A satellite fire point detection method based on dynamic index features and deep learning
By constructing a fire detection method based on dynamic index features and deep learning, and utilizing multi-threshold screening and a fire spatiotemporal network model, the accuracy and generalization problems of fire detection in geostationary satellite data are solved, achieving efficient and accurate fire detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU UNIV OF INFORMATION TECH
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-19
Smart Images
Figure CN122244715A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a satellite fire detection method based on dynamic exponential features and deep learning. Background Technology
[0002] The multidimensional destructiveness of forest fires makes the development of efficient and accurate fire detection technologies an urgent need. The development of fire detection technology has progressed from physical thresholding methods to machine learning models. Early thresholding methods based on the thermal infrared band identified fire points by setting fixed brightness temperature thresholds. While this laid a physical foundation, it ignored the spatial heterogeneity of the environmental background, resulting in a high false alarm rate. Subsequent context-based thresholding methods improved the robustness of the models to some extent through local environmental modeling and dynamic threshold adjustment. However, this method still heavily relies on manual prior knowledge to set parameters, has limited generalization ability, and is difficult to adapt to the high-frequency observation requirements of geostationary meteorological satellites.
[0003] With the development of machine learning technology, data-driven strategies have significantly improved fire detection performance. Traditional machine learning methods such as support vector machines and random forests have made significant progress in feature engineering and classification decisions; while deep learning models such as convolutional neural networks and Transformers have further realized end-to-end deep feature extraction and spatiotemporal context modeling. However, existing deep learning-based methods mostly use raw multi-band satellite data directly as input, relying excessively on the model to autonomously learn features from low signal-to-noise ratio data. This presents inherent challenges when processing fire monitoring based on geostationary meteorological satellite data: First, the coarse spatial resolution of satellites leads to severe pixel mixing problems, making weak fire signals easily submerged by background noise; second, fire marker samples are scarce and the positive and negative samples are extremely imbalanced, greatly limiting the model's feature learning ability and generalization performance.
[0004] Therefore, in practical applications with low signal-to-noise ratios and small sample sizes, existing deep learning models still exhibit significant shortcomings in their ability to detect weak and small fire targets, their spatial positioning accuracy, and their generalization performance. The key to overcoming this technical bottleneck lies in two aspects: firstly, constructing more discriminative high-dimensional features to enhance the model's sensitivity to fire signals and its anti-interference capabilities; and secondly, designing a dedicated deep learning architecture adapted to the spatiotemporal characteristics of geostationary satellite data. This is the core direction for improving the accuracy, reliability, and operational applicability of fire detection. Summary of the Invention
[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a satellite fire detection method based on dynamic exponential features and deep learning, which solves the problems of insufficient detection accuracy, excessive reliance on models, and high degree of manual intervention in existing fire detection technologies.
[0006] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a satellite fire detection method based on dynamic exponential features and deep learning, comprising: Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites; Candidate pixels are obtained by filtering the full-disk satellite data using multiple threshold conditions. Spatiotemporal input features and channel input features of candidate pixels are extracted based on multi-band spatiotemporal observation data from geostationary meteorological satellites; The spatiotemporal input features and channel input features of candidate pixels are input into a pre-trained fire spatiotemporal network model to obtain fire point detection results.
[0007] Furthermore, multi-threshold conditions are used to filter pixels in the full satellite disk data to obtain candidate pixels, including: Fire index and temperature difference are obtained based on the mid-infrared temperature of pixels in the satellite full disk data and the surface temperature. A first threshold is set based on the fire point index and temperature difference; A second threshold is set based on the red light band, the near-infrared band, and the surface temperature. A third threshold is set based on the shortwave infrared band; Based on the first threshold, the second threshold, and the third threshold, the satellite full disk data is filtered to obtain candidate pixels.
[0008] Furthermore, the spatiotemporal input features of candidate pixels are extracted by statistically analyzing the average range and duration of candidate pixels to obtain the spatiotemporal input features.
[0009] Furthermore, the feature is that the multi-band spatiotemporal observation data of the geostationary meteorological satellite includes remote sensing indices and band data; the specific method for extracting the channel input features of candidate pixels is as follows: Construct dynamic indices; By combining dynamic indices with remote sensing indices and band data, channel input characteristics are obtained. The construction of dynamic indices includes: The normalized combustion ratio (NBR) is calculated based on the near-infrared and short-infrared bands. Calculate the change in mid-infrared temperature in the spatiotemporal dimension to obtain the dynamic temperature change. Construct a temperature gain control function based on NBR; Dynamic thermal anomaly terms are obtained based on dynamic temperature change and temperature gain control function; Dynamic indices are generated based on NBR and dynamic outliers.
[0010] Constructing a dynamic index that integrates dynamic temperature changes and deterministic vegetation information can serve as a higher-dimensional feature during fire processes, thereby enabling more accurate fire point detection.
[0011] Furthermore, the fire spatiotemporal network model is characterized by comprising a spatial feature extraction module, a spatiotemporal collaborative attention module, a temporal dynamic modeling module, and a lightweight classifier connected in sequence. The spatial feature extraction module includes a 2D convolutional layer, a ReLU activation function layer, a first deformable lightweight residual block, a second deformable lightweight residual block, and a third deformable lightweight residual block connected in sequence. The spatiotemporal collaborative attention module includes two parallel attention branches and a weighted fusion module. The outputs of the two parallel attention branches serve as the inputs to the weighted fusion module. The two parallel attention branches are: a time-space collaborative branch and a channel-time collaborative branch. The temporal dynamic modeling module includes a global pooling layer and a bidirectional long short-term memory network connected in sequence.
[0012] Furthermore, the training process of the fire spatiotemporal network model is as follows: Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites, including remote sensing indices and band data; Construct a fire point sample set and a non-fire point sample set; Construct dynamic indices; The dynamic index is combined with the remote sensing index and band data to form the channel input feature; Determine the spatiotemporal input characteristics of the sample based on the characteristics of the fire point; The channel input features and spatiotemporal input features are input into the fire spatiotemporal network model to obtain the classification results; The fire spatiotemporal network model is trained by five-fold cross-validation, and the best set of model parameters is saved to obtain the pre-trained fire spatiotemporal network model.
[0013] Furthermore, the specific steps for constructing the fire point sample set include: Based on multi-band spatiotemporal observation data from geostationary meteorological satellites, combined with actual fire observation information, data correction, filling, and resampling are performed to obtain multi-band spatiotemporal data. Multispectral band features and high-resolution images are acquired, and the multispectral band spatiotemporal data are manually identified based on the multispectral band features and high-resolution images to obtain a fire point sample set.
[0014] Building a high-confidence fire point sample dataset can effectively improve the model's recall and reduce the number of missed fire points.
[0015] Furthermore, the specific steps for obtaining the non-fire point sample set include: Based on the typical underlying surface types in the multi-band spatiotemporal observation data of geostationary meteorological satellites in the detection area, pixel samples with and without cloud cover were selected, and the samples were uniformly distributed based on the observation time to obtain a non-fire point sample set.
[0016] Constructing diverse non-fire point sample datasets can effectively improve model accuracy and reduce false fire point detections. Average distribution of observation time can balance the differences in radiance caused by the angle of solar incidence throughout the day.
[0017] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described in claims 1 to 8.
[0018] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described in claims 1 to 8.
[0019] The beneficial effects of this invention are as follows: 1. By constructing a high-confidence, diverse dataset of fire and non-fire points, the model's recall and accuracy are significantly improved. A high-confidence fire point dataset is constructed by fusing geostationary meteorological satellite fire point products, multi-band remote sensing data, and high-resolution imagery, combined with manual correction. Simultaneously, a highly representative non-fire point dataset is constructed by selecting cloud-covered and cloudless samples from various underlying surface types and balancing their temporal distribution. This effectively alleviates the problems of fire point sample scarcity and class imbalance, enhances the model's generalization ability to complex fire point environments, and significantly reduces false negatives and false negatives.
[0020] 2. A dynamic index is proposed and a multi-dimensional feature engineering is constructed to enhance the discriminative power of fire signals. It effectively integrates rapid thermal anomalies and vegetation response characteristics during fire occurrence, overcoming the problems of traditional temperature features being affected by cloud cover and high-temperature ground surfaces, and the strong lag in vegetation indices. This significantly improves the sensitivity and anti-interference capability of the features to fire signals.
[0021] 3. In terms of model architecture, the progressive design of "spatial feature extraction - spatiotemporal collaborative attention - temporal dynamic modeling" solves the problem of "decoupling" of spatiotemporal processing in traditional methods. Specifically, a spatiotemporal collaborative attention module is introduced to directly model persistent anomalies on the feature layer, achieving accurate capture of the evolution trajectory of weak fire points.
[0022] 4. A pre-screening mechanism for potential fire points is introduced, significantly improving detection efficiency and system practicality. Multiple threshold conditions are used for rapid initial screening of the entire disk data, greatly reducing the number of pixels that the subsequent depth model needs to process. This makes the method highly applicable to the high-frequency, large-scale data processing needs of geostationary meteorological satellites, demonstrating significant engineering application value. Attached Figure Description
[0023] Figure 1 A flowchart of a satellite fire detection method based on dynamic index features and deep learning is provided for an embodiment. Figure 2 Example diagram of fire detection results provided for the embodiment; Figure 3 A schematic diagram of the fire spatiotemporal network model structure provided for an embodiment; Figure 4 A schematic diagram of the deformable lightweight residual block structure provided for the embodiment; Figure 5 A schematic diagram of the spatiotemporal collaborative attention module structure provided in the embodiment; Figure 6 A statistical graph of the high-temperature anomaly stability of DNBRT; Figure 7 A statistical chart of the initial fire response level in DNBRT; Figure 8 This is a statistical chart showing the sensitivity of DNBRT before and after a fire. Detailed Implementation
[0024] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0025] like Figure 1 As shown, in one embodiment of the present invention, a satellite fire detection method based on dynamic exponential features and deep learning includes the following steps: S1. Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites.
[0026] In this embodiment, data from the Fengyun-4B satellite's heat source / fire point products and fire events collected from the internet are used to record the time and location of the fires. L1 data from Fengyun-4B satellite for the corresponding time period is downloaded. Reflectivity correction is applied to bands 1-8 of the L1 data, and brightness temperature correction is applied to bands 9-15 to obtain radiometrically corrected data. Based on the radiometrically corrected data, atmospheric correction is performed using the dark pixel method to obtain atmospherically corrected data. A latitude and longitude lookup table is generated based on the L1 data. The radiometrically and atmospherically corrected data is converted into a VRT file, and the corresponding latitude and longitude information is inserted into the VRT file. The VRT file is then converted into a TIF file to complete geometric correction, obtaining corrected data. Linear interpolation is used to fill and resample the corrected data to obtain multi-band spatiotemporal data, including remote sensing indices and band data.
[0027] S2. Filter the satellite full disk data using multiple threshold conditions to obtain candidate pixels.
[0028] Specifically, including: Based on the mid-infrared temperature of pixels in the satellite full-disk data and the surface temperature, the fire index and temperature difference are obtained; the formulas for calculating the fire index and temperature difference are as follows:
[0029] In the formula, Fire point index, Mid-infrared temperature, For surface temperature, This represents the difference between the mid-infrared temperature and the surface temperature.
[0030] Based on this, a first threshold condition is established to filter the temperature features of pixels, and the rule is as follows: ; Based on the red light band, near-infrared band, and surface temperature, a second threshold is obtained to filter out pixels covered by clouds. The rules are as follows:
[0031] In the formula, NIR represents the reflectivity in the near-infrared band. The reflectivity is in the red light band.
[0032] A third threshold is obtained based on the shortwave infrared band, and water pixels are filtered out according to the following rules:
[0033] In the formula, SWIR is the reflectivity of the shortwave infrared band.
[0034] Based on the first threshold condition, the second threshold condition, and the third threshold condition, multiple threshold conditions are obtained. Based on the multiple threshold conditions, data in the entire disk of Fengyun-4B satellite are filtered to obtain potential fire point pixels.
[0035] S3. Extracting the spatiotemporal input features and channel input features of candidate pixels based on multi-band spatiotemporal observation data from geostationary meteorological satellites.
[0036] The specific steps for extracting the spatiotemporal input features of candidate pixels are as follows: statistically analyze the average range and duration of candidate pixels to obtain the spatiotemporal input features.
[0037] The specific steps for extracting the channel input features of candidate pixels are as follows: Construct dynamic indices; By combining dynamic indices with remote sensing indices and band data, channel input characteristics are obtained. The construction of dynamic indices includes: Based on the near-infrared and short-infrared bands, the normalized combustion ratio (NBR) is calculated, and its expression is as follows:
[0038] In the formula, NIR represents the reflectance in the near-infrared band, and SWIR represents the reflectance in the short-wave infrared band. The NBR index is mainly used to characterize the changes in vegetation moisture during combustion, and it has wide applications in fire intensity assessment and burned areas.
[0039] The change in mid-infrared temperature over time is calculated to obtain the dynamic temperature change, which is expressed as follows:
[0040] In the formula, This represents the dynamic temperature change. Indicates the temperature increment over time. This represents the temperature increment in space. This is the balance coefficient.
[0041] The temperature gain control function is constructed based on the NBR, and its expression is as follows:
[0042] In the formula, Here, is the temperature gain control function, and tanh is the hyperbolic tangent function.
[0043] The temperature gain control function can control the impact of temperature changes on the final index based on the NBR value. When the NBR is high, the vegetation is in a healthy state, and temperature changes may be considered noise. The hyperbolic tangent function can suppress high values and promote low values, thereby reducing the noise caused by temperature.
[0044] The dynamic thermal anomaly term is obtained based on the dynamic temperature change and the temperature gain control function. Its expression is as follows:
[0045] In the formula, e is the natural constant. This refers to dynamic thermal anomalies.
[0046] The dynamic thermal anomaly term substitutes the product of the temperature gain control function and the dynamic temperature change rate into an exponential function to promote temperature increases and suppress temperature decreases.
[0047] Based on NBR and dynamic outliers, a dynamic index is generated, and its expression is as follows:
[0048] In the formula, The DNBRT is a dynamic index that represents the severity of a fire; a higher DNBRT value indicates a more severe fire. Changing the NBR value range from [-1, 1] to [0, 2] can reduce the impact of sign changes on the DNBRT value.
[0049] In this embodiment, the channel feature selection includes the DNBRT index, as well as mid-infrared temperature, surface temperature, red light reflectance, near-infrared reflectance, short-wave infrared reflectance, normalized vegetation index, normalized combustion ratio, and the difference between mid-infrared temperature and surface temperature; the average coverage and duration of sample points in the fire point sample data are statistically analyzed; the spatial feature can be a 21*21 pixel background window with a resolution of 2km in geostationary meteorological satellites; and the temporal feature can be a time series of nine 15-minute observation intervals to obtain multi-dimensional features.
[0050] S4. Input the spatiotemporal input features and channel input features of the candidate pixels into the pre-trained fire spatiotemporal network model to obtain the fire point detection results, such as... Figure 2 As shown.
[0051] like Figure 3 As shown, the fire spatiotemporal network model includes a spatial feature extraction module, a spatiotemporal collaborative attention module, a temporal dynamic modeling module, and a lightweight classifier connected in sequence. The spatial feature extraction module includes a 2D convolutional layer, a ReLU activation function layer, a first deformable lightweight residual block, a second deformable lightweight residual block, and a third deformable lightweight residual block connected in sequence. The spatiotemporal collaborative attention module includes two parallel attention branches and a weighted fusion module, with the outputs of the two parallel attention branches serving as the input to the weighted fusion module. The two parallel attention branches are a temporal-spatial collaborative branch and a channel-temporal collaborative branch, respectively. The temporal dynamic modeling module includes a global pooling layer and a bidirectional long short-term memory network connected in sequence.
[0052] Spatial feature extraction is fundamental for models to understand single-temporal multispectral images. The spatial feature extraction module uses a lightweight ResNet as its backbone and incorporates two key improvements—deformable convolution and attention mechanisms—to adapt it to the detection requirements of fire points on low-resolution geostationary satellite images.
[0053] Specifically, in the spatial feature extraction module, the initial feature mapping layer performs preliminary transformations on the input multi-dimensional features. This layer uses a 3×3 convolutional kernel with a stride and padding of 1 to ensure that the spatial size of the output feature map remains unchanged at 21×21. The convolution operation maps the number of channels to 32, followed by batch normalization and ReLU activation to generate a feature map containing basic spectral and texture responses, laying the foundation for subsequent deep abstraction.
[0054] Deep feature extraction is achieved through a three-tiered, cascaded improved residual block architecture. The first deformable lightweight residual block increases the number of channels to 64 while maintaining spatial resolution; the second deformable lightweight residual block downsamples to 10×10 using a convolution with a stride of 2, increasing the number of channels to 128 to expand the receptive field; the third deformable lightweight residual block further downsamples to 5×5, reaching 256 channels, forming a highly abstract spatial representation. This progressive design enables the network to capture features at different scales, which is crucial for identifying small fire targets that occupy only a few pixels.
[0055] like Figure 4 As shown, each deformable lightweight residual block integrates deformable convolution and CBAM attention mechanisms. Deformable convolution predicts the spatial offset for each sampling point, enabling the convolution kernel to adaptively adjust the receptive field based on the input feature content. This mechanism allows the model to actively focus on high-temperature or abnormal regions within a pixel, improving the feature extraction accuracy for irregular targets.
[0056] Building upon the geometric adaptation provided by deformable convolution, each residual block further integrates a convolutional block attention module. CBAM sequentially performs channel and spatial attention calculations: channel attention generates weight vectors by aggregating spatial information, enhancing the response to fire-sensitive channels (such as mid-infrared and DNBRT); spatial attention generates a two-dimensional weight map, guiding the network to focus on spatial regions in the image where anomalies may occur. This design paradigm of "deformable convolution transformation - CBAM attention refinement - residual connection fusion" ensures that the network can efficiently learn highly discriminative spatial patterns.
[0057] After three levels of residual block processing, the input at each time step is converted into a deep feature map with 256 channels and a spatial size of 5×5. The feature maps from the nine time steps are stacked along the time dimension to obtain a spatiotemporal feature tensor of shape [B, 9, 256, 5, 5], which prepares for subsequent spatiotemporal co-processing.
[0058] like Figure 5 As shown, the spatiotemporal collaborative attention module receives a temporal feature tensor of shape [B, 9, 256, 5, 5]. Its core design principle is to establish a collaborative relationship between the temporal and spatial dimensions while maintaining spatial resolution, and to assign higher weights to spatiotemporal locations with a continuously increasing trend. The module achieves this goal through two parallel attention branches, ultimately fusing their outputs to modulate the original features.
[0059] The first is the temporal-spatial co-attention branch. This branch treats the input tensor as a spatiotemporal cube and uses a 3×3×3 three-dimensional convolutional kernel to directly extract local spatiotemporal patterns along the three dimensions of time, height, and width. This enables the model to perceive the evolutionary relationship of features between adjacent time intervals and adjacent spatial locations, thereby identifying spatial locations where the response value continuously increases across multiple frames.
[0060] The second is the channel-temporal collaborative attention branch. This branch focuses on analyzing the differences in importance of different feature channels in the temporal dimension. It first averages all spatial locations of each channel in the temporal dimension to obtain a temporal series overview of each channel; then, it learns the weights of each channel during the temporal evolution through fully connected layers. This allows the model to distinguish between critical channels that are sensitive to fire dynamics and relatively stable background channels, focusing on the specific patterns of the former's changes over time.
[0061] Two branches independently generate attention weight vectors for the channel dimension, which are then normalized using the Sigmoid function and summed to form the final spatiotemporal co-attention weights. Multiplying these weights by the original features enables adaptive modulation of the feature tensor. After processing by this module, the shape of the feature tensor remains unchanged, but the spatiotemporal patterns related to fire points are significantly enhanced, providing a more focused and discriminative input for subsequent temporal modeling. By introducing the spatiotemporal co-attention module, continuous anomalies are modeled directly on the feature layer, enabling accurate capture of the evolution trajectory of weak fire points.
[0062] To fully utilize the high temporal resolution of geostationary satellites and capture the complete life cycle of the fire point, the model introduces temporal dynamic modeling after the spatiotemporal collaborative attention module. First, global average pooling is performed on the modulated feature tensor, and the 5×5 spatial features of each time step are aggregated into a 256-dimensional feature vector, resulting in the sequence [B, 9, 256].
[0063] The sequence is fed into a single-layer bidirectional long short-term memory (BiLSTM) network. The BiLSTM contains two independent LSTM units, forward and backward, which parse the sequence from past to future and from future to past, respectively. Each LSTM unit selectively remembers, forgets, and outputs information through sophisticated input, forget, and output gates, thereby modeling long-term dependencies. The forward LSTM captures the influence of historical information on the current state, while the backward LSTM captures the constraints of future information; their combination enables the model to comprehensively understand the temporal context of the fire.
[0064] The BiLSTM ultimately outputs a feature sequence [B, 9, 256] that incorporates bidirectional temporal information. The feature vector from the last time step is taken as a summary of the entire temporal process. This 512-dimensional vector (256 dimensions each for both directions) encodes the complete dynamic evolution information of the target pixel over the past nine time steps. This design is crucial for distinguishing the continuous warming process of a real fire from instantaneous disturbances such as cloud movement, because real fires typically exhibit an ordered temporal evolution pattern.
[0065] The 512-dimensional feature vector after spatiotemporal fusion is fed into a lightweight classifier to make the final decision. In this embodiment, the classifier consists of two fully connected layers: the first layer reduces the dimension to 64 and introduces non-linearity through ReLU activation; the second layer further maps to 2 dimensions, corresponding to the two categories of "non-fire" and "fire". To alleviate overfitting, Dropout regularization is introduced between the two layers. Finally, the network output is converted into a probability distribution through the Softmax function to obtain the confidence score of each pixel belonging to each category.
[0066] The changes in data dimensions during the forward propagation of the entire model clearly demonstrate its processing logic: from the original input [B, 9, 11, 21, 21] to the spatial features [B, 9, 256, 5, 5], the shape is preserved after spatiotemporal collaborative attention modulation, then pooled into a sequence [B, 9, 256], and the spatiotemporal fusion features [B, 512] are obtained through BiLSTM, finally outputting the classification result [B, 2]. This design, while controlling the number of parameters, achieves accurate perception and discrimination of weak spatiotemporal signals of fire points, providing an effective deep learning solution for large-scale, near-real-time fire monitoring by geostationary satellites.
[0067] The training process of the fire spatiotemporal network model is as follows: Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites, including remote sensing indices and band data; Construct a fire point sample set and a non-fire point sample set.
[0068] The specific steps for constructing a fire point sample set include: Based on multi-band spatiotemporal observation data from geostationary meteorological satellites, combined with satellite fire point products and actual observed fire information, data correction, filling, and resampling are performed to obtain multi-band spatiotemporal data. Multispectral band features and high-resolution imagery are acquired, and fire point sample sets are obtained through manual identification of the multi-band spatiotemporal data based on these features and imagery. Specifically, based on the multi-band spatiotemporal data, combined with the fire point determination conditions in the traditional threshold method and the daily band changes of the sample, it is determined that the infrared temperature of fire point pixels suddenly increases compared to the surface temperature and remains high for a period of time, gradually decreasing towards the end of the fire. Simultaneously, high-resolution imagery from environmental satellites and Gaofen satellites can be used to observe the same area before and after a fire. Real fire point pixels show obvious burning and carbonization of vegetation after a fire, allowing for manual labeling of high-confidence fire point pixels to obtain a fire point sample dataset.
[0069] The specific steps for obtaining the non-fire point sample set include: Based on typical underlying surface types within the detection area, pixel samples with and without cloud cover are selected. These samples are then uniformly distributed based on observation time to obtain a non-fire point sample set. In this embodiment, equal amounts of cloud-covered and cloudless samples are selected from ten types of underlying surfaces in China (including urban areas, farmland, glaciers, wetlands, deserts, temperate grasslands, alpine grasslands, coniferous forests, broad-leaved forests, and mixed forests) to ensure the dataset includes cloud cover characteristics. Furthermore, the selected observation times are evenly distributed across morning, noon, and afternoon to balance the differences in radiance caused by the angle of solar incidence throughout the day. The same preprocessing method as the fire point sample dataset is used to obtain the non-fire point sample dataset.
[0070] Construct dynamic indices; The dynamic index is combined with the remote sensing index and band data to form the channel input feature; Determine the spatiotemporal input characteristics of the sample based on the characteristics of the fire point; The channel input features and spatiotemporal input features are input into the fire spatiotemporal network model to obtain the classification results; The fire spatiotemporal network model is trained by five-fold cross-validation, and the best set of model parameters is saved to obtain the pre-trained fire spatiotemporal network model.
[0071] like Figure 6 , Figure 7 , Figure 8 As shown, in order to verify the effectiveness of the dynamic index proposed in this invention, the stability of the DNBRT in the face of high temperature anomalies is verified based on the relative standard deviation; the responsiveness of the DNBRT in the early stage of a fire is verified based on the relative rate of change in the early stage of a fire; and the sensitivity of the DNBRT in fire events is verified based on the relative difference before and after a fire.
[0072] Samples of fire points that were falsely detected due to high-temperature anomalies were collected from the fire point sample dataset. The relative standard deviations of the mid-infrared temperature, NBR, and DNBRT for these samples on the same day were calculated. Based on the maximum, minimum, mean, median, and interquartile range of the three sets of data, the stability of these three indicators in the face of high-temperature anomaly noise was evaluated. The formula for calculating the relative standard deviation is:
[0073] Where RSD is the relative standard deviation. The standard deviation of the sample is 1. This is the sample average.
[0074] A fire point sample dataset was collected, focusing on samples taken at the moment of fire occurrence. The relative change rates of NBR and DNBRT were calculated after these samples transitioned from non-fire points to fire points. Based on the maximum, minimum, mean, median, and interquartile range of the two datasets, the responsiveness of these two indicators in the early stages of a fire was assessed. The formula for calculating the relative change rate is as follows:
[0075] Where RC is the relative rate of change. These are sample values taken at the time of the fire. These are sample values taken just before the fire occurred.
[0076] Complete fire samples were collected from the fire point sample dataset. The relative differences between NBR and DNBRT before and after the fire were statistically analyzed. Based on the maximum, minimum, mean, median, and interquartile range of the two sets of data, the sensitivity of the two indicators in fire events was evaluated. The formula for calculating the relative difference is as follows:
[0077] Where RD represents relative difference. This represents the sample mean during the period in which the fire occurred. This represents the sample mean before the fire. This represents the sample mean after the fire.
[0078] Furthermore, the capabilities of the fire spatiotemporal network model provided in this invention for fire point detection were compared with those of classic deep learning models to obtain verification results. Five classic deep learning models—ResNet, VGG, Vision Transformer, DenseNet, and MLP—were selected for comparison to verify the advantages of the fire spatiotemporal network model in fire point detection. The comparison results are shown in Table 1.
[0079] Table 1
[0080] As shown in the table above, the fire spatiotemporal network model provided by this invention significantly outperforms other comparative models in overall performance. Its recall rate is as high as 0.957, the best among all models, meaning it has the lowest false negative rate in actual fire detection and can capture real fire points to the greatest extent, which is crucial for fire early warning tasks. Simultaneously, its F1 score (0.819) and Fβ score (0.891) are both the highest, indicating that the model achieves the best balance between precision and recall, resulting in the strongest overall recognition capability. Furthermore, the model's inference speed (0.035 seconds / pixel) is second only to ViT and far faster than structures such as ResNet and DenseNet, ensuring high accuracy while maintaining high computational efficiency, making it very suitable for fire monitoring scenarios that require both real-time performance and reliability.
[0081] This method combines fire point products from geostationary meteorological satellites, daily band variation characteristics of pixels, and high-resolution satellite imagery to construct a high-confidence dataset of fire and non-fire point samples, thus providing a foundation for subsequent model training. The selection of geostationary meteorological satellites and high-resolution satellites is not specifically limited. Using mid-infrared temperature and surface temperature as features for fire point identification can lead to false positives due to cloud cover and high-temperature areas. While vegetation information such as NBR is primarily used for post-disaster assessment and burned areas, it lacks the ability to identify dynamic fires. Leveraging the advantages of geostationary meteorological satellites, this method combines dynamic, rapidly changing temperature information with lagging vegetation information to construct a more accurate dynamic index to characterize fire events. Vegetation information can include Normalized Difference Vegetation Index (NDI), Normalized Burn Ratio (NDR), and Enhanced Vegetation Index (EGI), among others, without specific limitations. Based on the constructed dynamic index... The system combines other remote sensing indices and band data to form channel features. Based on the spatiotemporal resolution of geostationary meteorological satellites and fire sample data, spatial and temporal features are determined. All feature dimensions can be adjusted based on instance data and are not specifically limited here. Features are extracted based on a fire spatiotemporal network model. A convolutional neural network structure is used to extract multi-level feature representations of spatial and channel dimensions from multi-band remote sensing images. The structure of the convolutional neural network can be adjusted according to data characteristics and is not specifically limited here. Subsequently, the CBAM attention module is used for adaptive optimization of features. The channel attention mechanism enhances the discriminative power of spectral features, and the spatial attention mechanism focuses on key spatial areas to achieve feature importance weighting. Finally, a BiLSTM module is used to process temporal features, modeling the dynamic evolution of the fire from both forward and reverse time dimensions. Simultaneously, a multi-threshold discrimination mechanism is used to pre-screen the full-disk data of geostationary meteorological satellites, quickly eliminating non-fire pixels based on prior physical knowledge, significantly reducing the data scale for subsequent deep model processing and comprehensively improving system computational efficiency.
[0082] This invention effectively overcomes the high false alarm and false negative rates caused by cloud cover, high-temperature ground surface, and vegetation background interference in traditional fire detection methods. By integrating dynamic index features and multi-dimensional remote sensing indicators, it significantly improves the discrimination ability and anti-interference capability of fire signals. Simultaneously, this invention fully utilizes the high spatiotemporal resolution advantage of geostationary meteorological satellites to construct a lightweight deep learning model adapted to small sample scenarios, solving the problem of weak fire signal extraction and spatial localization in mixed pixel backgrounds. The proposed technical solution can not only achieve large-scale, high-frequency automatic fire detection, but also accurately identify early fire points and weak fire targets. The detection results have high reliability, stability, and generalization ability, providing key technical support and scientific basis for forest fire monitoring, disaster assessment, emergency response decision-making, and ecological environmental protection.
Claims
1. A satellite fire detection method based on dynamic exponential features and deep learning, characterized in that, include: Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites; Candidate pixels are obtained by filtering the full-disk satellite data using multiple threshold conditions. Spatiotemporal input features and channel input features of candidate pixels are extracted based on multi-band spatiotemporal observation data from geostationary meteorological satellites; The spatiotemporal input features and channel input features of candidate pixels are input into a pre-trained fire spatiotemporal network model to obtain fire point detection results.
2. The method according to claim 1, characterized in that, Pixel filtering of the satellite full-disk data using multiple threshold conditions yielded candidate pixels, including: Fire index and temperature difference are obtained based on the mid-infrared temperature of pixels in the satellite full disk data and the surface temperature. A first threshold is set based on the fire point index and temperature difference; A second threshold is set based on the red light band, the near-infrared band, and the surface temperature. A third threshold is set based on the shortwave infrared band; Based on the first threshold, the second threshold, and the third threshold, the satellite full disk data is filtered to obtain candidate pixels.
3. The method according to claim 1, characterized in that, The specific steps for extracting the spatiotemporal input features of candidate pixels are as follows: statistically analyze the average range and duration of candidate pixels to obtain the spatiotemporal input features.
4. The method according to claim 1, characterized in that, Multi-band spatiotemporal observation data from geostationary meteorological satellites include remote sensing indices and band data; the channel input features for extracting candidate pixels specifically include: Construct dynamic indices; By combining dynamic indices with remote sensing indices and band data, channel input characteristics are obtained. The construction of dynamic indices includes: The normalized combustion ratio (NBR) is calculated based on the near-infrared and short-infrared bands. Calculate the change in mid-infrared temperature in the spatiotemporal dimension to obtain the dynamic temperature change. Construct a temperature gain control function based on NBR; Dynamic thermal anomaly terms are obtained based on dynamic temperature change and temperature gain control function; Dynamic indices are generated based on NBR and dynamic outliers.
5. The method according to claim 1, characterized in that, The fire spatiotemporal network model consists of a spatial feature extraction module, a spatiotemporal collaborative attention module, a temporal dynamic modeling module, and a lightweight classifier, which are connected in sequence. The spatial feature extraction module includes a 2D convolutional layer, a ReLU activation function layer, a first deformable lightweight residual block, a second deformable lightweight residual block, and a third deformable lightweight residual block connected in sequence. The spatiotemporal collaborative attention module includes two parallel attention branches and a weighted fusion module. The outputs of the two parallel attention branches serve as the inputs to the weighted fusion module. The two parallel attention branches are: a time-space collaborative branch and a channel-time collaborative branch. The temporal dynamic modeling module includes a global pooling layer and a bidirectional long short-term memory network connected in sequence.
6. The method according to claim 1, characterized in that, The training process of the fire spatiotemporal network model is as follows: Acquire multi-band spatiotemporal observation data from geostationary meteorological satellites, including remote sensing indices and band data; Construct a fire point sample set and a non-fire point sample set; Construct dynamic indices; The dynamic index is combined with the remote sensing index and band data to form the channel input feature; Determine the spatiotemporal input characteristics of the sample based on the characteristics of the fire point; The channel input features and spatiotemporal input features are input into the fire spatiotemporal network model to obtain the classification results; The fire spatiotemporal network model is trained by five-fold cross-validation, and the best set of model parameters is saved to obtain the pre-trained fire spatiotemporal network model.
7. The method according to claim 6, characterized in that, The specific steps for constructing a fire point sample set include: Based on multi-band spatiotemporal observation data from geostationary meteorological satellites, combined with actual fire observation information, data correction, filling, and resampling are performed to obtain multi-band spatiotemporal data. Multispectral band features and high-resolution images are acquired, and the multispectral band spatiotemporal data are manually identified based on the multispectral band features and high-resolution images to obtain a fire point sample set.
8. The method according to claim 1, characterized in that, The specific steps for obtaining the non-fire point sample set include: Based on the typical underlying surface types in the multi-band spatiotemporal observation data of geostationary meteorological satellites in the detection area, pixel samples with and without cloud cover were selected, and the samples were uniformly distributed based on the observation time to obtain a non-fire point sample set.
9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method described in claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The device stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described in claims 1 to 8.