A fire point identification method, device, medium and program product based on stationary satellite

Abstract

A method, device, medium, and program product for fire point identification based on geostationary satellites, relating to the field of early fire warning. Points where the difference between the brightness temperature of a suspected fire point and the background brightness temperature is below a preset threshold are identified as critical fire points. For newly appearing critical fire points, the method captures the unique temperature change characteristics of the fire point by analyzing the difference between its brightness temperature change rate and the background change rate. For critical fire points that have appeared before, the method effectively identifies the dynamic evolution characteristics of the fire point by constructing a high-temperature region and analyzing the time-series variation of its morphological measurement parameters. This multi-dimensional judgment mechanism, combining time-varying rate analysis and spatial morphological evolution analysis, improves the sensitivity of identifying initial fires and small-scale fire points while maintaining a low false alarm rate, providing more reliable technical support for early fire warning.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a fire point identification method, device, medium and program product based on geostationary satellites. Background Technology

[0002] With the intensification of global climate change, the frequency of fires has been increasing in recent years, posing a threat to the ecological environment and the safety of people's lives and property. Satellite remote sensing, with its wide-area, all-weather observation capabilities, has become an important means of fire monitoring.

[0003] In related technologies, high-frequency observation data from geostationary meteorological satellites (such as the Himawari series and GK series) are commonly used. By labeling the attributes of pixels in satellite remote sensing images, and based on physical parameters such as visible light reflectance, brightness temperature values ​​in different infrared bands, and flare angles, the image area is pre-divided into various types such as cloud areas, water bodies, flare areas, and desert areas, thus eliminating known interference sources in advance. Subsequently, in the remaining background area, suspected fire points are identified by setting fixed or semi-fixed brightness temperature thresholds.

[0004] However, the core of the relevant technology lies in identifying signals exhibiting strong thermal anomalies in space, which is effective for fires that have developed to a certain scale and whose thermal radiation signal intensity is sufficient to produce significant statistical differences. However, for some small-scale, slow-burning, or smoldering fires, as well as early-stage fires in their nascent stages, their small burning area and low temperature result in very weak thermal signals on satellite pixels with kilometer-level resolution. In the relevant technologies, their brightness temperature may only be slightly higher than the background area, insufficient to meet the aforementioned statistical identification conditions, and therefore may be repeatedly identified as low-confidence suspected fires. During subsequent early warning, these may be easily ignored as background fluctuations or noise, leading to the risk of missed or delayed reporting in small fire identification, thus missing the optimal opportunity for early warning and intervention at the initial stage of a fire. Summary of the Invention

[0005] This application provides a fire point identification method, device, medium, and program product based on geostationary satellites, which can improve the accuracy and real-time performance of fire point identification.

[0006] Firstly, this application provides a method for fire point identification based on geostationary satellites. The method includes: the device, based on suspected fire points pre-extracted from geostationary satellite remote sensing images, identifies suspected fire points whose brightness temperature difference with the corresponding background brightness temperature is lower than a preset brightness temperature difference threshold as critical fire points; the background brightness temperature is the average brightness temperature within a preset range of the area where the suspected fire point is located; if the device determines that the critical fire point has no historical correlation data within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than a preset change threshold, the device determines that the critical fire point is a high-confidence fire point; if the device determines that the critical fire point has historical correlation data within the preset period... The device uses data linking to identify pixel units with brightness temperatures higher than a preset average threshold within a preset neighborhood space, centered on the critical fire point, and combines these pixel units into a high-temperature region. Based on the current morphological measurement parameters of the high-temperature region at the current moment and the historical morphological measurement parameters of the critical fire point within the preset period prior to the current moment, the device calculates the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter. The morphological measurement parameter includes the area of ​​the high-temperature region and the compactness calculated from the area and the perimeter of the boundary of the high-temperature region. If the device determines that the time series variation coefficient is not less than a preset time series change threshold, it determines that the critical fire point is a high-confidence fire point.

[0007] By employing the aforementioned technical solution, points where the difference between the brightness temperature of a suspected fire point and the background brightness temperature is below a preset threshold are identified as critical fire points. This effectively addresses the technical problem of traditional fixed threshold methods' difficulty in identifying small-scale initial fires. For newly appearing critical fire points, the method captures the unique temperature change characteristics of the fire point by analyzing the difference between its brightness temperature change rate and the background change rate. For critical fire points that have appeared before, the method effectively identifies the dynamic evolution characteristics of the fire point by constructing a high-temperature region and analyzing the time-series variation of its morphological measurement parameters. This multi-dimensional judgment mechanism, combining time-varying rate analysis and spatial morphological evolution analysis, improves the sensitivity of identifying initial fires and small-scale fire points while maintaining a low false alarm rate, providing more reliable technical support for early fire warning.

[0008] In some embodiments, in conjunction with the first aspect, the method further includes: the device determining the classification result corresponding to the suspected fire point; the device inputting the multispectral data corresponding to the suspected fire point in the geostationary satellite remote sensing image into a pre-trained fire point recognition model to obtain the model identification confidence level that the suspected fire point is a real fire point; the device obtaining adaptive weights based on the environmental complexity index of the area where the suspected fire point is located by querying a preset complexity weight mapping table, the environmental complexity index being determined by comprehensive quantification based on land cover type data and historical thermal anomaly database, the adaptive weights including a first weight and a second weight, and the higher the environmental complexity index, the larger the ratio of the second weight to the first weight; the device weighting the classification result based on the first weight and weighting the model identification confidence level using the second weight, generating a fire point confirmation score through weighted sum calculation; and the device identifying the suspected fire point with a fire point confirmation score higher than a historical preset threshold as a real fire point.

[0009] By adopting the above technical solution and introducing the key parameter of environmental complexity index, the complexity of the environment in which suspected fire points are located can be quantitatively assessed, and the weighting of traditional classification results and deep learning model identification results can be dynamically adjusted accordingly. In areas with high environmental complexity, the reliance on the deep learning model is increased; while in areas with simpler environments, the reliance on traditional classification methods is increased. This adaptive weighting mechanism effectively solves the technical problem that a single discrimination method cannot adapt to diverse environments. By generating a fire point confirmation score through weighted fusion and comparing it with historical preset thresholds, more accurate fire point determination is achieved, improving the accuracy of fire point identification in complex and changing environments and reducing the false alarm rate.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the step of the device determining the classification result corresponding to the suspected fire point specifically includes: if the device determines that the critical fire point has no historical associated data within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than a preset change threshold, the device determines the numerical value of the classification result based on the magnitude by which the difference exceeds the preset change threshold; if the device determines that the critical fire point has historical associated data within the preset period, and the time series variation coefficient is not less than a preset time series change threshold, the device determines the numerical value of the classification result based on the magnitude by which the time series variation coefficient exceeds the preset time series change threshold; the greater the magnitude by which the time series variation coefficient exceeds the preset time series change threshold, the greater the numerical value of the classification result; if the device determines that the temperature change rate difference of the critical fire point does not reach the preset change threshold, or the time series variation coefficient is less than the preset time series change threshold, the classification result is not included in the calculation.

[0011] By employing the above technical solution, for the first occurrence of a critical fire point, the classification result is quantified based on the magnitude of the temperature change rate difference exceeding a preset threshold, ensuring that points with more significant temperature change characteristics receive higher classification result values. For critical fire points that are not first occurrences, the classification result is determined based on the magnitude of the time series coefficient of variation exceeding a threshold, ensuring that points with more obvious morphological evolution characteristics receive higher classification result values. When a critical fire point fails to pass either of these two criteria, the classification result is excluded from the calculation, ensuring the rigor of the judgment. This refined quantification mechanism based on the degree of feature significance enables the differentiation of fire points with different confidence levels, providing a richer information foundation for subsequent weight fusion, effectively improving the accuracy and reliability of fire point judgment, and reducing the uncertainty of judgment in boundary cases.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the environmental complexity index is generated by comprehensively quantifying land cover type data and a historical thermal anomaly database. Specifically, this includes: the device constructing a spatiotemporal correlation network based on historical thermal anomaly events in the database as nodes. This spatiotemporal correlation network includes establishing a directed edge between corresponding nodes when the spatial distance between two thermal anomaly events is less than a preset correlation distance threshold and the time interval is less than a preset correlation time window. The weight of the edge is determined by the reciprocal of the spatial distance and a decay function of the time interval. The device then executes [the necessary steps] within this spatiotemporal correlation network. A random walk algorithm is used to statistically determine the number of paths and the average path length reaching the suspected fire point within a preset neighborhood. The number of paths represents the activity level of the thermal anomaly, and the average path length represents the spatial clustering characteristics of the thermal anomaly. Based on the land cover type patches within the preset spatial range corresponding to the suspected fire point location, the device calculates a fragmentation index by weighting the ratio of the boundary length to the area of ​​the land cover type patches. The weight of the land cover type is obtained based on a preset land cover database. Based on the number of paths, the average path length, and the fragmentation index, the device determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean.

[0013] By employing the aforementioned technical solutions and constructing a spatiotemporal correlation network based on historical thermal anomalies, the spatiotemporal distribution patterns of thermal anomaly events within a region can be captured. A random walk algorithm is used to obtain the number of paths reflecting the activity level of thermal anomalies and the average path length characterizing spatial clustering. Furthermore, by analyzing the boundary length-to-area ratio of land cover type patches, a fragmentation index reflecting surface heterogeneity is calculated. These three dimensions collectively form the basis for assessing environmental complexity. By calculating the ratio of the weighted geometric mean to the arithmetic mean, an index comprehensively reflecting environmental complexity is obtained. This multi-dimensional quantification method effectively solves the technical problem that traditional single indicators are insufficient for comprehensively assessing environmental complexity, provides a scientific basis for adaptive weight allocation, and improves adaptability under different environmental conditions.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the step of determining the ratio of the weighted geometric mean to the arithmetic mean as the environmental complexity index based on the number of paths, the average path length, and the fragmentation index specifically includes: the device, based on the geographical location of the suspected fire point, retrieves historical verified fire point data within a preset spatiotemporal range, and statistically analyzes the distribution of the identification success rate of the historical verified fire point data within different time delay intervals, where the time delay is the difference between the actual occurrence time of the fire point and the first identification time of the algorithm; the device obtains the regional identification difficulty coefficient by performing a weighted integral of the identification success rate according to the time delay based on the identification success rate distribution; and the device determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean based on the number of paths, the average path length, the regional identification difficulty coefficient, and the fragmentation index.

[0015] By employing the aforementioned technical solution, historical verification fire point data within a preset spatiotemporal range is retrieved, and the distribution of recognition success rates across different time delay intervals is analyzed to quantitatively assess the regional recognition difficulty coefficient. This coefficient reflects the historical success rate of fire point recognition within a specific region, providing empirical evidence for environmental complexity assessment. This regional recognition difficulty coefficient, along with the number of paths, average path length, and fragmentation index, is incorporated into the environmental complexity calculation, and the final index is obtained by comparing the weighted geometric mean with the arithmetic mean. This environmental complexity assessment method, which incorporates historical verification data, effectively addresses the technical problem that purely theoretical models may not accurately reflect actual recognition difficulty, making the environmental complexity index more aligned with practical application needs and further improving the accuracy of adaptive weight allocation.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after determining that historical correlation data exists within the preset period for the critical fire point, identifying pixel units with brightness temperatures higher than a preset average threshold in a preset neighborhood space centered on the critical fire point, and combining these pixel units into a high-temperature region, the method further includes: the device performing spectral analysis on the brightness temperature time series of the critical fire point to obtain periodic and aperiodic components, the brightness temperature time series being composed of brightness temperatures at consecutive preset observation times; the device calculating the proportion of the aperiodic component in the time series and extracting pulse features from the aperiodic component, the pulse features including pulse amplitude and pulse interval; the device determining an anomaly score based on the statistical distribution of the pulse features, the greater the variability of the pulse amplitude and the weaker the regularity of the pulse interval, the higher the anomaly score; when the anomaly score exceeds a preset anomaly threshold and the proportion exceeds a preset energy threshold, the device performing time window expansion on the preset period, the time window expansion being to extend the preset period to a preset multiple of the original period.

[0017] By employing the aforementioned technical solution, spectral analysis is performed on the brightness temperature time series of critical fire points to separate periodic and aperiodic components, with a focus on analyzing the pulse characteristics within the aperiodic components. By calculating the proportion of aperiodic components and the anomaly score of the pulse characteristics, fire points exhibiting abnormal temperature change patterns can be identified. When both the anomaly score and the proportion of aperiodic components exceed preset thresholds, the time window is automatically extended, prolonging the analysis period. This dynamic feature recognition mechanism based on spectral analysis effectively solves the technical problem that traditional fixed-time-window analysis may miss long-period abnormal changes, enhances the ability to identify fire points with complex temporal evolution characteristics, improves the detection sensitivity for intermittent and slowly developing fires, and reduces the false negative rate.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, the step of calculating the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter based on the current morphological measurement parameter of the high-temperature region at the current moment and the historical morphological measurement parameter corresponding to the critical fire point within the preset period before the current moment specifically includes: the device constructs a morphological evolution trajectory sequence of the historical morphological measurement parameters in chronological order, and then divides the morphological evolution trajectory sequence into multiple overlapping sub-sequence segments through a sliding window, each sub-sequence segment containing a consecutive preset number of historical morphological measurement parameters; the device obtains the local variation coefficient of the sub-sequence segment by calculating the ratio of the standard deviation to the mean of the morphological measurement parameters within the sub-sequence segment; the device extracts a morphological stability benchmark value based on the distribution characteristics of the local variation coefficient; the device extracts the nearest local variation coefficient corresponding to the last sub-sequence segment in chronological order among the multiple sub-sequence segments, and determines the ratio of the nearest local variation coefficient to the morphological stability benchmark value as the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter.

[0019] By employing the aforementioned technical solution, the morphological evolution trajectory sequence is divided into multiple overlapping sub-sequence segments using a sliding window. The local coefficient of variation for each sub-sequence segment is calculated, thereby capturing the local dynamic characteristics of morphological changes. By analyzing the distribution characteristics of the local coefficient of variation, a morphological stability benchmark value is extracted, establishing a reference standard for morphological changes. Finally, by comparing the local coefficient of variation of the most recent sub-sequence segment with the morphological stability benchmark value, the time series coefficient of variation is calculated, achieving an accurate assessment of the significance of current morphological changes. This morphological evolution assessment method based on sliding windows and local variation analysis effectively solves the technical problem that global statistics may mask significant local changes, enhances the sensitivity to subtle but continuous morphological changes, improves the early identification capability of slowly expanding fires, and provides a more reliable basis for fire point determination based on dynamic morphological characteristics.

[0020] In a second aspect, this application provides an apparatus comprising: one or more processors and a memory; the memory being coupled to the one or more processors, the memory being used to store computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the apparatus to perform the method as described in the first aspect and any possible implementation thereof.

[0021] Thirdly, this application provides a computer program product containing instructions that, when run on a device, cause the device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a device, cause the device to perform the method described in the first aspect and any possible implementation thereof.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] 1. By adopting the above technical solution, due to the use of critical fire point identification and multi-dimensional judgment mechanism, including brightness temperature difference analysis, temperature change rate comparison and morphological measurement parameter time series variation analysis, the technical problem that the fixed threshold method in the existing technology is difficult to identify small-scale fires in the early stage is effectively solved. Thus, the technical effect of improving the identification sensitivity of initial fires and small-scale fire points is achieved while maintaining a low false alarm rate.

[0025] 2. By adopting the above technical solution, due to the use of an adaptive weight allocation and multi-source information fusion mechanism based on environmental complexity, including quantitative assessment of environmental complexity, dynamic weight adjustment and weighted sum calculation, the technical problem that the single discrimination method in the existing technology is difficult to adapt to diverse environments is effectively solved. Thus, the technical effects of significantly improving the accuracy of fire point identification, reducing false alarm rate and enhancing environmental adaptability in complex and changeable environments are achieved.

[0026] 3. By adopting the above technical solution, the fine quantification mechanism of classification results based on feature significance is used, including the quantification of temperature change rate difference exceeding the amplitude and the quantification of time series variation coefficient exceeding the amplitude. Therefore, the technical problems of binary fire point determination results and lack of confidence differentiation in the existing technology are effectively solved, thereby achieving the technical effect of improving the accuracy of fire point determination and reducing the uncertainty of boundary condition determination. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a fire point identification method based on geostationary satellites in an embodiment of this application.

[0028] Figure 2 This is another flowchart illustrating a fire point identification method based on geostationary satellites in an embodiment of this application;

[0029] Figure 3 This is a schematic diagram of an exemplary hardware structure of a remote sensing monitoring device in the embodiments of this application. Detailed Implementation

[0030] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0031] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0032] Please see Figure 1 This is a flowchart illustrating a fire point identification method based on geostationary satellites in an embodiment of this application.

[0033] S101. Based on suspected fire points pre-extracted from geostationary satellite remote sensing images, the equipment identifies suspected fire points whose brightness temperature difference with the corresponding background brightness temperature is lower than a preset brightness temperature difference threshold as critical fire points.

[0034] Geostationary satellite remote sensing imagery refers to remote sensing data acquired by meteorological or resource satellites located in geosynchronous orbit (at an altitude of approximately 36,000 kilometers). These satellites remain stationary relative to the Earth's surface, enabling high-frequency and continuous observation of the same area. Typical geostationary meteorological satellites include China's Fengyun-4, the US GOES series, and Japan's Himawari series. Suspected fire points refer to a set of pixels in remote sensing imagery that exhibit thermal anomalies and are preliminarily identified as potential fire heat sources. These points typically show high brightness temperature values ​​in the thermal infrared band (such as 3.9μm mid-wave infrared or 10.8μm long-wave infrared). Brightness temperature refers to the equivalent blackbody temperature obtained by inverting the radiant energy received by the remote sensor using Planck's law, measured in Kelvin (K), or simply brightness temperature, reflecting the intensity of thermal radiation from ground features. Equipment can be a dedicated server cluster deployed in a data center, a graphics workstation with high-performance computing capabilities, or a cloud-based virtual computing instance, responsible for processing massive amounts of satellite remote sensing data.

[0035] Specifically, the equipment receives and processes multi-band remote sensing data from geostationary satellites, which typically includes multiple bands such as visible light, near-infrared, and thermal infrared. The equipment generates standardized remote sensing products through preprocessing steps (including radiometric calibration, geometric correction, and cloud detection). Subsequently, based on the brightness temperature anomaly characteristics of the thermal infrared band (primarily the 3.9μm mid-wave infrared band, due to its high sensitivity to fire points), the equipment, combined with multi-band thresholding or contextual analysis, initially extracts a set of all possible suspected fire points. For each suspected fire point, the equipment accurately calculates its brightness temperature value and defines a spatial window (such as a 5×5 or 7×7 pixel matrix) centered on that point as the background region. The equipment statistically analyzes the brightness temperature values ​​of the effective background pixels within this window (excluding clouds, water bodies, and other suspected fire points) to calculate a background brightness temperature value representing the normal temperature level of the region. Then, the equipment calculates the difference between the brightness temperature of the suspected fire point and the background brightness temperature and compares this difference with a preset brightness temperature difference threshold. If the difference is higher than the threshold (e.g., difference > 20K), the suspected fire point is directly identified as a high-confidence fire point; if the difference is positive but lower than the threshold (e.g., 0K < difference ≤ 20K), the suspected fire point is marked as a critical fire point and needs to enter the subsequent spatiotemporal analysis process; if the difference is negative or zero, the point is excluded from the fire point candidate set.

[0036] In some embodiments, the device receives remote sensing images from geostationary satellites (such as Himawari-8 / 9, GK2A, etc.) and identifies suspected fire points from the images. Specifically, this includes: identifying and removing pixels covered by clouds using a cloud detection algorithm, marking the image data as necessary, and assigning metadata such as geographic coordinates to each pixel unit; then performing preliminary threshold screening based on the spectral anomalies of the pixels (e.g., abnormally high radiance values ​​in the mid-infrared band) to identify all thermodynamically suspicious targets, i.e., the set of suspected fire points.

[0037] It is understandable that the specific implementation methods of the preprocessing stage, including cloud detection algorithms (such as thresholding, decision tree, or deep learning), pixel tag encoding methods, and spectral feature selection for initial screening of potential fire points (such as single-band brightness temperature threshold, multi-band difference, or ratio index), are all relevant technologies and are not limited here; the shape (rectangular, circular, or irregular) and size (fixed or adaptive) of the background window are not limited here; the calculation method of background brightness temperature (such as arithmetic mean, weighted average, median, or other statistics) is not limited here; the preset brightness temperature difference threshold can be a global constant or can be set according to land cover type, season, and solar zenith angle, and is not limited here.

[0038] S102. If the equipment determines that there is no historical associated data for the critical fire point within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than the preset change threshold, the critical fire point is determined to be a high-confidence fire point.

[0039] Among them, the brightness temperature change rate refers to the amount of change in brightness temperature value per unit time, usually expressed as the amount of temperature change per hour or per observation cycle (such as every 15 minutes) (K / h or K / cycle).

[0040] The equipment queries historical monitoring records. If the critical fire point location has not been detected as a critical fire point or fire point within a preset time window (e.g., the past 24 hours or longer), it is considered to have no historical associated data within the preset period, indicating its first occurrence. The equipment then extracts the brightness temperature data sequence of the critical fire point location for the most recent consecutive observation times (e.g., for a geostationary satellite with a 15-minute observation interval, this could be the most recent 4-6 times, covering approximately 1-1.5 hours). Based on this time sequence, the equipment calculates the brightness temperature change rate of the critical fire point, which can be obtained through a simple linear regression method (least squares method) or a difference method (latest brightness temperature minus the earliest brightness temperature, then divided by the time interval). Simultaneously, the equipment also extracts the brightness temperature sequence for the same time period for the background region defined around the critical fire point (usually the same as the background window used in step S101) and calculates the average brightness temperature change rate of the background region. The equipment calculates the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate, which reflects the additional warming rate of the critical fire point relative to its environment. If this difference is greater than a preset change threshold (e.g., for a 15-minute observation interval, the threshold might be set to 1.5K / 15min), it indicates that the temperature rise rate of the critical fire point is significantly faster than its surrounding environment, consistent with the thermodynamic characteristics of the early stage of a fire, and therefore it is determined to be a high-confidence fire point. The equipment calculates the difference between these two rates of change. If this difference is greater than a preset change threshold (e.g., 2K / h), it indicates that the temperature rise rate of the critical fire point is significantly faster than its surrounding environment, consistent with the thermodynamic characteristics of the early stage of a fire, and therefore it is determined to be a high-confidence fire point.

[0041] Understandably, in the early stages of a fire, the temperature in the fire zone usually rises faster than the surrounding environment. This dynamic change is an important basis for distinguishing a real fire from other stable heat sources (such as industrial facilities and urban heat islands).

[0042] S103. If the device determines that there is historical correlation data in the critical fire point within the preset period, it identifies the pixel units with brightness and temperature higher than the preset average threshold in the preset neighborhood space, centered on the critical fire point, and combines the pixel units into a high-temperature area.

[0043] If the device determines that there is historical associated data for the critical fire point within the preset period, it takes the currently detected critical fire point pixel as the center and determines a preset neighborhood space, which is usually a rectangular or circular area of a fixed size (such as a rectangular window of 7×7 pixels or a circular area with a radius of 3 pixels). Then, the device checks the brightness temperature values of each pixel within this neighborhood space and compares them with a preset average threshold. This preset average threshold is usually an absolute temperature value (such as 320K) or a temperature increment relative to the background (such as the background average temperature plus 10K). For pixels within the neighborhood space whose brightness temperature is higher than the preset average threshold, the device marks them as high-temperature pixel units. These high-temperature pixel units may be part of the fire point, or the surrounding area affected by the fire point, or other types of thermal anomalies. Subsequently, the device combines all adjacent high-temperature pixel units together to form one or more high-temperature regions. If there are multiple non-connected high-temperature regions, usually the region containing the original critical fire point is selected as the object for subsequent analysis.

[0044] Specifically, after the device confirms that there is historical associated data for the critical fire point within the preset period, it defines a fixed-size window centered on the critical fire point (such as 11×11 pixels) as the analysis area. By calculating the histogram of the brightness temperature of all pixels within the analysis area and based on the distribution characteristics of the histogram, it determines N temperature thresholds (such as N = 3), dividing the brightness temperature range into N + 1 levels. For example, the Otsu multi-threshold method or the K-means clustering method can be used to determine these thresholds, obtaining T_1 < T_2 <... < T_N. The device starts from the highest threshold T_N, marks all pixels within the analysis area whose brightness temperature is higher than T_N, and combines adjacent marked pixels into connected regions to form the high-temperature core area of the highest level. Then, the device reduces the threshold to T_(N - 1), marks the pixels whose brightness temperature is higher than T_(N - 1) but lower than T_N, and combines these pixels with the existing high-temperature core area to form the high-temperature region of the second level. This process is repeated step by step until the lowest threshold T_1 is processed.

[0045] It can be understood that the specific shape and size of the preset neighborhood space are not limited here; the specific value of the preset average threshold and its calculation method (which can be an absolute value or a relative value) are not limited here; the connection rule for combining high-temperature pixel units into high-temperature regions (such as four-connected or eight-connected) is not limited here; the selection strategy when there are multiple non-connected high-temperature regions is not limited here.

[0046] In some embodiments, during the construction of the high-temperature region, situations may arise where the high-temperature region is too large or irregularly shaped. This could be due to multiple adjacent but independent fire points being incorrectly merged into one region, or due to natural variations in surface temperature (such as the temperature difference between valleys and ridges). The device can introduce a morphological watershed segmentation mechanism. After completing the initial construction of the high-temperature region, gradient calculation is applied to the brightness temperature distribution within the region to obtain a temperature gradient image. On the temperature gradient image, the device identifies local minima as seed points for the watershed algorithm and applies a watershed transformation to segment the high-temperature region into multiple sub-regions. For each sub-region, the device calculates the difference between its center brightness temperature and its edge brightness temperature, as well as the region's compactness (the ratio of area to the square of the perimeter). Based on these characteristics, sub-regions with large differences in brightness temperature between the center and edge and high compactness are retained; these sub-regions are more likely to correspond to real independent fire points.

[0047] In some embodiments, the device applies spectral analysis methods to the preprocessed brightness temperature time series. The most commonly used method is the Fast Fourier Transform (FFT), which decomposes the time series into a combination of sine and cosine waves of different frequencies. The device calculates the power spectral density (PSD) of the time series, displaying the energy distribution of different frequency components. In the PSD plot, peaks correspond to the main periodic components in the time series. For example, for a diurnal variation with a 24-hour cycle, a significant peak will appear at a frequency of approximately 1 / 24 h⁻¹. Based on the power spectral analysis results, the device decomposes the brightness temperature time series into periodic and aperiodic components. Periodic components typically include identifiable natural periodic variations, such as diurnal, semi-diurnal, multi-diurnal, and seasonal variations. The device extracts the frequency components corresponding to these known periods using frequency domain filtering methods to construct a periodic model. Specifically, the device can design a series of bandpass filters, each corresponding to a known natural period, and then apply these filters to the spectrum of the original signal to extract the corresponding periodic components. After subtracting the periodic components from the original brightness temperature time series, the remaining portion is the aperiodic component. This section includes random noise, sudden events (such as fires), and other variations that do not conform to known periodic patterns. For real fires, especially initial or intermittent fires, significant anomalous pulses or abrupt changes are typically observed in the non-periodic components. The device then calculates the proportion of the non-periodic components in the time series and extracts pulse features, including pulse amplitude (intensity of the temperature anomaly) and pulse interval (temporal distribution of the temperature anomaly event). Based on the statistical distribution of these pulse features, the device calculates an anomaly score—a higher anomaly score is awarded when the pulse amplitude variability is greater (indicating unstable combustion) and the pulse interval regularity is weaker (consistent with natural fire characteristics). When the anomaly score exceeds a preset anomaly threshold and the proportion of the non-periodic component exceeds a preset energy threshold, it is determined that more comprehensive historical data is needed to support the analysis. Therefore, the time window is extended for the preset period, extending the analysis time range to a preset multiple of the original period to obtain more sufficient time-series information to assist in the judgment.

[0048] S104. The equipment calculates the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter based on the current morphological measurement parameter of the high-temperature region at the current moment and the historical morphological measurement parameter of the critical fire point within a preset period before the current moment.

[0049] The device calculates a series of morphological metrics that describe the geometric and spatial characteristics of the high-temperature region from different perspectives. Typical morphological metrics include: area (the number of pixels contained in the high-temperature region, reflecting the size of the thermal anomaly), perimeter (the length of the boundary of the high-temperature region, reflecting the boundary complexity), compactness (usually defined as 4π × area / perimeter², with values ​​closer to 1 indicating a shape closer to a circle), directionality (the principal axis direction obtained through principal component analysis, reflecting the extension direction of the high-temperature region), eccentricity (reflecting the elongation of the shape), and fractal dimension (reflecting the complexity of the boundary), etc. Simultaneously, the device retrieves historical observation records of the critical fire point from the database within a preset period (e.g., the past 24 hours) before the current moment, extracting the historical morphological metrics for the corresponding time. These historical parameters may come from records previously identified as critical fire points or fire points. Subsequently, the device calculates the time series coefficient of variation based on the current and historical morphological metrics.

[0050] Understandably, if observational data is missing at certain historical moments (e.g. due to cloud cover), the device can fill in the missing values ​​through interpolation methods, or construct time series with unequal intervals using only valid observations, which is not limited here.

[0051] In some embodiments, the device arranges historical morphological metrics retrieved from the database in chronological order to construct a sequence of morphological evolution trajectories. This sequence may contain data from multiple time points, each corresponding to a set of morphological metrics (such as area, perimeter, compactness, etc.). The device then applies a sliding window technique, setting a fixed-length window (e.g., containing data from 10 consecutive time points), moving the window forward by one or more time points each time point from the beginning of the sequence, and extracting data within the window's coverage area to form a sub-sequence segment. These sub-sequence segments typically overlap to ensure the continuity and smoothness of the analysis. For each sub-sequence segment, the device calculates the mean (μ) and standard deviation (σ) of the morphological metrics within the sub-sequence segment, and then calculates the ratio of the standard deviation to the mean (σ / μ) to obtain the local coefficient of variation for that sub-sequence segment. For multidimensional morphological metrics (such as considering multiple features such as area and perimeter simultaneously), the local coefficient of variation for each dimension can be calculated separately, and then a comprehensive local coefficient of variation can be obtained through methods such as weighted average or principal component analysis. The device extracts a morphological stability baseline value based on the distribution characteristics of the local coefficients of variation (COPs) of all subsequence segments. Extraction methods may include calculating the median of all local COPs (insensitive to outliers), calculating the mean after removing outliers, or finding the dominant peak of the distribution through kernel density estimation. The device extracts the COPs corresponding to the subsequence segment closest in time to the current moment (i.e., the subsequence segment containing data from the most recent historical time points), defining it as the nearest local COP. The device calculates the ratio of the nearest local COP to the morphological stability baseline value to obtain the time series COP.

[0052] In other embodiments, the device can apply linear regression analysis to the time series of historical morphological measurement parameters to obtain the slope characterizing the trend of change. Then, the device calculates the deviation between the current morphological measurement parameter and the expected value predicted based on the historical trend, and standardizes this deviation (dividing by the standard deviation of the historical parameters) to obtain the standardized deviation. This standardized deviation can serve as the time series coefficient of variation, reflecting the degree of anomalousness of the current morphology relative to the historical trend; this is not limited here.

[0053] S105. If the equipment determines that the coefficient of variation of the time series is not less than the preset time series change threshold, the critical fire point is determined to be a high confidence fire point.

[0054] Specifically, the equipment first obtains the time series coefficient of variation value calculated in step S104. A higher coefficient of variation indicates that the morphological characteristics of the current high-temperature area have changed significantly, which may indicate the sudden outbreak or spread of a fire; while a lower coefficient of variation indicates that the morphological changes are in line with historical patterns and may be a stable non-fire heat source.

[0055] The device then compares this coefficient of variation with a preset time-series change threshold. The threshold is typically set based on statistical analysis of a large amount of historical verification data, determining the optimal value while balancing the false alarm and false alarm rates. Different threshold standards may need to be set for different types of land cover (such as forests, grasslands, farmland, and cities) and different climatic conditions (such as dry seasons and rainy seasons) to accommodate the differences in fire morphology evolution under various scenarios. If the calculated time-series coefficient of variation is not less than (i.e., greater than or equal to) the preset time-series change threshold, it indicates that the morphological changes in the current high-temperature area significantly exceed the normal range of its historical evolution, consistent with the dynamic characteristics of a real fire. In this case, the device classifies the critical fire point as a high-confidence fire point. This determination result will be recorded in the system database and may trigger a series of subsequent operations, such as generating a fire alarm, notifying relevant departments, and initiating emergency response procedures.

[0056] If the coefficient of variation is less than the preset threshold, it indicates that the morphological changes in the current high-temperature area are within the normal range, and may be a stable non-fire heat source (such as industrial facilities, urban heat islands, etc.) or low-intensity controlled combustion (such as agricultural straw burning). In this case, the critical fire point will remain in its critical state and will continue to be monitored but will not trigger a high-level alarm.

[0057] By employing a critical fire point intelligent determination technology based on spatiotemporal feature fusion, this technology effectively solves the technical problems of balancing false alarms and missed detections, and the inability of static determination to capture the dynamic characteristics of fires. This results in improved fire point identification accuracy, reduced false alarm rates, and enhanced detection capabilities for small-scale initial fires. Specifically, by organically combining brightness-temperature difference analysis, time-varying rate analysis, spatial morphology construction, and temporal morphological evolution analysis, it can consider both the static characteristics and dynamic evolution patterns of fire points. In particular, it accurately determines critical fire points that, while not exhibiting strong thermal signals, possess clear spatiotemporal evolution characteristics. This maintains high detection sensitivity while reducing false alarm rates, providing more reliable technical support for early fire warning and emergency response.

[0058] In the above embodiments, the device can achieve high-precision fire point identification by fusing spatiotemporal feature analysis with intelligent critical fire point determination technology. However, in practical applications, when implementing the aforementioned geostationary satellite-based fire point identification method, it still faces the technical challenge of unifying and optimizing the discrimination strategy due to differences in environmental complexity. This geostationary satellite-based fire point identification method can solve this technical problem by fusing adaptive weight allocation based on environmental complexity, thereby improving the accuracy and reliability of fire point identification under different environmental conditions.

[0059] Please see Figure 2 This is another flowchart illustrating a fire point identification method based on geostationary satellites in an embodiment of this application.

[0060] S201. Based on suspected fire points pre-extracted from geostationary satellite remote sensing images, the equipment identifies suspected fire points whose brightness temperature difference with the corresponding background brightness temperature is lower than a preset brightness temperature difference threshold as critical fire points.

[0061] S202. If the equipment determines that there is no historical associated data for the critical fire point within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than the preset change threshold, the critical fire point is determined to be a high-confidence fire point.

[0062] S203. If the device determines that there is historical correlation data for the critical fire point within a preset period, it identifies pixel units with brightness and temperature higher than the preset average threshold in the preset neighborhood space, centered on the critical fire point, and combines the pixel units into a high-temperature region.

[0063] S204. The equipment calculates the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter based on the current morphological measurement parameter of the high-temperature region at the current moment and the historical morphological measurement parameter of the critical fire point within a preset period before the current moment.

[0064] S205. If the equipment determines that the coefficient of variation of the time series is not less than the preset time series change threshold, the critical fire point is determined to be a high confidence fire point.

[0065] Steps S201 to S205 are similar to steps S101 to S105, and will not be described in detail here.

[0066] S206. The equipment determines the classification result corresponding to the suspected fire point.

[0067] The equipment determines the classification results corresponding to suspected fire points. These results represent the output of classifying suspected fire points based on brightness temperature differences, temporal change rates, and morphological evolution analyses as described in steps S201-S205. Typically, these include three categories: high-confidence fire points, critical fire points, and non-fire points. The equipment first integrates the judgment results from the preceding steps and classifies each suspected fire point pre-extracted from the geostationary satellite remote sensing image. For suspected fire points whose brightness temperature difference with the background brightness temperature exceeds a preset brightness temperature difference threshold in step S201, they are directly classified as high-confidence fire points due to their significant thermal anomaly characteristics. For suspected fire points identified as critical fire points, the equipment further checks whether they meet the judgment conditions S202 or S205: if the critical fire point has no historical correlation data within a preset period and the difference between its brightness temperature change rate and background change rate is greater than a preset change threshold (condition S202 is met), or if the critical fire point has historical correlation data within a preset period and its time series variation coefficient is not less than a preset time series change threshold (condition S205 is met), then it is classified as a high-confidence fire point; otherwise, its critical fire point classification result is maintained. Suspected fire points that do not meet either the high-confidence fire point condition or the critical fire point condition are classified as non-fire points. Then, the equipment converts these classification results into numerical form.

[0068] In some embodiments, the determination of suspected fire point classification results can be achieved in multiple ways:

[0069] Optionally, if a critical fire point is determined to be a high-confidence fire point in S202 (first occurrence and temperature change rate difference exceeding a threshold), the classification result value is calculated based on the magnitude of its exceeding the preset change threshold; if a critical fire point is determined to be a high-confidence fire point in S205 (not first occurrence and time series coefficient of variation exceeding a threshold), the classification result value is calculated based on the magnitude of the coefficient of variation exceeding the preset time series change threshold. For critical fire points not determined to be high-confidence fire points, their classification results are not included in the calculation and can be represented by a score of 0.

[0070] Optionally, the device can implement a weighted classification method based on decision paths. First, it analyzes historical validation data and calculates accuracy statistics (such as precision and recall) for each decision path. Then, the device assigns a reliability weight to each decision path based on these statistics. When a suspected fire point is identified as a fire point by a certain path, its classification score not only considers the basic category score but also multiplies it by the reliability weight of that path. For example, if historical data indicates that a decision based on brightness-temperature difference is more reliable than a decision based on morphological evolution, the former might have a weight of 0.9, and the latter 0.7. This weighted method, which considers the reliability of decision paths, can generate classification results that better reflect historical experience. It is understood that other methods can also be used to determine the classification results; this is not limited here.

[0071] S207. The equipment inputs the multispectral data corresponding to the suspected fire point in the geostationary satellite remote sensing image into the pre-trained fire point recognition model to obtain the model identification confidence level of the suspected fire point as the real fire point.

[0072] Multispectral data refers to electromagnetic radiation information collected by geostationary satellite sensors in different wavelength ranges. It typically includes image data in multiple bands such as visible light (e.g., red, green, and blue bands), near-infrared, short-wave infrared, mid-wave infrared, and long-wave infrared, with each band reflecting different physical characteristics of ground objects. A pre-trained fire point recognition model represents a deep learning model trained on a large amount of labeled data, specifically designed to identify fire point features from multispectral remote sensing data. Typical model architectures may include convolutional neural networks (CNN), recurrent neural networks (RNN), or combinations thereof. Model identification confidence refers to the probability value or score output by the deep learning model, typically ranging from 0 to 1. It represents the model's confidence that a suspected fire point is a real fire point; a higher value indicates that the model is more confident that the point is a real fire point.

[0073] The device first extracts multispectral data corresponding to each suspected fire point from a geostationary satellite remote sensing image database. This data typically includes multiple channels covering the visible to thermal infrared bands, each capturing the radiation characteristics of ground features within a specific wavelength range. The device extracts not only the spectral information of the suspected fire point pixels themselves, but also spatial context information within a certain range around them (e.g., a 64×64 pixel window), enabling the model to analyze the relationship between the fire point and its surrounding environment. Furthermore, the device may extract multispectral data for that location over several past time periods to construct temporal features, allowing the model to capture the dynamic changes of the fire point. The device preprocesses the extracted multispectral data, including radiometric correction (eliminating atmospheric and sensor effects), geometric correction (ensuring accurate spatial location), and data standardization (unifying the numerical ranges of different bands), to make the data conform to the model's input requirements. For temporal data, time alignment and missing value handling are also required. The preprocessed data is organized into the input format required by the model, typically a multidimensional tensor containing spatial dimensions (height, width), spectral dimensions (number of bands), and possibly a temporal dimension. The device inputs the prepared data into a pre-trained fire detection model. A lightweight MobileNetV3 backbone network can be used, integrating a Transformer encoder to capture temporal features and introducing a special layer to describe the fire spread pattern. The model's input includes multispectral band slices (64×64 pixels) and temporal difference features (changes over three frames). After processing the input data, the model generates a probability value between 0 and 1 through its output layer.

[0074] In some embodiments, the model identification confidence score can be calculated in a variety of ways:

[0075] Optionally, the device can employ a multi-scale feature fusion network. First, it constructs a pyramid-structured feature extraction network to process the input multispectral data in parallel at different spatial resolution levels. For example, an original 64×64 pixel input can generate feature maps at three scales: 64×64, 32×32, and 16×16. Each scale's feature map extracts features at that scale through independent convolutional layers. Then, upsampling or downsampling operations are used to align features at different scales to the same resolution, and attention mechanisms or feature fusion layers are used to combine them into a unified multi-scale feature representation. For temporal features, the device uses a Transformer encoder to process changes between three consecutive frames, capturing temporal dependencies through a self-attention mechanism. Finally, a multilayer perceptron maps the fused spatial-spectral-temporal features to the final confidence output.

[0076] Optionally, the device can implement an ensemble learning framework to integrate the judgments of multiple expert models. The device can simultaneously deploy multiple expert models, such as a CNN-based spatial feature model, an RNN-based temporal feature model, and a Transformer-based attention model. Each model independently processes the input multispectral data and generates its own fire point confidence prediction. Then, the device fuses these predictions using weighted averaging, voting, or more complex meta-learners (such as random forests or neural networks) to obtain the final model judgment confidence. It is understood that other methods can also be used, and this is not limited here.

[0077] Furthermore, it should be noted that the training process of the pre-trained fire detection model can be optimized using various strategies. For example, the model training uses Focal Loss as the loss function, and the training data includes multispectral band slices and temporal difference features. Data augmentation techniques (such as rotation, flipping, and noise addition) are used to expand the training set, and the model is deployed on an NVIDIA A10 GPU to achieve an inference speed of no more than 1.5 seconds per frame, meeting the requirements of real-time monitoring.

[0078] It is understandable that the above fire detection model is a related technology, and will not be elaborated here.

[0079] S208. The device obtains adaptive weights by querying a preset complexity weight mapping table based on the environmental complexity index of the suspected fire point area.

[0080] Among them, the environmental complexity index is a numerical indicator that quantifies the complexity of environmental conditions in the area where a suspected fire point is located. It comprehensively considers various factors such as topographic relief, surface cover heterogeneity, intensity of human activities, and changes in meteorological conditions. The higher the value, the more complex the environment and the greater the difficulty in identifying the fire point. The complexity weight mapping table represents a pre-established database of correspondences that maps different ranges of environmental complexity indices to corresponding weight coefficients. It is used to dynamically adjust the relative importance of traditional methods and deep learning methods in the fusion decision. The adaptive weight refers to the weight coefficients that are automatically adjusted according to specific environmental conditions. It includes a first weight (used for traditional feature discrimination results) and a second weight (used for deep learning model identification results). The sum of these two weights is usually 1, reflecting the relative reliability of the two methods in different environments.

[0081] After obtaining the classification results of feature discrimination and the identification confidence of the deep learning model, the device first needs to assess the environmental complexity of the area where each suspected fire point is located. The device extracts environmental information related to the area from multi-source databases, including digital elevation model (DEM) data (used to calculate topographic relief, slope, aspect, etc.) and land cover classification data (used to assess the diversity and heterogeneity of land surface types).

[0082] The equipment extracts land cover type distribution maps of the region from a land cover database, including the spatial distribution and area proportions of different types such as forests, grasslands, farmland, urban areas, and water bodies. It calculates land cover type diversity indices (e.g., the Shannon diversity index) and fragmentation indices (e.g., the ratio of boundary length to area). These indices reflect the degree of heterogeneity of land cover; higher heterogeneity generally indicates a more complex background temperature distribution, making fire point identification more difficult. Next, the equipment retrieves historical thermal anomaly records for the region over a past period (e.g., the last 3 years) from a historical thermal anomaly database, analyzing the spatiotemporal distribution characteristics of thermal anomaly events. The equipment calculates the density (number of events per unit area), periodicity (whether seasonal or annual patterns exist), and clustering (whether the spatial distribution of thermal anomalies is random or clustered) of thermal anomaly events. These characteristics reflect the history of thermal anomaly activity in the region and help assess whether current suspected fire points conform to the typical thermal anomaly patterns of the region. Finally, the equipment extracts the region's topographic features from a digital elevation model (DEM) database, including average elevation, slope distribution, and aspect distribution. The device calculates terrain complexity indices (such as terrain undulation or terrain roughness), which reflect the degree of terrain variation. More complex terrain typically indicates more variable microclimate and radiation conditions, making fire point identification more difficult. Finally, the device calculates an environmental complexity index by weighted fusion of these multi-dimensional environmental characteristic indices. This index can be a weighted average of the individual indices or a combination based on a specific function (such as geometric mean or harmonic mean). The device retrieves the first weight (used for traditional feature discrimination results) and the second weight (used for deep learning model discrimination results) by querying a pre-defined complexity weight mapping table.

[0083] The structure of the mapping table can be a discrete interval mapping (such as dividing the complexity index into several intervals, with each interval corresponding to a fixed weight value) or a continuous function mapping (such as continuously mapping the complexity index to weight values ​​through mathematical functions), without any limitation here.

[0084] In some embodiments, the environmental complexity index can be calculated in a variety of ways:

[0085] Optionally, the device constructs a spatiotemporal correlation network based on historical thermal anomaly events in the historical thermal anomaly database as nodes. This network includes establishing a directed edge between corresponding nodes when the spatial distance between two thermal anomaly events is less than a preset correlation distance threshold and the time interval is less than a preset correlation time window. The weight of the edge is determined by the inverse of the spatial distance and the decay function of the time interval. The device executes a random walk algorithm in the spatiotemporal correlation network to statistically obtain the number of paths and the average path length reaching the preset neighborhood of the suspected fire point. The number of paths represents the activity level of the thermal anomaly, and the average path length represents the spatial clustering characteristic of the thermal anomaly. Based on the land cover type patches within the preset spatial range corresponding to the suspected fire point location, the device calculates a fragmentation index by weighting the ratio of the boundary length to the area of ​​the land cover type patches. The weight of the land cover type is obtained based on a preset land cover database. Based on the number of paths, the average path length, and the fragmentation index, the device determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean.

[0086] Optionally, based on the geographical location of suspected fire points, the device retrieves historical verified fire point data within a preset spatiotemporal range and statistically analyzes the distribution of the identification success rate of the historical verified fire point data within different time delay intervals. This time delay is the difference between the actual occurrence time of the fire point and the time of the algorithm's first identification. Based on the identification success rate distribution, the device obtains the regional identification difficulty coefficient by weighted integral of the identification success rate over time delay. The device then determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean based on the number of paths, average path length, regional identification difficulty coefficient, and fragmentation index. It is understood that other methods can also be used to calculate the environmental complexity index; this is not limited here.

[0087] Furthermore, it should be noted that the calculation of the environmental complexity index can take into account the influence of time factors. For example, different calculation parameters or weights can be set for different seasons and different time periods (such as day / night) to reflect the dynamic changes in environmental complexity over time. At the same time, the environmental complexity index can also be customized according to actual application needs; for example, in forest reserves, the impact of vegetation type may be more important, while in urban areas, the impact of human activities may be more important.

[0088] S209. The device weights the classification results based on the first weight and uses the second weight to weight the model identification confidence, and generates a fire point confirmation score through weighted summation.

[0089] After obtaining the adaptive weights, the device applies the first weight to the classification result and the second weight to the model's confidence score, calculating the fire point confirmation score through a weighted sum. The formula for the weighted sum might be: Fire Point Confirmation Score = First Weight × Classification Result + Second Weight × Model Confidence Score. The sum of the two weights is usually set to 1; a larger value indicates a higher probability of a true fire point.

[0090] S210, The equipment identifies suspected fire points with fire point confirmation scores higher than the historical preset threshold as real fire points.

[0091] The device compares the calculated fire point confirmation score with a historical preset threshold. This threshold is typically determined based on historical verification data and aims to balance the false alarm rate and the missed alarm rate. If the fire point confirmation score is higher than the historical preset threshold, the suspected fire point is ultimately identified as a real fire point, which may trigger corresponding alarms and notifications; if the score is lower than the threshold, the point is determined to be a non-fire point or continues to be monitored.

[0092] Furthermore, it should be noted that the historical preset threshold can be determined in several ways. Optionally, based on ROC curve analysis, by plotting true positive and false positive rate curves at different thresholds on the validation dataset, the threshold point that maximizes a certain performance metric (such as the Youden index or F1 score) can be selected. Optionally, based on cost-sensitivity analysis, considering the different costs of false negatives and false positives, the threshold that minimizes the total expected cost can be selected. Additionally, the threshold can be dynamically adjusted according to the specific needs of the application scenario. For example, a lower threshold may be preferred in high-value protection areas to reduce the risk of false negatives, while a higher threshold may be used in resource-constrained wide-area monitoring to control the number of false positives; this is not limited here.

[0093] In this embodiment, by employing an adaptive weight allocation technique based on environmental complexity, the technical problem of related technologies being unable to adapt to complex and ever-changing environments is effectively solved. This results in improved accuracy of fire point identification in complex environments, reduced false alarm rates, and enhanced system environmental adaptability. Furthermore, the adoption of a fire point confirmation mechanism based on multi-source information weighted fusion achieves more refined judgment results, a more transparent decision-making process, and improved system reliability.

[0094] The following describes an exemplary remote sensing monitoring device 300 provided in an embodiment of this application. Figure 3 This is an exemplary hardware structure diagram of the remote sensing monitoring device 300 provided in this application embodiment.

[0095] In some embodiments, the remote sensing monitoring device 300 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0096] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0097] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0098] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0099] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A fire point identification method based on geostationary satellites, characterized in that, The method includes: The device identifies suspected fire points that are pre-extracted from geostationary satellite remote sensing images as critical fire points if the difference between the brightness temperature of the suspected fire point and the corresponding background brightness temperature is lower than a preset brightness temperature difference threshold. The background brightness temperature is the average brightness temperature within a preset range of the area where the suspected fire point is located. If the device determines that the critical fire point has no historical associated data within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than a preset change threshold, the critical fire point is determined to be a high-confidence fire point. If the device determines that the critical fire point has historical associated data within the preset period, it identifies pixel units with brightness and temperature higher than a preset average threshold in a preset neighborhood space centered on the critical fire point, and combines the pixel units into a high-temperature region. The device calculates the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter based on the current morphological measurement parameter of the high-temperature region at the current moment and the historical morphological measurement parameter of the critical fire point within the preset period before the current moment. The morphological measurement parameter includes the area of ​​the high-temperature region and the compactness calculated from the area and the perimeter of the boundary of the high-temperature region. If the device determines that the coefficient of variation of the time series is not less than a preset time series change threshold, the critical fire point is a high-confidence fire point.

2. The method according to claim 1, characterized in that, The method further includes: The device determines the classification result corresponding to the suspected fire point; The device inputs the multispectral data corresponding to the suspected fire point in the geostationary satellite remote sensing image into a pre-trained fire point recognition model to obtain the model identification confidence level that the suspected fire point is a real fire point; The device obtains an adaptive weight based on the environmental complexity index of the area where the suspected fire point is located by querying a preset complexity weight mapping table. The environmental complexity index is determined by comprehensive quantification based on land cover type data and historical hot anomaly database. The adaptive weight includes a first weight and a second weight, and the higher the environmental complexity index, the greater the ratio of the second weight to the first weight. The device weights the classification results based on the first weight and the model identification confidence based on the second weight, and generates a fire point confirmation score by weighting and calculating the weighted sum. The device identifies suspected fire points whose fire point confirmation scores are higher than a historical preset threshold as actual fire points.

3. The method according to claim 2, characterized in that, The step of determining the classification result corresponding to the suspected fire point by the device specifically includes: If the device determines that the critical fire point has no historical associated data within a preset period, and the difference between the brightness temperature change rate of the critical fire point and the background brightness temperature change rate of the critical fire point is greater than a preset change threshold, the device determines the numerical value of the classification result based on the magnitude of the difference exceeding the preset change threshold. The greater the magnitude of the difference in temperature change rate, the greater the numerical value of the classification result. If the device determines that the critical fire point has historical associated data within the preset period, and the time series variation coefficient is not less than the preset time series change threshold, the device determines the numerical value of the classification result based on the magnitude of the time series variation coefficient exceeding the preset time series change threshold. The greater the magnitude of the time series variation coefficient exceeding the preset time series change threshold, the greater the numerical value of the classification result. If the device determines that the temperature change rate difference of the critical fire point does not reach the preset change threshold, or the time series variation coefficient is less than the preset time series change threshold, the classification result will not be included in the calculation.

4. The method according to claim 2, characterized in that, The environmental complexity index is generated by comprehensively quantifying land cover type data and historical thermal anomaly databases, specifically including the following steps: The device constructs a spatiotemporal association network based on historical hot anomaly events in the historical hot anomaly database as nodes. The spatiotemporal association network includes establishing a directed edge between corresponding nodes when the spatial distance between two hot anomaly events is less than a preset association distance threshold and the time interval is less than a preset association time window. The weight of the edge is determined by the reciprocal of the spatial distance and the decay function of the time interval. The device executes a random walk algorithm in the spatiotemporal correlation network to statistically obtain the number of paths and the average path length within a preset neighborhood of the suspected fire point. The number of paths represents the activity level of the thermal anomaly, and the average path length represents the spatial clustering characteristics of the thermal anomaly. The device calculates a fragmentation index based on the weighted average of the ratio of the boundary length to the area of ​​the land cover type patches within a preset spatial range corresponding to the suspected fire point location. The weight of the land cover type is obtained based on a preset land cover database. The device determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean based on the number of paths, the average path length, and the fragmentation index.

5. The method according to claim 4, characterized in that, The step of determining the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean based on the number of paths, the average path length, and the fragmentation index specifically includes: Based on the geographical location of the suspected fire point, the device retrieves historical verified fire point data within a preset time and space range, and statistically analyzes the recognition success rate distribution of the historical verified fire point data in different time delay intervals. The time delay is the difference between the actual occurrence time of the fire point and the first recognition time of the algorithm. The device obtains the regional recognition difficulty coefficient by weighting the recognition success rate with time delay based on the recognition success rate distribution. The device determines the environmental complexity index by the ratio of the weighted geometric mean to the arithmetic mean based on the number of paths, the average path length, the region identification difficulty coefficient, and the fragmentation index.

6. The method according to claim 1, characterized in that, After determining that historical correlation data exists for the critical fire point within the preset period, identifying pixel units with brightness temperatures higher than a preset average threshold within a preset neighborhood space centered on the critical fire point, and combining the pixel units into a high-temperature region, the method further includes: The device performs spectral analysis on the brightness temperature time series of the critical fire point to obtain periodic and non-periodic components. The brightness temperature time series is based on the brightness temperature at a consecutive preset number of observation times. The device calculates the proportion of the aperiodic component in the time series and extracts the pulse features from the aperiodic component, the pulse features including pulse amplitude and pulse interval; The device determines anomaly score based on the statistical distribution of the pulse characteristics. The greater the variability of the pulse amplitude and the weaker the regularity of the pulse interval, the higher the anomaly score. When the anomaly score exceeds a preset anomaly threshold and the proportion exceeds a preset energy threshold, the device performs a time window extension on the preset period, wherein the time window extension is to extend the preset period to a preset multiple of the original period.

7. The method according to claim 1, characterized in that, The step of calculating the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter based on the current morphological measurement parameter of the high-temperature region at the current moment and the historical morphological measurement parameter of the critical fire point within the preset period before the current moment specifically includes: After constructing a morphological evolution trajectory sequence of the historical morphological measurement parameters in chronological order, the device divides the morphological evolution trajectory sequence into multiple overlapping sub-sequence segments through a sliding window. Each sub-sequence segment contains a consecutive preset number of historical morphological measurement parameters. The device obtains the local coefficient of variation of the subsequence segment by calculating the ratio of the standard deviation to the mean of the morphological metric parameters within the subsequence segment; The device extracts a morphological stability benchmark value based on the distribution characteristics of the local coefficient of variation; The device extracts the nearest local variation coefficient corresponding to the last sub-sequence segment in the multiple sub-sequence segments sorted by time, and determines the ratio of the nearest local variation coefficient to the morphological stability benchmark value as the time series variation coefficient of the current morphological measurement parameter relative to the historical morphological measurement parameter.

8. A device, characterized in that, The device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the device to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on the device, the device causes the device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the device, the device causes the device to perform the method as described in any one of claims 1-7.

Images