnormal fluctuation filtering algorithm. When the wind speed sensor is instantly disturbed and causes a brief jump in the risk value, the re-evaluation operation is automatically delayed until the data becomes stable again.

[0096] Step 146: Adjust the geo-fence radius and evacuation route planning of the warning instruction set according to the re-evaluation result.

[0097] In step 146, the fire monitoring and early warning device adjusts the parameters of the early warning instruction set according to the re-evaluation results. The fire monitoring and early warning device calls the terrain flame retardant coefficient model to recalculate the fire spread resistance and expands the original geographic fence radius from 3 kilometers to 4.2 kilometers. For example, in the southeast of the target area, because the wind direction was detected to change from northwest to southeast and the wind speed increased by 2.4m / s, the evacuation route planning module automatically adjusted the evacuation route originally distributed along the contour lines to a fast channel perpendicular to the dominant wind direction. The adjustment process uses the A* optimization algorithm to avoid the real-time fire wave prediction path to ensure that the newly planned evacuation route No. 17 maintains a safe distance of more than 2.8 kilometers from the fire front.

[0098] Step 147: Match the adjusted warning instruction set with the emergency resource database to generate an emergency plan including a material dispatch plan and a rescue force deployment diagram.

[0099] In step 147, the fire monitoring and early warning device performs an emergency plan generation operation. The fire monitoring and early warning device topologically associates and matches the adjusted early warning instruction set with the emergency resource spatial database, wherein the material dispatching plan obtains the optimal solution by solving the vehicle path problem with a time window. For example, for the expanded geographic fence area, the fire monitoring and early warning device retrieves the location information of two fire stations, five sprinkler trucks, and three material storage depots within a radius of 15 kilometers, and generates a multi-objective optimization dispatching plan: fire trucks A1 and A3 arrive at the isolation zone on the west side of the fire scene along road G358 first, and sprinkler trucks B2 and B5 perform vegetation humidification operations through the emergency passage in the forest area. The rescue force deployment map uses a Voronoi diagram to divide the responsibility area to ensure that each sub-area has at least two rescue units that can implement cross-support.

[0100] Step 148: Use the digital twin system to simulate the implementation effects of different emergency plans, and select the emergency plan with the smallest quantitative loss as the target response strategy.

[0101] In step 148, the fire monitoring and early warning device simulates and selects the best emergency plan through the digital twin system. Specifically, the fire monitoring and early warning device constructs a three-dimensional digital twin model of the target area with millimeter-level accuracy, and imports real-time meteorological data, combustible distribution layers, and building structure parameters.

[0102] For example, the two alternative plans generated in step 147 can be simulated: Plan A adopts a frontal interception strategy, and the digital twin simulation shows that the fire breaks through the second isolation belt in 2 hours and 43 minutes; Plan B adopts a flanking tactic, and the simulation results show that the fire area can be controlled to a safe range in 3 hours and 15 minutes. The fire monitoring and early warning device calculates that Plan B can reduce the burned area by 12.7% and the exposure risk of rescue personnel by 23.4% through the loss assessment algorithm. Finally, Plan B is selected as the target response strategy, and the decision-based data packet is synchronized to the emergency command center verification system.

[0103] In an alternative embodiment, the method further comprises: Step 410: embedding an attention mechanism into the target fire spread prediction model to dynamically adjust the contribution weights of different environmental parameters to fire spread prediction.

[0104] In step 410, the fire monitoring and early warning device embeds an attention mechanism in the target fire spread prediction model to optimize the feature weight distribution. In specific implementation, the fire monitoring and early warning device adds a multi-head attention layer in the hybrid neural network architecture of the model, which dynamically adjusts the contribution weights of different environmental parameters to fire prediction by calculating the correlation matrix between satellite heat source intensity distribution data, wind speed dynamic matrix and terrain slope characteristics.

[0105] For example, when a sustained wind speed increase occurs in the southeast of the target area, the attention mechanism automatically increases the weight coefficient of the wind speed and direction correlation factor from the baseline value of 0.3 to 0.52, while reducing the weight of the humidity parameter in the adjacent area, so that the model prediction focuses more on the wind-driven fire spread effect. The attention weight matrix is ​​updated every five minutes to ensure that the model adapts to real-time environmental parameter changes.

[0106] Step 420: extracting meteorological mutation characteristics at the current moment, wherein the meteorological mutation characteristics include the probability of dry thunderstorm occurrence and the air humidity drop rate.

[0107] In step 420, the fire monitoring and early warning device extracts the meteorological mutation characteristics at the current moment. The fire monitoring and early warning device calculates two core indicators, the probability of dry thunderstorms and the rate of sudden drop in air humidity, by analyzing the minute-level data stream uploaded by the meteorological observation station network.

[0108] For example, in the northwest of the target area, the fire monitoring and early warning device detected that the relative humidity dropped from 65% to 28% within three hours, and the humidity drop rate reached 0.2% / minute. At the same time, the atmospheric electric field intensity monitoring value exceeded 4kV / m, triggering a dry thunderstorm probability warning. The fire monitoring and early warning device called the lightning location data within the sliding time window to verify that three cloud-to-ground discharge events had occurred in the area, confirming that the meteorological mutation characteristics reached the preset sensitive threshold.

[0109] Step 430: When it is detected that the meteorological mutation feature exceeds the preset sensitivity threshold, the activation strength of the corresponding feature channel in the target fire spread prediction model is enhanced: according to the causal relationship diagram between the meteorological parameter mutation event and the historical fire case, the key trigger factor combination is identified; the sensitivity coefficient of each feature channel to the model output result is calculated by the gradient back propagation algorithm; an exponential weight gain is applied to the feature channel whose sensitivity coefficient exceeds the sensitivity threshold; during the model reasoning process of the target fire spread prediction model, the output fluctuation of the gain channel is monitored in real time, and the model self-correction mechanism is initiated when the fluctuation amplitude exceeds the safety range.

[0110] In step 430, the fire monitoring and early warning device starts the model feature channel enhancement mechanism. When the humidity drop rate is detected to exceed 0.15% / minute, the fire monitoring and early warning device retrieves 37 fire events under similar meteorological conditions in the historical fire case library, constructs a causal relationship diagram, and identifies the combination of vegetation dryness index and wind speed mutation as the key trigger factor. The sensitivity coefficient of each feature channel in the target fire spread prediction model to the output fire risk index is calculated by the gradient back propagation algorithm, and it is found that the sensitivity coefficient of the surface combustible moisture content channel reaches 0.78, exceeding the threshold value of 0.6. The fire monitoring and early warning device applies an exponential weight gain to the channel, increasing the gain factor from 1.0 to 2.3.

[0111] During the model inference process, the output fluctuation of the gain channel is monitored in real time. When the predicted fire spread rate fluctuates by more than ±25% within ten minutes, the model self-correction mechanism is triggered and the gain factor is adjusted to 1.8 through the online learning algorithm to maintain the prediction stability.

[0112] Step 440: Based on the enhanced target fire spread prediction model output, the calculation indication of the fire risk index is modified to generate an emergency warning signal under mutation conditions.

[0113] In step 440, the fire monitoring and early warning device generates an emergency early warning signal under sudden change conditions. Based on the enhanced target fire spread prediction model, the fire monitoring and early warning device recalculates the fire risk index and upgrades the southeastern area, which was originally assessed as a second-level early warning, to a special-level early warning.

[0114] For example, the risk index value of the area was revised from 0.62 to 0.85, triggering a new purple warning signal. The fire monitoring and early warning device generates a warning instruction containing a mutation condition identifier, and sends enhanced warning information containing lightning protection instructions and rapid evacuation routes to patrol personnel in the area through the emergency broadcast system. The three-dimensional visualization interface of the regional emergency response platform is updated simultaneously, and the impact range of the meteorological mutation is marked with pulse flashing special effects.

[0115] In an alternative embodiment, the method further comprises: Step 510: Deploy a swarm of drones to conduct real-time thermal imaging scans of areas with frequent fire risks to obtain a surface temperature distribution map with sub-meter accuracy.

[0116] In the embodiment of the present invention, the fire monitoring and early warning device deploys a drone swarm to perform thermal imaging scanning tasks. The fire monitoring and early warning device dispatches 6 rotary-wing drones equipped with dual-spectrum thermal imagers to conduct grid inspections of the three designated fire risk-prone areas. The drone swarm flies along a preset route and collects surface temperature data at an altitude of 150 meters with a spatial resolution of 0.3 meters.

[0117] For example, in the coniferous forest belt in the northeast of the target area, the sub-meter-level temperature distribution map obtained by the drone showed 12 local temperature anomalies, with the highest temperature reaching 52°C, forming an 8°C temperature difference with the surrounding environment. The drone swarm transmits temperature data in real time to the fire monitoring and early warning device through the 5G private network, with a data collection frequency of 2 frames per second.

[0118] Step 520: Perform multi-scale feature matching on the surface temperature distribution map and the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points.

[0119] In the embodiment of the present invention, the fire monitoring and early warning device implements multi-scale feature matching to identify hidden fire points. The fire monitoring and early warning device spatially registers the sub-meter surface temperature distribution map obtained by the drone with the 500-meter resolution heat source image of the Fengyun-4 satellite, and uses a pyramid matching algorithm to compare heat source features at multiple scales.

[0120] For example, in an area that appears as a single hot spot in satellite images, the drone data analyzed three discrete high-temperature points, two of which were confirmed to be surface fires and one to be a hidden fire point with no visible flames through feature matching. The fire monitoring and early warning device established a hidden fire point feature library, recording the periodic fluctuation characteristics of its temperature curve, which matched the typical smoldering fire pattern by 87%.

[0121] Step 530: Simulate the diffusion path of the potential hidden fire point coordinates through the target fire spread prediction model to generate a hot spot map of the early intervention area.

[0122] In an embodiment of the present invention, the fire monitoring and early warning device generates a hot spot map of the early intervention area. The target fire spread prediction model simulates the diffusion path of the identified potential hidden fire point coordinates, comprehensively considers the current wind speed of 3.2m / s, slope of 15° and other terrain and meteorological conditions, and predicts that the hidden fire point may ignite the surrounding 300 meters of bushes within 6 hours. The hot spot map generated by the fire monitoring and early warning device divides the risk area into a core disposal area (radius of 50 meters), a buffer monitoring area (radius of 150 meters) and an outer warning area (radius of 300 meters), and each partition is marked with recommended disposal measures and resource demand levels.

[0123] Step 540: dispatch firefighting resources to carry out preventive isolation zone construction according to the heat map, and update the preventive isolation zone construction strategy to the regional emergency response platform.

[0124] In the embodiment of the present invention, the fire monitoring and early warning device performs preventive isolation zone construction scheduling. According to the core disposal area coordinates marked in the hot spot map, the fire monitoring and early warning device retrieves the road layer and fire resource location in the geographic information grid data to generate the optimal construction path.

[0125] For example, two bulldozers were dispatched to build a 30-meter-wide isolation belt along the contour line, and three sprinkler trucks were called to carry out infiltration operations in the buffer monitoring area. The construction plan was sent to the engineering vehicle terminal through the digital work order system to monitor the construction progress in real time, and the completed isolation belt vector data was updated to the basic layer database of the regional emergency response platform.

[0126] In a preferred embodiment, the step 520 of performing multi-scale feature matching on the surface temperature distribution map and the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points includes: Step 521: Perform wavelet transform analysis on the surface temperature distribution map to extract frequency domain characteristics of abnormal temperature fluctuations.

[0127] In step 521, the fire monitoring and early warning device performs wavelet transform analysis on the surface temperature distribution map, and can use the Daubechies wavelet basis function to perform a five-layer decomposition of the temperature matrix to extract the energy distribution characteristics of each frequency band. For example, in the target hidden fire point area, the high-frequency component energy accounts for more than 65%, which is consistent with the frequency domain characteristic pattern of the smoldering fire. The fire monitoring and early warning device constructs a temperature fluctuation spectrum histogram and marks three suspected hidden fire areas with continuous high-frequency oscillation characteristics to provide feature input for subsequent detection.

[0128] Step 522: Create a hidden fire point detection model based on a residual neural network, and input the abnormal temperature fluctuation frequency domain features and the acquired visible light image texture features into the hidden fire point detection model.

[0129] In step 522, the fire monitoring and early warning device creates a hidden fire point detection model based on a residual neural network. The model input layer receives the temperature frequency domain features processed by wavelet transform and the texture features of the visible light image of the drone, where the visible light image extracts texture parameters in 8 directions through the grayscale co-occurrence matrix. The residual module is designed as a deep structure containing 15 convolutional layers, and each residual block introduces a jump connection to prevent the gradient from disappearing. For example, for an area with no obvious smoke on the surface but with high temperature and high frequency fluctuations, the model identifies the accumulation of dead branches through texture features, and together with the temperature features, supports the determination of hidden fire points.

[0130] Step 523: Migrate the pre-trained fire point identification knowledge to the hidden fire point detection model through a transfer learning algorithm.

[0131] In step 523, the fire monitoring and early warning device implements transfer learning to optimize the detection model. The parameters of the open fire recognition model that has been trained in other forest areas are migrated to the hidden fire point detection model, the weight parameters of the underlying feature extraction layer are retained, and the top classifier is retrained to adapt to the hidden fire detection task. Exemplarily, the parameters of the first 10 convolutional layers can be frozen during the migration process, and only the last 5 residual blocks are fine-tuned, so as to improve the accuracy of the model in the hidden fire recognition of specific vegetation types in the target area.

[0132] Step 524: Use the hidden fire point detection model to process the temperature anomaly fluctuation frequency domain characteristics, the visible light image texture characteristics and the fire point identification knowledge, and output a probabilistic hidden fire point location confidence map.

[0133] In step 524, the fire monitoring and early warning device generates a confidence map of the hidden fire point location. The hidden fire point detection model outputs a 256×256 probability matrix, and each pixel value represents the confidence that a hidden fire exists at that location. The fire monitoring and early warning device uses a non-maximum suppression algorithm to eliminate redundant detections in adjacent high-confidence areas, such as merging three high-confidence points with a spacing of less than 10 meters into a single hidden fire point. The confidence map finally generated is aligned with the geographic coordinate system, and after superimposing the terrain elevation data, a three-dimensional probability distribution model is formed, marking 5 high-risk hidden fire points with a confidence level exceeding 0.9.

[0134] Step 525: Perform spatial intersection and union operations on the hidden fire point location confidence map and the combustible material distribution layer in the preset geographic information system to determine the target hidden fire point coordinate set.

[0135] In step 525, the fire monitoring and early warning device determines the target hidden fire point coordinate set. The fire monitoring and early warning device imports the hidden fire point location confidence map into the geographic information system and performs spatial intersection operations with the vegetation flammability classification layer. For example, a point with a disposal confidence of 0.85 completely overlaps with the distribution area of ​​highly flammable coniferous forests and is confirmed as a valid hidden fire point; while another point with a confidence of 0.78 is located in a rock exposed area and is determined to be a false alarm through spatial intersection analysis and is eliminated. The fire monitoring and early warning device ultimately outputs a list of hidden fire points containing geographic coordinates, confidence values, and surrounding combustible material types as the core basis for early intervention decisions.

[0136] In a preferred embodiment, the step 524 uses the hidden fire point detection model to process the temperature anomaly fluctuation frequency domain features, the visible light image texture features, and the fire point identification knowledge to output a probabilistic hidden fire point location confidence map, including: Step 5241: Input the temperature anomaly fluctuation frequency domain features into the first convolutional layer of the residual neural network for local gradient enhancement to generate a high-frequency temperature anomaly feature map; perform multi-channel spatial pooling processing on the visible light image texture features to extract vegetation coverage density and surface crack distribution characteristics to generate a texture enhancement feature matrix; perform channel cascading on the high-frequency temperature anomaly feature map and the texture enhancement feature matrix to generate a multi-scale fusion feature vector.

[0137] In step 5241, the fire monitoring and early warning device performs multi-source feature fusion processing, inputs the temperature abnormality fluctuation frequency domain features into the first convolution layer of the residual neural network, uses a 3×3 convolution kernel for local gradient enhancement, and highlights the high-frequency temperature abnormality area.

[0138] For example, an abnormal increase in energy in the 52Hz frequency band was detected in the northwest of the target area, and after convolution operation, a patchy high-frequency temperature anomaly feature map with obvious edge features was formed. At the same time, the fire monitoring and early warning device performs multi-channel spatial pooling processing on the visible light image of the drone, compresses the image resolution through the maximum pooling layer, extracts the texture features of the area with a vegetation coverage density of 85% and the surface crack length of more than 2 meters, and generates a texture enhancement feature matrix.

[0139] Finally, the fire monitoring and early warning device cascades the high-frequency temperature anomaly feature map and the texture enhancement feature matrix to form a multi-scale fusion feature vector containing 128 feature channels, providing a cross-modal data basis for subsequent analysis.

[0140] Step 5242: Load the thermal radiation pattern library in the pre-trained fire point identification knowledge through the transfer learning algorithm, perform similarity matching on the typical fire point radiation profile in the thermal radiation pattern library and the multi-scale fusion feature vector, and generate a set of fire point feature matching coefficients; superimpose the fire point feature matching coefficient set and the multi-scale fusion feature vector into the residual block of the residual neural network, perform cross-layer feature fusion, and generate a temporally and spatially correlated latent fire point enhanced feature tensor.

[0141] In step 5242, the fire monitoring and early warning device implements cross-layer feature fusion optimization and loads the thermal radiation pattern library in the pre-trained fire point recognition knowledge, which contains 237 typical fire point radiation profile data. Through the similarity matching algorithm, the multi-scale fusion feature vector of the target area is compared with the radiation characteristics of the surface smoldering fire in the pattern library to generate a matching coefficient matrix.

[0142] For example, the cosine similarity between a certain feature vector and the radiation profile of No. 153 reaches 0.89, triggering a high-confidence match mark. The fire monitoring and early warning device superimposes the matching coefficient matrix and the multi-scale fusion feature vector into the fifth residual block of the residual neural network, and uses a jump connection mechanism to fuse shallow texture features and deep semantic features to generate a temporally and spatially associated hidden fire point enhanced feature tensor. This tensor retains the spatial resolution while encoding the time-dependent characteristics of the evolution of hidden fire points.

[0143] Step 5243: Use a bidirectional long short-term memory network to perform time series modeling on the hidden fire point enhanced feature tensor, capture the temporal correlation between temperature anomalies and texture changes, and generate a dynamic hidden fire point evolution sequence.

[0144] In step 5243, the fire monitoring and early warning device constructs a time series evolution model, uses a bidirectional long short-term memory network to model the time series of the hidden fire point enhanced feature tensor, and analyzes the evolution law of five consecutive frames of temperature anomaly data. For example, it is detected that the temperature fluctuation frequency of a certain coordinate point gradually increases from 45Hz to 62Hz within 30 minutes, which is consistent with the typical frequency shift characteristics of the spread of smoldering fire. The network captures the spatial correlation between this time series change and the expansion of the carbonization traces of dead branches in the visible light image, and generates a dynamic hidden fire point evolution sequence containing timestamp marks, which accurately reflects the stage characteristics of the hidden fire point from germination to stable combustion.

[0145] Step 5244: Perform spatial attention weighting on the dynamic hidden fire point evolution sequence to generate a fire point existence probability value for each geographic unit; perform Gaussian kernel density estimation on the target area based on the fire point existence probability value to generate a hidden fire point location confidence map with a smooth transition.

[0146] In step 5244, the fire monitoring and early warning device generates a spatial probability distribution map, applies a spatial attention mechanism to the dynamic hidden fire point evolution sequence, and adjusts the attention weight of each geographical unit according to the terrain slope and wind speed vector. For example, the northeast slope area is affected by strong winds, and its attention weight is increased to 1.3 times the baseline value. The weighted feature data is input into the Gaussian kernel density estimation algorithm to generate a smoothly transitioned hidden fire point location confidence map. The map shows three high-probability core areas in the eastern woodland of the target area, with confidence values ​​of 0.92, 0.87 and 0.79, respectively, which accurately correspond to the abnormal temperature rise areas discovered by drone thermal imaging.

[0147] Step 5245: Input the hidden fire point location confidence map into the output layer of the residual neural network for normalization processing to generate a normalized probability distribution matrix; perform peak screening on the overlapping probability areas in the normalized probability distribution matrix through a non-maximum suppression algorithm to generate a de-redundant hidden fire point location confidence map.

[0148] In step 5245, the fire monitoring and early warning device optimizes the output quality of the confidence map, inputs the hidden fire point location confidence map into the residual neural network output layer, and uses the Softmax function to perform global probability normalization to eliminate the dimensional differences between different regions. For example, the core area with an original confidence of 0.92 maintains the highest value of 0.98 after normalization, while the surrounding transition area drops below 0.45.

[0149] Subsequently, the fire monitoring and early warning device applied the non-maximum suppression algorithm to perform peak screening on the normalized probability distribution matrix, and eliminated redundant detection points with a 3×3 neighborhood window scan. In the southwest of the target area, three high-confidence points with a spacing of less than 5 meters were merged into a single hidden fire point, and finally a redundancy-free hidden fire point location confidence map was generated, which contained 17 independent high-probability areas.

[0150] In a preferred embodiment, the step 525 of performing spatial intersection and union operation on the hidden fire point location confidence map and the combustible material distribution layer in a preset geographic information system to determine the target hidden fire point coordinate set includes: Step 5251: rasterize the hidden fire point location confidence map to generate a first raster data layer, wherein each raster unit in the first raster data layer contains a corresponding hidden fire point existence probability value.

[0151] In step 5251, the fire monitoring and early warning device performs raster data conversion, converting the hidden fire point location confidence map into a 1-meter resolution raster data layer in the UTM coordinate system, and each raster unit stores a probability value in the range of 0-1. For example, the raster unit numbered G327541 records a confidence value of 0.87, which corresponds to the coniferous forest area with the actual geographic coordinates of 118.72 degrees east longitude and 32.15 degrees north latitude. The rasterization process uses a bilinear interpolation algorithm to ensure spatial continuity and ensure that the probability distribution of steep slope terrain is not distorted.

[0152] Step 5252: Perform equal-resolution raster conversion on the combustible material distribution layer to generate a second raster data layer, wherein each grid cell in the second raster data layer contains a combustible material density level value.

[0153] In step 5252, the fire monitoring and early warning device unifies the data space benchmark, performs raster conversion on the combustible distribution vector layer in the preset geographic information system, and generates a second raster data layer with the same resolution as the hidden fire point confidence map. During the conversion process, the combustible density is divided into five levels of values ​​according to the vegetation type comparison table, for example, the coniferous forest area is assigned a value of 4 (highly flammable) and the broad-leaved forest area is assigned a value of 3. The raster conversion uses the nearest neighbor algorithm to maintain the original classification accuracy, ensuring that each 1-meter grid unit accurately reflects the state of the surface combustibles.

[0154] Step 5253: Using the set hidden fire point existence probability threshold and combustible material density level threshold, perform binary filtering on the first raster data layer and the second raster data layer respectively to generate frequent probability hidden fire point area masks and frequent combustible material density area masks.

[0155] In step 5253, the fire monitoring and early warning device implements spatial threshold filtering, sets the probability threshold of hidden fire points to 0.7, the combustible density level threshold to level 3, and performs binarization on the two grid layers. For example, in the eastern part of the target area, 12 of the 17 high-probability hidden fire point areas pass the probability threshold filtration, and the combustible distribution layer screens out 8 areas that meet the density requirements. After logical AND operations, a frequent probability hidden fire point area mask is generated, which contains 6 grid clusters that meet the dual requirements at the same time, forming a potential hidden fire core area to be analyzed.

[0156] Step 5254: Perform spatial overlay analysis on the frequent probability hidden fire point area mask and the frequent combustible material density area mask to extract a set of overlapping grid cells that simultaneously meet the hidden fire point existence probability threshold and the combustible material density level threshold; perform morphological closing operation on the overlapping grid cell set to eliminate discretely distributed isolated grid cells and generate a continuous space area cluster; execute a connected domain labeling algorithm on the continuous space area cluster to identify the circumscribed rectangular bounding box of each independent connected domain, and calculate the weighted centroid coordinates of the hidden fire point existence probability of the grid cells within each circumscribed rectangular bounding box.

[0157] In step 5254, the fire monitoring and early warning device performs spatial morphological optimization, performs morphological closing operations on the overlapping grid unit sets, and uses 5×5 structural elements to eliminate discretely distributed isolated grids. For example, three grids with a spacing of 2 meters in a certain area are connected into a continuous area through closing operations. Subsequently, the eight-neighborhood connected domain labeling algorithm is applied to identify two independent connected domains, and their circumscribed rectangular bounding boxes are calculated respectively. In the connected domain numbered C15, the fire monitoring and early warning device calculates the weighted centroid of the probability of existence of hidden fire points in each grid unit, and finally determines the position of the core hidden fire point with coordinates of 118.7236 degrees east longitude and 32.1532 degrees north latitude.

[0158] Step 5255: Map the weighted centroid coordinates of the hidden fire point existence probability to the geographic coordinate system of the preset geographic information system to generate a set of candidate hidden fire point coordinates; based on the combustible material distribution gradient direction of each coordinate in the candidate hidden fire point coordinate set, perform a direction consistency check on adjacent coordinates, eliminate abnormal coordinate points whose gradient direction does not match the hidden fire point diffusion trend, and generate the target hidden fire point coordinate set.

[0159] In step 5255, the fire monitoring and early warning device verifies the spread trend of hidden fire points, imports the candidate hidden fire point coordinate set into the terrain dynamic model, and analyzes the combustible distribution gradient direction of each point. For example, the combustible density gradient of candidate point number P09 points to the southeast, and the angle with the fire spread direction predicted by the current wind speed and direction is less than 15 degrees, passing the direction consistency check. However, candidate point number P12 was judged as an abnormal coordinate and eliminated because the gradient direction deviated by 45 degrees from the predicted path. The verified target hidden fire point coordinate set contains 9 valid points, and the spatial distribution conforms to the law of fire spread.

[0160] Step 5256: spatially match the target hidden fire point coordinate set with the historical hidden fire point verification database, filter out the newly added hidden fire point coordinates that do not exist in the historical hidden fire point verification database, and update them to the target hidden fire point coordinate set.

[0161] In step 5256, the fire monitoring and early warning device completes the update of hidden fire point data, spatially matches the target hidden fire point coordinate set with the historical verification database of the past three years, and finds that 3 of the points overlap with the records of the smoldering fire points that have been dealt with, and are automatically marked as recurring potential risk points. The remaining 6 newly added coordinates are manually verified and confirmed, updated to the target hidden fire point coordinate set, and synchronously entered into the historical database. The updated set contains 15 valid coordinates, 9 of which are located in key monitoring areas, triggering preventive disposal plans.

[0162] As an optional but non-limiting embodiment, the step 124 performs dynamic weighted fusion of the fire point pixel coordinates, the continuous meteorological parameter surface and the probability density map through a spatiotemporal sliding window algorithm to generate the fire dynamic data set, including: Step 1241: Set the window size parameters and sliding step parameters of the spatiotemporal sliding window based on the preset spatiotemporal resolution, wherein the window size parameters include the time window length and the spatial grid coverage, and the sliding step parameters include the time advancement interval and the spatial offset.

[0163] In step 1241, the fire monitoring and early warning device sets the core parameters of the spatiotemporal sliding window algorithm. According to the terrain complexity of the target area and the speed of fire evolution, the time window length is set to 30 minutes, and the spatial grid coverage is configured as a 500m×500m unit. The time advancement interval in the sliding step parameter is set to 15 minutes, and the spatial offset is set to 50 meters to ensure that there is a 75% spatial overlap rate between adjacent windows. For example, in the hilly area in the eastern part of the target area, the fire monitoring and early warning device uses the above parameters to achieve fine capture of the fire evolution process based on the characteristics of the average historical fire spread speed of 1.2 kilometers per hour in the area, while avoiding the fuzziness of local features caused by excessive window size.

[0164] Step 1242: According to the window size parameter and the sliding step parameter, the fire point radiation intensity value of the fire point pixel coordinates, the meteorological parameter value of the continuous meteorological parameter surface and the fire occurrence probability value of the probability density map are cut into a set of local data blocks aligned in time and space.

[0165] In step 1242, the fire monitoring and early warning device performs a multi-source data cutting and alignment operation, maps the pixel coordinates of the fire point monitored by the satellite to a 500-meter spatial grid, and extracts the maximum radiation intensity of the fire point in each grid unit. The continuous meteorological parameter surface is gridded and resampled to obtain the wind speed, temperature and humidity parameter values ​​at the center of each grid. At the same time, the historical fire probability density map is downsampled to the same resolution, and the historical occurrence probability value of each grid is extracted. For example, in the grid unit numbered G3275, the fire monitoring and early warning device recorded a radiation intensity peak of 235W / m², a wind speed of 4.2m / s, and a historical fire probability of 0.67, forming a local data block aligned in time and space.

[0166] Step 1243: Perform range normalization processing on the fire point radiation intensity values, meteorological parameter values ​​and fire occurrence probability values ​​in the local data block set respectively to generate normalized fire point radiation intensity values, normalized meteorological parameter values ​​and normalized fire occurrence probability values.

[0167] In step 1243, the fire monitoring and early warning device implements multi-parameter normalization processing, normalizes the radiation intensity value of the fire point using the range method, and linearly maps the 80-320W / m² interval observed in the target area to the 0-1 range. The wind speed value in the meteorological parameter is normalized according to the 0-10m / s benchmark, and the temperature and humidity parameters are processed using the absolute temperature scale and percentage scale respectively. The historical fire probability value maintains the original 0-1 value unchanged. For example, the original radiation intensity of a grid unit of 280W / m² is converted to 0.83 after normalization, the wind speed of 5.6m / s is converted to 0.56, and the historical probability of 0.72 is directly retained to eliminate the influence of different dimensions on the fusion calculation.

[0168] Step 1244: Dynamically allocate the fire point weight coefficient of the normalized fire point radiation intensity value, the meteorological weight coefficient of the normalized meteorological parameter value, and the probability weight coefficient of the normalized fire occurrence probability value according to the real-time data quality assessment index; wherein the real-time data quality assessment index includes the timeliness of satellite image update, the error rate of meteorological sensors, and the spatial coverage of historical data.

[0169] In step 1244, the fire monitoring and early warning device dynamically calculates the feature weight coefficient. When it is detected that the current satellite image update delay exceeds 8 minutes, the fire point weight coefficient is reduced from the baseline value of 0.6 to 0.4. The error rate report transmitted in real time by the meteorological sensor network shows that the humidity parameter error is less than 2%, and the corresponding meteorological weight coefficient is increased to 0.7. The historical data spatial coverage assessment shows that data is missing in the northwest of the target area, and the probability weight coefficient of the area is reduced by 0.2. For example, during the time window of 14:30-15:00, the fire point weight of the eastern grid unit is 0.5, the meteorological weight is 0.7, and the probability weight is 0.3, forming a dynamic balance to ensure the reliability of fusion when data quality fluctuates.

[0170] Step 1245: Perform a weighted product calculation on the normalized fire point radiation intensity value and the fire point weight coefficient to generate a weighted fire point intensity value; perform a weighted product calculation on the normalized meteorological parameter value and the meteorological weight coefficient to generate a weighted meteorological parameter value; perform a weighted product calculation on the normalized fire occurrence probability value and the probability weight coefficient to generate a weighted fire probability value.

[0171] In step 1245, the fire monitoring and early warning device completes the weighted feature calculation, multiplies the normalized fire point radiation intensity value 0.83 by the fire point weight coefficient 0.5, and obtains a weighted fire point intensity value of 0.415. The normalized wind speed value 0.56 is combined with the meteorological weight 0.7 to generate a weighted wind speed value of 0.392. The historical probability value 0.72 is multiplied by the probability weight 0.3 to obtain a weighted probability value of 0.216. The three weighted values ​​form a feature vector [0.415, 0.392, 0.216] in grid unit G3275, which characterizes the comprehensive fire status of the unit in a specific time and space window.

[0172] Step 1246: Perform pixel-by-pixel superposition operations on the weighted fire point intensity values, the weighted meteorological parameter values, and the weighted fire probability values ​​within the same space-time window to generate a preliminary fused fire indicator matrix; perform a space-time continuity check on the preliminary fused fire indicator matrix, including detecting the indicator mutation rate of adjacent windows and the consistency of the spatial gradient direction, and eliminating abnormal fused pixel points that do not meet the continuity constraints.

[0173] In step 1246, the fire monitoring and early warning device performs fusion and verification, superimposes the weighted values ​​of three consecutive spatiotemporal windows in the eastern region pixel by pixel, and generates a preliminary fused fire indicator matrix. The spatial continuity check found that the fusion value of the grid number G3280 suddenly increased to 0.92, which was 3.8 times different from the average value of 0.48 of the adjacent grids, triggering the anomaly detection mechanism. The gradient direction analysis showed that the mutation point was inversely sheared with the dominant wind direction, and it was determined to be caused by instantaneous interference in the satellite data. After performing data repair, the fusion value of the point was corrected to 0.51 to maintain the rationality of the spatial distribution.

[0174] Step 1247: The verified preliminary fused fire indicator matrix is ​​spatially spliced ​​and time series connected according to the sliding step parameter to generate a complete spatiotemporal fire dynamic data set.

[0175] In step 1247, the fire monitoring and early warning device constructs a complete spatiotemporal data set, splices the verified local data blocks at a spatial offset of 50 meters, and eliminates the boundary effect of overlapping areas. In the time dimension, continuous windows are connected at intervals of 15 minutes to form a continuous data set of 4 frames per hour. For example, the target area generates a 240×180 grid time series data cube in the period of 14:00-15:00, and each grid contains the fusion index values ​​of 48 time nodes, which fully describes the spatiotemporal evolution of the fire.

[0176] Step 1248: attach a spatiotemporal index tag to each data unit in the spatiotemporally complete fire dynamics dataset, wherein the spatiotemporal index tag includes a timestamp code and a geographic grid code.

[0177] In step 1248, the fire monitoring and early warning device adds a spatiotemporal index tag and uses the ISO8601 standard to format the timestamp code, such as "2023-08-15T14:30:00Z" to identify the specific window period. The geographic grid code uses the UTM partition number combined with the grid row and column number, for example, "50S_3275_5412" represents the 500-meter grid at 118.7 degrees east longitude and 32.1 degrees north latitude. Each data unit is quickly retrieved through the spatiotemporal index, supporting the regional emergency response platform to extract fire dynamic data of specific spatiotemporal nodes according to geographic coordinates or time ranges.

[0178] Further, Figure 2 FIG. 2 is a schematic diagram of a fire monitoring and early warning device 200 provided in an embodiment of the present invention. Figure 2 The fire monitoring and early warning device 200 shown includes a processor 210, and the processor 210 can call and run a computer program from a memory to implement the method in the embodiment of the present invention.

[0179] Alternatively, if Figure 2 As shown, the fire monitoring and early warning device 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention.

[0180] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0181] Alternatively, if Figure 2 As shown, the fire monitoring and early warning device 200 may also include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices.

[0182] Optionally, the fire monitoring and early warning device 200 can implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present invention. For the sake of brevity, they will not be repeated here.

[0183] It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capability.

[0184] It is understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0185] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0186] It should be noted that, in this article, the term "comprises", "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0187] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned implementation methods can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the embodiment of the present invention.

[0188] The embodiments of the embodiments of the present invention are described above in conjunction with the accompanying drawings, but the embodiments of the present invention are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present invention and the scope of protection of the embodiments of the present invention, all of which are within the protection of the embodiments of the present invention.

Claims

1. A fire monitoring and early warning method based on big data analysis, characterized in that: include: Collecting multi-source fire monitoring data, the multi-source fire monitoring data including satellite remote sensing heat source images, meteorological environmental parameters, historical fire event records and geographic information raster data; Performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporally associated fire dynamic data set, wherein the spatiotemporal fusion processing includes timestamp alignment, geocoding mapping, and abnormal data removal; The trained target fire spread prediction model is called to process the fire dynamic data set, and the fire risk index and the spread trend probability distribution of the target area are output; wherein the target fire spread prediction model includes a spatiotemporal convolution layer, a three-dimensional convolution kernel and a graph convolution network node processing module, and the multi-dimensional fire feature matrix of the fire dynamic data set is extracted through the spatiotemporal convolution layer, and is spliced ​​with the fire feature vector set to generate a model input tensor, and the model input tensor is traversed according to the three-dimensional convolution kernel to generate a three-dimensional fire spread field, and the graph convolution network node processing module is used to calculate the fire transmission intensity between each geographic grid unit and the adjacent units based on the three-dimensional fire spread field, and the fire transmission probability matrix and the fire spread prediction path between the nodes are generated, and the fire risk index and the spread trend probability distribution are generated based on the fire transmission probability matrix and the fire spread prediction path; A multi-level warning signal is generated according to the fire risk index and the probability distribution of the spreading trend, and is dynamically updated to the regional emergency response platform.

2. The method according to claim 1, characterized in that The step of performing spatiotemporal fusion processing on the multi-source fire monitoring data to generate a spatiotemporally correlated fire dynamic data set includes: Extracting the fire point pixel coordinates and the radiation intensity value in the satellite remote sensing heat source image, and mapping the fire point pixel coordinates and the radiation intensity value to the geographic coordinate system of the target area based on a geographic coordinate conversion algorithm; The meteorological environment parameters are interpolated piecewise according to the timestamps to generate a continuous meteorological parameter surface matching the geographic coordinate system; Performing spatial cluster analysis on the historical fire event records to generate a probability density map of fire-frequently occurring areas, and superimposing the probability density map on the geographic coordinate system; The fire point pixel coordinates, the continuous meteorological parameter surface and the probability density map are dynamically weighted fused through a spatiotemporal sliding window algorithm to generate the fire situation dynamic data set.

3. The method according to claim 2, characterized in that Before calling the trained target fire spread prediction model to process the fire dynamic data set, the method further includes: Generate a fire characteristic vector set, wherein the fire characteristic vector set includes a real-time fire point spread rate, a surface vegetation flammability index, a wind speed and wind direction correlation factor, and a terrain slope flame retardancy coefficient; A convolutional neural network is used to extract features from the fire dynamic data set to generate a multi-dimensional fire feature matrix; Performing tensor splicing on the multidimensional fire condition feature matrix and the fire condition feature vector set to generate model input data; Among them, the target fire spread prediction model is trained through a generative adversarial network, using historical fire data as real samples to generate simulated fire spread paths, and the model parameters are optimized through a discriminant network.

4. The method according to claim 3, characterized in that The training process of the target fire spread prediction model includes: Obtaining fire spread trajectory data of a historical time series and a snapshot of environmental parameters at a time point corresponding to the fire spread trajectory data; Creating a three-dimensional fire diffusion field, the three-dimensional fire diffusion field including a spatial diffusion velocity field, a combustible consumption rate field and a thermal radiation attenuation field; A graph convolutional network is used to model the spatiotemporal relationship of the three-dimensional fire diffusion field, and a fire transmission probability matrix between nodes is generated; Optimizing the weight parameters of the fire transmission probability matrix by a reinforcement learning algorithm so as to maximize the similarity between the predicted path output by the target fire spread prediction model and the actual fire spread trajectory; The optimized weight parameters are solidified into the reasoning module of the target fire spread prediction model.

5. The method according to claim 1, characterized in that The generating of a multi-level warning signal according to the fire risk index and the probability distribution of the spreading trend includes: The warning level is divided according to the numerical range of the fire risk index; wherein the first-level warning corresponds to a risk threshold in the first proportional range of the historical peak value, the second-level warning corresponds to a risk threshold in the second proportional range of the historical peak value, and the third-level warning corresponds to a risk threshold in the third proportional range of the historical peak value; Calculating the time evolution curve of the fire wave and range based on the probability distribution of the spreading trend, and determining the effective time window of each warning level according to the time evolution curve; Combining and encoding the warning level and the effective time window to generate a warning instruction set including the coordinates of the geo-fence; The warning instruction set is distributed to the terminal devices within the corresponding geographic fence through the edge computing node, and the emergency response protocol is activated.

6. The method according to claim 5, characterized in that The dynamic update to the regional emergency response platform includes: Real-time monitoring of the rate of change of the fire risk indicator, and triggering a re-evaluation of the warning level when the rate of change of the fire risk indicator exceeds a preset critical value; Adjusting the geo-fence radius and evacuation route planning of the warning instruction set according to the re-evaluation result; Match the adjusted warning instruction set with the emergency resource database to generate an emergency plan including a material dispatch plan and a rescue force deployment diagram; The digital twin system is used to simulate the implementation effects of different emergency plans, and the emergency plan with the smallest quantitative loss is selected as the target response strategy.

7. The method according to claim 1, characterized in that The method further comprises: An attention mechanism is embedded in the target fire spread prediction model to dynamically adjust the contribution weights of different environmental parameters to fire spread prediction; Extracting meteorological mutation characteristics at the current moment, wherein the meteorological mutation characteristics include the probability of dry thunderstorm occurrence and the air humidity drop rate; When it is detected that the meteorological mutation characteristic exceeds a preset sensitivity threshold, the activation intensity of the corresponding characteristic channel in the target fire spread prediction model is enhanced; Based on the enhanced target fire spread prediction model output, the calculation indication of the fire risk index is modified to generate an emergency warning signal under mutation conditions; The step of enhancing the activation intensity of the corresponding feature channel in the target fire spread prediction model includes: Based on the causal relationship diagram between meteorological parameter mutation events and historical fire cases, identify the key trigger factor combination; The sensitivity coefficient of each feature channel to the model output result is calculated through the gradient back propagation algorithm; Applying an exponential weight gain to the feature channels whose sensitivity coefficients exceed a sensitivity threshold; During the model reasoning process of the target fire spread prediction model, the output fluctuation of the gain channel is monitored in real time, and the model self-correction mechanism is started when the fluctuation amplitude exceeds the safety range.

8. The method according to claim 1, characterized in that The method further comprises: Deploy drone swarms to conduct real-time thermal imaging scans of areas prone to fire risk, and obtain surface temperature distribution maps with sub-meter accuracy; Performing multi-scale feature matching on the surface temperature distribution map and the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points; The target fire spread prediction model is used to simulate the diffusion path of the potential hidden fire point coordinates to generate a hot spot map of the early intervention area; Dispatching firefighting resources to carry out preventive isolation zone construction according to the hot spot map, and updating the preventive isolation zone construction strategy to the regional emergency response platform; The multi-scale feature matching of the surface temperature distribution map with the satellite remote sensing heat source image to identify the coordinates of potential hidden fire points includes: Performing wavelet transform analysis on the surface temperature distribution map to extract frequency domain characteristics of abnormal temperature fluctuations; Creating a hidden fire point detection model based on a residual neural network, and inputting the abnormal temperature fluctuation frequency domain features and the acquired visible light image texture features into the hidden fire point detection model; Migrating the pre-trained fire point recognition knowledge to the hidden fire point detection model through a transfer learning algorithm; The hidden fire point detection model is used to process the temperature anomaly fluctuation frequency domain characteristics, the visible light image texture characteristics and the fire point identification knowledge, and output a probabilistic hidden fire point location confidence map; The hidden fire point location confidence map is spatially intersected and merged with the combustible material distribution layer in a preset geographic information system to determine a target hidden fire point coordinate set.

9. The method according to claim 8, characterized in that The method of using the hidden fire point detection model to process the temperature abnormal fluctuation frequency domain characteristics, the visible light image texture characteristics and the fire point identification knowledge to output a probabilistic hidden fire point location confidence map includes: The temperature anomaly fluctuation frequency domain feature is input into the first convolutional layer of the residual neural network for local gradient enhancement to generate a high-frequency temperature anomaly feature map; multi-channel spatial pooling is performed on the texture feature of the visible light image to extract vegetation coverage density and surface crack distribution features to generate a texture enhancement feature matrix; the high-frequency temperature anomaly feature map and the texture enhancement feature matrix are channel-concatenated to generate a multi-scale fusion feature vector; The thermal radiation pattern library in the pre-trained fire point identification knowledge is loaded through the transfer learning algorithm, and the typical fire point radiation profile in the thermal radiation pattern library is matched with the multi-scale fusion feature vector for similarity to generate a fire point feature matching coefficient set; the fire point feature matching coefficient set and the multi-scale fusion feature vector are superimposed on the residual block of the residual neural network to perform cross-layer feature fusion to generate a temporally and spatially associated latent fire point enhanced feature tensor; A bidirectional long short-term memory network is used to perform time series modeling on the hidden fire point enhanced feature tensor, capture the temporal correlation between temperature anomalies and texture changes, and generate a dynamic hidden fire point evolution sequence; Performing spatial attention weighting on the dynamic hidden fire point evolution sequence to generate a fire point existence probability value for each geographic unit; performing Gaussian kernel density estimation on the target area according to the fire point existence probability value to generate a hidden fire point location confidence map with a smooth transition; Inputting the hidden fire point location confidence map into the output layer of the residual neural network for normalization processing to generate a normalized probability distribution matrix; performing peak screening on the overlapping probability regions in the normalized probability distribution matrix by a non-maximum suppression algorithm to generate a de-redundant hidden fire point location confidence map; The step of performing spatial intersection and union operations on the hidden fire point location confidence map and the combustible material distribution layer in a preset geographic information system to determine a target hidden fire point coordinate set includes: Performing rasterization processing on the hidden fire point location confidence map to generate a first raster data layer, wherein each raster unit in the first raster data layer contains a corresponding hidden fire point existence probability value; Performing equal-resolution raster conversion on the combustible distribution layer to generate a second raster data layer, wherein each raster cell in the second raster data layer contains a combustible density level value; Using the set hidden fire point existence probability threshold and combustible material density level threshold, the first raster data layer and the second raster data layer are respectively subjected to binary filtering to generate a frequent probability hidden fire point area mask and a frequent combustible material density area mask; Perform spatial overlay analysis on the frequent probability hidden fire point area mask and the frequent combustible material density area mask to extract a set of overlapping grid cells that simultaneously meet the hidden fire point existence probability threshold and the combustible material density level threshold; perform morphological closing operation on the overlapping grid cell set to eliminate discretely distributed isolated grid cells and generate a continuous space area cluster; perform a connected domain labeling algorithm on the continuous space area cluster to identify the circumscribed rectangular bounding box of each independent connected domain, and calculate the weighted centroid coordinates of the hidden fire point existence probability of the grid cells in each circumscribed rectangular bounding box; The weighted centroid coordinates of the existence probability of the hidden fire point are mapped to the geographic coordinate system of the preset geographic information system to generate a candidate hidden fire point coordinate set; based on the combustible distribution gradient direction of each coordinate in the candidate hidden fire point coordinate set, the direction consistency check is performed on adjacent coordinates, and the abnormal coordinate points whose gradient direction does not match the diffusion trend of the hidden fire point are eliminated to generate the target hidden fire point coordinate set; The target hidden fire point coordinate set is spatially matched with the historical hidden fire point verification database, and the newly added hidden fire point coordinates that do not exist in the historical hidden fire point verification database are screened out and updated to the target hidden fire point coordinate set.

10. A fire monitoring and early warning device, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images