A time-space multi-modal based satellite fire point identification method for kuaishahao No. 9

Abstract

This invention discloses a method for identifying fire points on the Himawari-9 satellite based on spatiotemporal multimodal data, relating to the field of Himawari-9 satellite fire point identification technology. The method includes the following steps: S1, excluding non-fire point pixels based on the physical characteristics of the Himawari-9 satellite and publicly available data. Non-fire point pixels include cloud pixels, water body pixels, flare pixels, urban built-up area pixels, and fixed heat source pixels; S2, performing spatial context potential fire point detection on the image processed in S1, first filtering suspected fire point pixels using a contextual physical threshold method. This invention utilizes a multi-level filtering framework, a thermal infrared-terrain-meteorological multimodal feature fusion strategy, and a combination of Otsu adaptive thresholding and random forest algorithms with spatiotemporal dual-dimensional identification. This solves the problems of poor generalization, insufficient sensitivity in identifying small fire points, high false alarm rate, and insufficient fusion of heterogeneous data in existing methods, achieving high-precision, high-sensitivity, and low-false-alarm efficient identification of Himawari-9 satellite fire points.

Description

Technical Field

[0001] This invention relates to the field of Himawari-9 satellite fire point identification technology, specifically a Himawari-9 satellite fire point identification method based on spatiotemporal multimodality. Background Technology

[0002] Forest fires and other disasters severely threaten the ecological environment and the safety of human life and property. Satellite remote sensing technology, with its advantages of wide coverage and high timeliness, has become a core means of fire point monitoring. As a geostationary satellite, Himawari-9 possesses high temporal resolution characteristics. The optimization of its fire point identification technology can theoretically enrich the satellite remote sensing fire point monitoring methodology and, in practice, provide precise support for disaster prevention and control, thus possessing significant research value.

[0003] In recent years, various algorithms and technologies have emerged in the field of satellite remote sensing fire detection. Chinese patent application CN116758403A discloses a method for constructing a BP neural network based on Himawari-8 satellite band data. While this method solves some problems in straw burning monitoring, its model structure is simple, its expressive power is limited, it relies on manually set multiple thresholds for feature extraction, lacks the ability to model spatial context and temporal series changes, and it does not fully integrate multimodal information, resulting in poor generalization. Another type of multimodal fire detection method based on deep learning such as UNet, while improving spatial feature extraction capabilities, focuses primarily on single-phase images, failing to effectively utilize the high temporal resolution advantage of the Himawari satellite for dynamic identification. It lacks sensitivity to small fires, and its complex models rely on large amounts of labeled data, limiting its real-time performance and adaptability. In related literature reviews, Cunhui Zhang et al. focused on the brightness and temperature changes in Himawari-8 satellite images to detect spatiotemporal fire spots, but did not disclose specific methods and conclusions, making it difficult to assess reliability. Xiaohong Zhu and Lin Sun proposed a short-term time series window method based on adaptive thresholds, which reduces dependence on historical data, but the adaptive threshold has limitations under different regional and environmental conditions. Regarding spectral analysis and thresholding methods, Afifah et al. used the Otsu thresholding method to detect photovoltaic module hotspots, achieving high accuracy but with a small sample size. Fitriana et al. detected fire hotspots in Sumatra using spectral analysis of Himawari-8 satellite data, but this method was limited to a specific area, and its generalizability needs to be verified.

[0004] Overall, existing technologies have significant shortcomings: they are not optimized for the characteristics of the Himawari-9 satellite sensors; multimodal fusion is limited to multi-band spectral information and lacks deep fusion of heterogeneous spatiotemporal data such as meteorological and topographical data; methodologically, they lack comprehensive research on the optimal fusion methods for various spatiotemporal sequence features and multi-source data; theoretically, they lack comparative analysis of different algorithms in complex scenarios; empirical studies show insufficient verification of model generalization ability and inadequate adaptability to different regions and environmental conditions, making it difficult to achieve high-precision and robust fire point identification in complex contexts. Therefore, this paper proposes a spatiotemporal multimodal fire point identification method based on the Himawari-9 satellite to overcome these problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodality, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides a method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodality, comprising the following steps: S1. Based on the physical characteristics of the Himawari-9 satellite and publicly available data, non-fire point pixels are excluded. Non-fire point pixels include cloud pixels, water body pixels, flare pixels, urban built-up area pixels, and fixed heat source pixels. S2. Perform spatial context potential fire detection on the image processed by S1. First, use the context physical thresholding method to filter suspected fire pixels, and then use the Otsu adaptive thresholding algorithm to perform secondary identification. S3. Based on the random forest algorithm, a multi-feature fire point identification model is constructed through feature engineering and a multi-modal feature fusion strategy. The model integrates thermal infrared, terrain, and meteorological multi-modal features to confirm the potential fire point pixels obtained in S2. S4. Perform the final determination of fire points based on the time series method. Generate regional adaptive thresholds through historical data and combine the pixel brightness temperature change rate and the continuous time change trend to complete the confirmation of real fire points.

[0007] Furthermore, in S1: Cloud pixels pass through the equation: The symbols are defined as follows: Apparent reflectance of the Himawari-9 satellite in the third band (near-infrared band, center wavelength approximately 1.6µm); Apparent reflectance of the Himawari-9 satellite in the 4th band (shortwave infrared band, center wavelength approximately 2.3µm); Brightness temperature of the Himawari-9 satellite in band 15 (thermal infrared band, center wavelength approximately 12.4µm), in Kelvin (K). Identification; Water pixels pass through the equation: Identify and apply water masking only to daytime pixels with a solar zenith angle ≤ 85°; Flare pixels pass through the system of equations: Identification; The pixels of the urban built-up area were identified based on MODIS 30m land classification data, and the pixels of the fixed heat source were identified based on the global remote sensing industrial heat source annual dynamic dataset from 2012 to 2021. The definitions of each are as follows: Apparent reflectance of the Himawari-9 satellite in the 6th band (visible band, center wavelength approximately 0.64µm); The geometric angle between the sun, the target, and the sensor, i.e., the flare angle; : Satellite observation of zenith angle; : Solar zenith angle; : Relative azimuth angle between the sun and the sensor; Apparent reflectance of the Himawari-9 satellite in the 12th band (near-infrared band, center wavelength approximately 1.38µm); Apparent reflectance of the Himawari-9 satellite in band 15 (thermal infrared band, center wavelength approximately 12.4µm).

[0008] Furthermore, in S2: Contextual physics thresholding is achieved through the equation: Filter pixels suspected of being fire points; The Otsu adaptive thresholding algorithm includes calculating the gray-level histogram, probability intensity values, class weights and means, and inter-class variance to determine the optimal threshold k*, and then applying the following equation: Perform binarization segmentation; The symbols are defined as follows: Pixel brightness temperature values ​​detected by the AHI sensor of the Himawari-9 satellite in band 11 (center wavelength 10.4μm); The brightness temperature difference between the mid-infrared band and the thermal infrared band, i.e. and The difference ( ), used to enhance the difference between fire signals and background features; Apparent reflectivity of the second band (center wavelength 0.86μm) of the AHI sensor of the Himawari-9 satellite, used to eliminate cloud interference; The coordinates in the fire detection image are: The pixel grayscale value or radiation intensity value; The optimal segmentation threshold calculated by the Otsu algorithm; The output result after binarization segmentation, where "1" indicates that the pixel is identified as a suspected fire point, and "0" indicates that the pixel is a background pixel.

[0009] Furthermore, the feature engineering of S3 includes: Twelve original features, including thermal infrared brightness temperature (Tbb07-Tbb16) and latitude and longitude, were extracted. The radiance of the original bands is calculated, and the radiance is combined with the luminance temperature bands to obtain 25 bands of calculated features. Five spatial features were extracted, including the mean absolute deviation of background pixel band values ​​and the difference between the pixel and the background mean.

[0010] Furthermore, S3's multimodal feature fusion strategy includes: Extracting terrain features: Obtaining elevation, slope, and aspect based on SRTM data; Extracting meteorological features: fusing ERA5 or GFS reanalysis data to obtain near-surface air temperature, relative humidity, wind speed, and wind direction; Z-score normalization is used to normalize features, and feature importance scores based on random forest are applied. Dynamically allocate fusion weights: .

[0011] Furthermore, S3 also includes a feature importance calculation step, specifically evaluating features by ranking their importance. Importance is defined as: ; in The F1 score is the model validation set before ranking. The F1 score is the result after sorting.

[0012] Furthermore, the region-adaptive threshold generation of S4 includes: Based on MODIS land cover data, the study area was divided into different underlying surface types, and fire-free brightness temperature data for the same season and time period in the past year were extracted. Calculate the brightness temperature change rate mean and standard deviation Construct confidence intervals ; Adjustments are made based on current meteorological conditions to generate an adaptive judgment threshold. ,in The value is between 2 and 3.

[0013] Furthermore, the brightness temperature change rate in S4 is calculated according to the equation: ; calculate, The difference in brightness between the two periods is the temperature difference between the two periods. For the observation time difference; when The confidence level of the fire point is enhanced, and the final identification is completed by combining the changing trends of multiple consecutive time periods.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. A rigorous multi-level filtering framework was constructed, greatly reducing the false alarm rate: This invention adopts a four-step identification process (S1-S4) from coarse to fine, sequentially excluding non-fire pixels, detecting potential fire points, confirming through machine learning, and verifying through temporal changes. This progressive, multi-verification framework design systematically eliminates major interference factors such as clouds, water bodies, flares, and urban heat sources, ensuring the reliability of the results from the root and significantly reducing the false alarm phenomenon common in traditional methods.

[0015] 2. Solid Scientific Foundation Through Integration of Multi-Source Data and Physical Mechanisms: This invention does not rely on a single data source or algorithm, but rather deeply integrates satellite physical characteristics, publicly available data from multiple sources (global industrial heat source dataset, MODIS land classification data), and heat conduction physical models. This design based on multi-source data and physical laws ensures that the fire point identification process is evidence-based, and the conclusions are scientifically reliable, surpassing methods that rely solely on statistics or thresholds.

[0016] 3. Introducing advanced machine learning and image processing algorithms to improve recognition accuracy and automation: This invention introduces the Otsu adaptive thresholding algorithm (S2.2) and the random forest model (S3) into the field of fire detection. The Otsu algorithm can automatically find the optimal segmentation threshold based on the image histogram, adapting to different scenarios; the random forest model, through multi-feature training and ensemble learning, effectively solves the problem of poor adaptability of traditional thresholding methods in complex scenarios, significantly improving the model's generalization ability and recognition accuracy.

[0017] 4. Combining spatiotemporal dual-dimensional information to achieve a balance between high sensitivity and high reliability: This invention not only utilizes the spatial context information of fire point pixels (differences from background pixels) but also introduces time series analysis (S4). By analyzing the brightness temperature change rate in 10-minute high-frequency observation data, it can effectively distinguish between normal surface warming caused by solar radiation and abnormal heat release caused by actual fires, greatly improving the detection sensitivity for small and weak fires while avoiding misjudgments caused by changes in sunlight.

[0018] 5. Meticulous feature engineering design effectively supports high-performance machine learning models: In the feature engineering (S3.2) of this invention, not only are the original band values ​​considered, but also radiance and spatial background statistical features (mean absolute deviation, difference from the background mean, etc.) are calculated, and fire point count labels are constructed. This meticulous design ensures the richness and discriminativeness of the input features, providing a solid foundation for the high-precision classification of the random forest model.

[0019] 6. The model design and validation process fully considers the complexity of actual business and data imperfections: The method is designed to fully consider the complexity of real-world business. For example, to alleviate the problem of imbalanced training data, non-fire point samples are selected at a ratio of 1:10; to avoid the striping and positioning errors of other fire point products, the eight adjacent pixels of the fire point label are not selected. These details reflect the rigor of the method and ensure the practicality and robustness of the model in real-world scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 Example of a full-disk image of Himawari-9; Figure 3 This includes measured data and satellite imagery of fire events. Figure 4 This is a schematic diagram of the suspected fire point assessment results. The right side represents the spatial context physical threshold determination results, and the left side represents the Otsu assessment results based on the physical assessment results. Figure 5 Cross-correlation diagram of features in the initial random forest fire detection model; Figure 6A schematic diagram illustrating the importance of features in the initial model for determining fire points in a random forest. Figure 7 The cross-validation learning curve of the initial random forest fire detection model; Figure 8 ROC curve of the initial random forest fire detection model; Figure 9 This is a schematic diagram of the fire point assessment results of the time-series fire point assessment model, where (a) represents the final fire point assessment result of this invention, and (b) represents the NASA fire point production disk assessment result. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-9 The present invention provides a technical solution: See Figures 1-9 As shown, an embodiment of a method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodality is presented: S1. Elimination of non-fire point pixels based on the physical characteristics of the Himawari-9 satellite and publicly available data: S1.1 Cloud and Water Pixel Identification: Clouds are a key component of active fire detection systems. Cloud reflections in the MIR band may be falsely detected as fire pixels. Furthermore, solar flares originating from water bodies may also be falsely identified as fire pixels. Excluding cloud and water pixels can reduce false fire detections caused by their presence. In this patent, we reference the method of Xu et al. To detect clouds and water in Himawari-9 images, Himawari-9 image data is as follows: Figure 2 As shown. Pixels that satisfy the following equation are defined as cloud pixels.

[0023] The daytime cloud pixels are shown in equation (1): The nighttime cloud pixels are shown in equation (2): in, and These are the reflectances of band 3 and band 4 of the Himawari-9, respectively. This is the brightness temperature of band 15. Pixels that satisfy equation (3) are defined as water pixels: in This is the albedo of the Himawari-9 Band 6. Water masking is not required at night because the false positives from water are lower. We define pixels with a solar zenith angle greater than 85° as night pixels.

[0024] S1.2 Solar Flare Pixel Identification: Differences in brightness temperature may exist at the boundary between water and land, leading to solar flares. Additionally, the high reflectivity of the land surface to sunlight in acquired remote sensing images can also cause flares. Both phenomena can produce false fire points, interfering with correct fire points. Therefore, it is necessary to eliminate misjudgments caused by solar flares to improve identification accuracy.

[0025] Flare identification is shown in equation set (4): In the formula: Indicates the zenith angle observed by the satellite. Indicates the zenith angle of the sun. Indicates relative azimuth. It is the angle between the ground and the direction of satellite and specular reflection. If all three conditions are met, it is marked as a flare pixel.

[0026] S1.3 Identification of urban built-up areas and fixed heat sources: This invention uses the experimental results provided by the "annual dynamics of global remote industrial heat sources dataset from 2012 to 2021" as the identification condition for fixed heat sources, and uses the 30m land classification data based on MODIS provided by NASA as the identification condition for built-up areas.

[0027] Based on step S1 above, non-fire pixels are excluded, greatly reducing false positives for fire points. This is then used as input for step S2.

[0028] S2, Spatial Context Potential Fire Detection: Context-based fire detection using physical thresholds is the most commonly used monitoring algorithm and is widely applied in the automated identification of MODIS data. This method evaluates the brightness temperature of background pixels centered on potential hotspot pixels. Once a potential hotspot pixel meets the data filtering criteria, it can be identified as a valid hotspot pixel. Based on this, this invention introduces the Otsu method for secondary identification of valid hotspot pixels to ensure accuracy and reduce false identification of potential fire points. Figure 4 As shown, the right side represents the spatial context physical threshold determination result, and the left side represents the Otsu judgment result based on the physical judgment result.

[0029] S2.1 Contextual physical threshold fire point identification: Pixels in the image that simultaneously satisfy equations (5) to (8) are identified as suspected fire point pixels.

[0030] in, , These represent the brightness temperature values ​​for channels 7 and 11, respectively. This represents the reflectivity of the 4th channel.

[0031] S2.2, Otsu's secondary identification of suspected fire points: Otsu is applied to automatically obtain threshold values. The concept of Otsu is a discriminant analysis method designed to identify variables that can classify data into certain categories.

[0032] Step 1: Calculate the histogram values: Before performing the Otsu method, a histogram of the grayscale image needs to be calculated. A histogram is a graphical representation of the intensity distribution in an indexed grayscale image. It shows the frequency of each pixel value in the image. An ideal threshold can be derived by verifying the pixel values ​​in the histogram by checking the highest-level inter-class variance (BCV) between background and foreground pixels.

[0033] Step 2: Calculate the probability strength value: The pixel values ​​of grayscale image data can be represented by the grayscale range. To represent. Each pixel in The potential of the stage is shown in formula (9), where Represents pixel count Pixel intensity is used express, As shown in formula (9).

[0034] Step 3: Calculate the frequency of category occurrences and the category mean level: Divide the pixels into two groups: As background, As foreground, through a horizontal The threshold at each point is used for division; The displayed pixel value is... The pixels within the range, and What is displayed is the pixel value. The pixels within the range. Next, the probability of each category occurring is calculated in the system of equations (10), and the average level of each category is calculated in the system of equations (10).

[0035] Step 4: Calculate the between-class variance: Between-class variance The value of otsu is determined; Step 5: Find the maximum value in the variance: Finally, the optimal threshold is shown in equation (12). This can be expressed as follows.

[0036] The thresholding operation is shown in (13). If the grayscale image is binary image The definition in (13) is as follows.

[0037] Where th is derived from the optimal threshold. The threshold value is used for image elements in the foreground or hotspot regions. As for image elements in the background area, Accurately selecting the threshold is crucial for profitable threshold segmentation.

[0038] Based on step S2 above, after excluding non-fire point pixels, potential fire point pixels are identified in the image according to the physical characteristics of the fire point pixels. This ensures the authenticity and objectivity of potential fire points. The potential fire point identification results are as follows: Figure 3 As shown, it is used as the input for step S3.

[0039] S3. Fire detection based on random forest: S3.1 Random Forest Algorithm: Randomization (RF) employs two randomization strategies: randomly selecting training samples for each tree and randomly selecting input variables for each node of the tree. By developing many independent trees with different sets of training samples and input variables, RF attempts to provide relatively unbiased results, preventing model overfitting and sensitivity to the configuration of training data. Assume there is a set of trees... The training set consists of training samples, each sample having _ ... One feature. Random forest will be composed of It consists of several decision trees or estimators. The specific classification process is as follows: 1) Use the Bootstrap autosampling method to randomly sample data with replacement from... Extracted from a sample dataset Each sample constitutes a single training set, repeated... Next, constitute Includes The training set of samples.

[0040] 2) For each containing The training set of n samples is randomly selected to be of size n. To train each estimator. Created A decision tree.

[0041] 3) Make predictions for newly input samples. The relevant features of the sample are passed to each decision tree to obtain the prediction result for each decision tree. Finally, a majority vote is used to determine the predicted class.

[0042] S3.2 Feature Engineering and Feature Selection: S3.2.1 Feature Engineering: In this invention, training samples for the Himawari-9 detection model are constructed using multi-source fire point products. The fire point counting concept involves first considering the Himawari-9 element as composed of 5×5 small grids of 0.004×0.004°. If the center of a fire point falls within a certain grid, that grid is marked. Finally, the total number of marked grids in the Himawari-9 element is taken as the fire point count. Simultaneously, non-fire point tags are randomly selected at a ratio of 1:10. To reduce the influence of striping phenomena and positioning errors in the fire point products, the eight adjacent pixels of each fire point tag are not selected.

[0043] Based on the initial Himawari-9 data, a total of 12 features were selected, including thermal infrared brightness temperature (Tbb07-Tbb16) and latitude and longitude data. To achieve consistent retrieval of fire point pixel features, solar zenith angle data strongly correlated with time were excluded. Radiance (Rad) was calculated for the original bands, and the radiance and brightness temperature bands were combined to obtain 25 calculated features. Furthermore, the spatial features of the pixels reflect the differences between fire point pixels and background pixels. In this invention, five spatial features were extracted using the MODIS thermal anomaly algorithm: the mean absolute deviation of background pixel band values, the difference between the mean values ​​of the pixel band values ​​and the mean values ​​of the background pixel band values, and the cross-correlation results of the features are shown below. Figure 5 As shown.

[0044] S3.2.2 Feature Selection: The ranking-based feature importance method is a way to directly measure the impact of each feature on model accuracy. The main idea is to shuffle the order of each feature value and assess feature importance by measuring the impact of this order change on model accuracy. The more severe the decrease in model accuracy after shuffling, the greater the importance of that feature in classification. Now, assume we have features... Then features The ranking importance in the random forest model is defined as shown in equation (14).

[0045] and These are the accuracy metrics for the model on the validation set before and after ranking. Considering the imbalance of the dataset, the F1-score is chosen as the accuracy metric. The feature importance measurement results are as follows: Figure 6 As shown.

[0046] The potential fire point pixels obtained in the initial judgment are fed into the multi-feature fire point recognition and judgment model constructed in step S3 to confirm the potential fire points, and non-potential fire point pixels are further eliminated.

[0047] S3.2.3 Multimodal Feature Fusion Strategy: To further improve the model's generalization ability and recognition accuracy in complex environments, this invention introduces multimodal data such as terrain and meteorology into feature engineering and designs a fusion strategy based on weighted feature fusion and attention mechanism.

[0048] 1) Multimodal feature extraction: Thermal infrared characteristics: as mentioned above, including raw brightness temperature, radiance, spatial context features, etc.

[0049] Topographic features: Based on SRTM data, topographic factors such as elevation, slope, and aspect of pixels are extracted.

[0050] Meteorological characteristics: By integrating ERA5 or GFS reanalysis data, meteorological variables such as near-surface temperature, relative humidity, wind speed, and wind direction corresponding to the pixels are extracted.

[0051] 2) Feature normalization and alignment: All features were uniformly resampled to the Himawari-9 pixel scale.

[0052] The Z-score standardization method is used to normalize various features and eliminate the influence of dimensions.

[0053] 3) Weighting strategy: The fusion weights are dynamically allocated using a feature importance feedback mechanism based on random forest. First, the three types of features—thermal infrared, topographic, and meteorological—are input into the random forest model for initial training. The learning curve of the random forest model is shown below. Figure 7 As shown, the ROC curve is as follows Figure 8 As shown.

[0054] Calculate the overall importance score of each type of feature in the model. .

[0055] Normalization yields the fusion weights: S4. Fire point determination based on timing method: S4.1, Regional adaptive threshold generation based on historical statistics: To further improve the adaptability of fire point identification and reduce misjudgments caused by differences in surface type, season, and meteorological conditions, this invention introduces a regional adaptive threshold calculation module based on historical data. This module statistically analyzes satellite observation data from historical fire-free periods to construct benchmark intervals for brightness temperature change rates of different underlying surface types within typical time periods, and dynamically generates a judgment threshold applicable to the current pixel.

[0056] The specific steps are as follows: Historical data partitioning and classification: The study area was divided into different underlying surface types (such as forest, farmland, grassland, urban, bare soil, etc.) using MODIS land cover data (MCD12Q1).

[0057] For each type of underlying surface, extract the Himawari-9 brightness temperature data (mainly Band7) of the same season and time period (e.g., 10:00-14:00) in the past year (or many years) without fires.

[0058] Statistical modeling of rate of change: For each type of underlying surface, calculate its brightness temperature change rate ΔA over consecutive observation periods (every 10 minutes) (the calculation method is the same as Formula 15).

[0059] Analyze the distribution characteristics of its rate of change and calculate the mean. and standard deviation Construct a confidence interval for the rate of change of the underlying surface under a steady state, such as: This range covers approximately 95% of normal variations.

[0060] Dynamic threshold generation: For the current pixel to be identified, the corresponding historical change rate confidence interval is called according to its land cover type.

[0061] Fine-tuning can be made based on current meteorological conditions (such as cloud cover and atmospheric humidity), for example, the upper limit of the threshold can be appropriately relaxed under hot and dry weather conditions.

[0062] The final adaptive decision threshold for this pixel is generated. in This is an adjustable parameter (2-3 is recommended) used to control the balance between sensitivity and robustness.

[0063] Enhanced fire timing determination: In the subsequent S4 step, the calculated current pixel brightness temperature change rate ΔA is compared with the adaptive threshold of that pixel: like If so, it is determined to be an abnormal heat release, thus increasing the confidence level of the fire point.

[0064] At the same time, it can make a comprehensive judgment by combining the changing trends of multiple consecutive time periods, and further improve the ability to identify low heat and weak heat.

[0065] S4.2 Timing-based fire detection: The time series method mainly uses the brightness temperature change rate of the identified pixel in the time series to determine whether there is a fire in the pixel, as shown in equation (15).

[0066] In the formula, It is the brightness temperature change rate. It is the difference in brightness temperature between the previous and subsequent times. It is the time difference between the observations. By analyzing the rate of change and comparing it with the rate of change of the normal background, information on whether there are other released heat sources within the identified pixel can be obtained. Currently, geostationary meteorological satellites can acquire an observation every 10 minutes, while the surface temperature changes little within this time period, generally less than 0.5K. By using the brightness temperature change threshold of the same pixel before and after the observations, the sensitivity of satellite remote sensing fire point identification can be improved. Under the condition that there is no fire, the surface heating energy mainly comes from the sun, according to the thermal conductivity equation (16).

[0067] In the formula, It is thermodynamic temperature. It is time. It is the thermal conductivity, which is related to the thermal conductivity, density, specific heat, or heat capacity of the material. The change in surface temperature is driven by periodic solar radiation, which provides periodic heat flux. The heating effect of solar radiation on the surface is shown in equation (17).

[0068] In the formula, It is the effective longwave radiation temperature of the sky. It is the incident solar radiation that reaches the ground through the atmosphere. It is the thermal conductivity. It is the distance of penetration into the ground surface. This refers to time. Items 1 and 3 represent the radiative flux incident from the sky and the sun, respectively, while item 2 represents the emitted radiative flux emitted from the Earth's surface. Solar radiation. It is a function of the surface reflectance, solar declination, ecliptic latitude and local slope of the solar spectral region (most of which is visible light and near infrared), as shown in equation (18).

[0069] In the formula, It is the surface reflectance. It is the solar constant. It is a factor that reduces solar radiation from clouds.

[0070] In the formula, It is the local zenith angle of an inclined surface. It is the zenith corner. It's atmospheric attenuation; it's the zenith angle. The function. Time. It is related to the sun's angle. For a clear daytime atmosphere, under ideal conditions, surface temperature changes are mainly caused by solar radiation. The formula for the effective solar radiation reaching the surface and being absorbed can be converted to... In the formula, Atmospheric transmittance, For surface absorption rate, This is the solar altitude angle. For a fixed region at a specific time, assuming... , Assuming δ is constant and the underlying surface type remains unchanged, the energy absorbed by the Earth's surface depends only on the solar altitude angle. As the solar altitude angle increases, the absorbed energy also gradually increases; that is, the energy absorbed by the Earth's surface is directly proportional to the trigonometric function of the solar altitude angle. By comparing the rate of temperature change before and after different time periods with the background rate of change in a non-fired state, it can be determined whether there is a fire during the warming (or cooling) process of the identified pixel.

[0071] Finally, the potential fire point pixels identified in S3 are fed into the temporal method fire point pixel final identification model in step S4 to confirm the final real fire points.

[0072] Summarize: A rigorous multi-level filtering framework was constructed, which greatly reduced the false alarm rate: This invention adopts a four-step identification process (S1-S4) from coarse to fine, successively excluding non-fire pixels, detecting potential fire points, confirming through machine learning, and verifying through temporal changes. This progressive and multi-verification framework design systematically eliminates major interference factors such as clouds, water bodies, flares, and urban heat sources, ensuring the reliability of the results from the root and significantly reducing the false alarm phenomenon common in traditional methods.

[0073] Integrating multi-source data with physical mechanisms, this invention boasts a solid scientific foundation: It does not rely on a single data source or algorithm, but rather deeply integrates satellite physical characteristics (such as brightness temperature and reflectivity), publicly available multi-source data (global industrial heat source datasets, MODIS land classification data), and heat conduction physical models. This design, based on multi-source data and physical laws, ensures that the fire point identification process is evidence-based, and the conclusions are scientifically reliable, surpassing methods that rely solely on statistics or thresholds.

[0074] This invention introduces advanced machine learning and image processing algorithms to improve recognition accuracy and automation: It introduces the Otsu adaptive thresholding algorithm (S2.2) and the random forest model (S3) into the field of fire detection. The Otsu algorithm can automatically find the optimal segmentation threshold based on the image histogram, adapting to different scenarios; the random forest model, through multi-feature training and ensemble learning, effectively solves the problem of poor adaptability of traditional thresholding methods in complex scenarios, significantly improving the model's generalization ability and recognition accuracy.

[0075] By combining spatiotemporal information, a balance between high sensitivity and high reliability is achieved: This invention not only utilizes the spatial context information of fire point pixels (differences from background pixels) but also introduces time series analysis (S4). By analyzing the brightness temperature change rate in 10-minute high-frequency observation data, it can effectively distinguish between normal surface warming caused by solar radiation and abnormal heat release caused by actual fires, greatly improving the detection sensitivity for small and weak fires while avoiding misjudgments caused by changes in solar radiation.

[0076] The meticulous feature engineering design effectively supports the high performance of the machine learning model: In the feature engineering (S3.2) of this invention, not only are the original band values ​​considered, but also the radiance and spatial background statistical features (mean absolute deviation, difference from the background mean, etc.) are calculated, and a "fire point count" label is constructed. This meticulous design ensures the richness and discriminative power of the input features, providing a solid foundation for the high-precision classification of the random forest model.

[0077] The model design and validation process fully considers the complexity of real-world business and data imperfections: the method is designed to take into full account the complexity of real-world business. For example, to alleviate the problem of imbalanced training data, non-fire point samples are selected at a ratio of 1:10; to avoid the striping and positioning errors of other fire point products, the eight adjacent pixels of the fire point label are not selected. These details reflect the rigor of the method and ensure the practicality and robustness of the model in real-world scenarios.

Claims

1. A time-space multi-modal based satellite fire point identification method for sunflower No. 9, characterized by, Includes the following steps: S1. Based on the physical characteristics of the Himawari-9 satellite and publicly available data, non-fire point pixels are excluded. Non-fire point pixels include cloud pixels, water body pixels, flare pixels, urban built-up area pixels, and fixed heat source pixels. S2. Perform spatial context potential fire detection on the image processed by S1. First, use the context physical thresholding method to filter suspected fire pixels, and then use the Otsu adaptive thresholding algorithm to perform secondary identification. S3. Based on the random forest algorithm, a multi-feature fire point identification model is constructed through feature engineering and a multi-modal feature fusion strategy. The model integrates thermal infrared, terrain, and meteorological multi-modal features to confirm the potential fire point pixels obtained in S2. S4. Perform the final determination of fire points based on the time series method. Generate regional adaptive thresholds through historical data and combine the pixel brightness temperature change rate and the continuous time change trend to complete the confirmation of real fire points.

2. The sun-synchronous multi-modal based fire spot identification method for sun-synchronous satellites according to claim 1, wherein, In S1: Cloud pixels pass through the equation: The symbols are defined as follows: : apparent reflectance of the 3rd band of the sun-synchronous satellite Landsat 9; Apparent reflectivity of the 4th band of the Himawari-9 satellite; Brightness temperature of the 15th band of the Sunflower-9 satellite; Identification; Water pixels pass through the equation: Identify and apply water masking only to daytime pixels with a solar zenith angle ≤ 85°; Flare pixels pass through the system of equations: Identification; The pixels of the urban built-up area were identified based on MODIS 30m land classification data, and the pixels of the fixed heat source were identified based on the global remote sensing industrial heat source annual dynamic dataset from 2012 to 2021. The definitions of each are as follows: Apparent reflectivity of the 6th band of the Himawari-9 satellite; The geometric angle between the sun, the target, and the sensor, i.e., the flare angle; : Satellite observation of zenith angle; : Solar zenith angle; : Relative azimuth angle between the sun and the sensor; Apparent reflectivity of the 12th band of the Himawari-9 satellite; Apparent reflectivity of the 15th band of the Himawari-9 satellite.

3. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodal analysis as described in claim 1, characterized in that, In S2: Contextual physics thresholding is achieved through the equation: Filter pixels suspected of being fire points; Otsu's adaptive thresholding algorithm includes calculating the gray-level histogram, probability intensity values, class weights and means, and inter-class variance to determine the optimal threshold k*, and then applying the equation... Perform binarization segmentation; The symbols are defined as follows: Pixel brightness temperature values ​​detected by the AHI sensor of the Himawari-9 satellite in band 11; The brightness temperature difference between the mid-infrared band and the thermal infrared band, i.e. and The difference is used to enhance the distinction between the fire signal and background features; Apparent reflectivity of the second band of the AHI sensor on the Himawari-9 satellite, used to eliminate cloud interference; The coordinates in the fire detection image are: The pixel grayscale value or radiation intensity value; The optimal segmentation threshold calculated by the Otsu algorithm; : The output result after binarization segmentation, where 1 indicates that the pixel is judged as a suspected fire point, and 0 indicates that the pixel is a background pixel.

4. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodal analysis as described in claim 1, characterized in that, S3's feature engineering includes: Twelve original features were extracted, including thermal infrared brightness temperature (Tbb07-Tbb16) and latitude / longitude. The radiance of the original bands is calculated, and the radiance is combined with the luminance temperature bands to obtain 25 bands of calculated features. Five spatial features were extracted, including the mean absolute deviation of background pixel band values ​​and the difference between the pixel and the background mean.

5. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodal analysis as described in claim 1, characterized in that, S3's multimodal feature fusion strategy includes: Extracting terrain features: Obtaining elevation, slope, and aspect based on SRTM data; Extracting meteorological features: fusing ERA5 or GFS reanalysis data to obtain near-surface air temperature, relative humidity, wind speed, and wind direction; Z-score normalization is used to normalize features, and feature importance scores based on random forest are applied. Dynamically allocate fusion weights: 。 6. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodal analysis as described in claim 1, characterized in that, S3 also includes a feature importance calculation step, specifically evaluating features by ranking their importance. Importance is defined as: ； in The F1 score is the model validation set before ranking. The F1 score is the result after sorting.

7. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodal analysis as described in claim 1, characterized in that, S4's region-adaptive threshold generation includes: Based on MODIS land cover data, the study area was divided into different underlying surface types, and fire-free brightness temperature data for the same season and time period in the past year were extracted. Calculate the brightness temperature change rate mean and standard deviation Construct confidence intervals ; Adjustments are made based on current meteorological conditions to generate an adaptive judgment threshold. ,in The value is between 2 and 3.

8. The method for identifying the fire point of the Himawari-9 satellite based on spatiotemporal multimodality as described in claim 1, characterized in that: The brightness temperature change rate in S4 is calculated according to the equation: ； calculate, The difference in brightness between the two periods is the temperature difference between the two periods. For the observation time difference; when The confidence level of the fire point is enhanced, and the final identification is completed by combining the changing trends of multiple consecutive time periods.

Images