A forest fire satellite monitoring method and system

Abstract

The present application relates to the technical field of mountain fire monitoring, and discloses a mountain fire satellite monitoring method and system. The mountain fire satellite monitoring method comprises the following steps: obtaining a satellite image to be analyzed of a target region; using a pre-trained target detection model to identify and locate target object pixels containing target objects in the satellite image to be analyzed, wherein the target objects include interference sources, and the interference sources include non-fire point high-reflection artificial ground objects and / or fixed heat objects; monitoring and locating potential fire point pixels in the satellite image to be analyzed; spatially superimposing the target object pixels and the potential fire point pixels, and determining that any potential fire point pixel is a false fire point pixel in response to the potential fire point pixel being overlapped or partially overlapped with the target object pixels, or otherwise determining that the potential fire point pixel is a real fire point pixel. The mountain fire satellite monitoring method and system can identify and exclude non-interference source interference, thereby improving the accuracy of fire point monitoring.

Description

Technical Field

[0001] This invention relates to the field of wildfire monitoring technology, and in particular to a wildfire satellite monitoring method and system. Background Technology

[0002] Satellite-based wildfire monitoring technology boasts advantages such as wide monitoring range, high observation frequency, and low cost, making it one of the primary methods for monitoring wildfires in forests, grasslands, and other areas. According to Wien's displacement law, the peak wavelength of blackbody radiation is inversely proportional to temperature. The peak wavelength of radiation at a normal Earth surface (approximately 290K) is about 11μm, while wildfire temperatures typically exceed 750K, with their peak thermal radiation wavelength approaching the 4μm mid-infrared channel. Therefore, meteorological satellites identify fire points by capturing the radiation difference between the 4μm and 11μm bands.

[0003] However, traditional satellite wildfire monitoring methods are severely affected by ground-based interference sources in practical applications, mainly including man-made features that highly reflect sunlight, such as photovoltaic panels, and fixed heat sources such as thermal plants, resulting in significant false alarms.

[0004] The existing technologies offer the following main solutions to the above problems:

[0005] The first category involves fusing other monitoring methods and meteorological environmental factors to reduce false alarms. Chinese patents CN107038833B, CN110634258A, and CN 110334655A utilize multi-source data such as historical meteorological data, vegetation index, and aerosol concentration to further identify preliminarily identified fire points. However, this type of method relies on indirect environmental correlation characteristics, limiting its effectiveness in suppressing systematic false alarms. Chinese patents CN116403354A, CN115577322A, and CN116047546B reduce false alarm rates by combining satellite monitoring with image monitoring devices, radar, and other satellite monitoring methods. The monitoring range and observation frequency of different methods vary significantly, making fusion difficult and failing to fundamentally address the problem of interference from ground-based sources affecting satellite monitoring.

[0006] The second method involves filtering based on the type of ground cover at the initial fire location. Chinese patent CN 104269012 B proposes that if the Normalized Difference Vegetation Index (NDVI) at the initial fire location is <0.3, it can be considered a non-actual fire point, belonging to the category of industrial fire points. The problem with this method is that NDVI represents an average value over a certain range, with a resolution typically ranging from several hundred meters to 1 kilometer. Wildfires do not require large amounts of vegetation; scattered bushes and nurseries can also ignite them. Therefore, relying solely on NDVI to determine whether an actual wildfire will occur is highly inaccurate. Furthermore, this method has limited effectiveness in filtering out highly reflective interference sources.

[0007] The third type involves filtering out potential fire points by comparing their initial fire location with the location of ground interference sources. Chinese patents CN113361323 B, CN 111783560 A, CN 104537792 A, and CN 103455804 B propose comparing the coordinates of the initial fire point with the coordinates of ground interference sources; if the fire point is in or near a ground interference source area, it is determined to be a non-real fire point. Chinese patents CN111858813 A and CN 112562244 B further propose using electronic maps to check for ground interference sources near the initial fire point location. Theoretically, this method is effective in resolving false alarms in satellite wildfire monitoring caused by ground interference sources. However, there are over 100,000 ground interference source areas nationwide, including photovoltaic panels and thermal plants, and publicly available location data is limited. How to efficiently and accurately obtain the location of ground interference sources is the key to implementing this method, but currently, no publicly available technical solutions provide a method for establishing a ground interference source location database. Summary of the Invention

[0008] Therefore, the purpose of this invention is to efficiently and accurately obtain the location of ground interference sources, overcome the problem of high false alarm rate caused by the high reflectivity of sunlight of non-fire point high reflective man-made objects (such as photovoltaic panels) and / or the self-heating characteristics of fixed heat-generating objects (such as thermal plants) in the prior art, and provide a wildfire satellite monitoring method and system to identify and eliminate ground interference sources, thereby improving the accuracy of fire point monitoring.

[0009] To address the aforementioned technical problems, this invention provides a method for satellite monitoring of wildfires, comprising the following steps:

[0010] Acquire satellite images of the target area to be analyzed;

[0011] Using a pre-trained target detection model, target pixels containing target objects in the satellite image to be analyzed are identified and located. The target objects include interference sources, which include non-fire point high reflectivity artificial ground features and / or fixed heat-generating objects.

[0012] Monitor and locate potential fire pixels in the satellite image to be analyzed;

[0013] The target object pixel and the potential fire point pixel are spatially superimposed. In response to any potential fire point pixel overlapping or partially overlapping with the target object pixel, the potential fire point pixel is determined to be a false fire point pixel; otherwise, the potential fire point pixel is determined to be a real fire point pixel.

[0014] Preferably, the training method for the target detection model includes:

[0015] Construct a dataset that includes training images and validation images, both of which consist of satellite images under diverse geographical environments, lighting conditions, and climate characteristics.

[0016] The dataset is preprocessed to identify and label targets in the satellite images;

[0017] The target detection model is trained by taking the dataset as input and the preprocessed dataset as output.

[0018] Preferably, the preprocessing of the dataset includes labeled bounding boxes, the labeled bounding boxes including:

[0019] Calculate the distribution density and arrangement characteristics of the targets in the satellite image;

[0020] Based on the distribution density and arrangement characteristics, the bounding box is labeled using either the region labeling mode or the independent labeling mode.

[0021] The region labeling modes include:

[0022] When the target objects are densely and continuously distributed, a single bounding box is used to cover the continuous distribution area of ​​the target objects;

[0023] When the target objects are arranged in a dense array, the array area of ​​the target objects is covered by the minimum bounding rectangle.

[0024] The independent annotation modes include:

[0025] When the target objects are distributed independently or arranged in an independent array, each target object is labeled with a bounding box, and the proportion of the background area within each bounding box is minimized.

[0026] Preferably, the preprocessing of the dataset includes data augmentation, which includes:

[0027] Based on the distribution density of the target object, the satellite image is cropped preferentially in the distribution area of ​​the target object, so that the cropped image contains at least one target object.

[0028] Each cropped image to be stitched is subjected to color balance processing to simulate color distribution under different lighting conditions;

[0029] Multiple cropped images are stitched together according to a preset layout to generate a composite image, and pixel fusion transition processing is performed on the stitching edges of adjacent cropped images to eliminate stitching marks.

[0030] An enhancement operation is performed on the synthesized image, and the synthesized image is adjusted to the input size required by the target detection model;

[0031] The enhancement operation includes at least one of photometric adjustment, geometric transformation, and spatial deformation.

[0032] Preferably, the preprocessing of the dataset includes setting prior boxes, which include:

[0033] Based on the geometric information of the bounding box, a clustering algorithm is used to analyze the bounding box and obtain several cluster centers;

[0034] The bounding box size corresponding to the cluster center is used as the prior box size of the target detection model;

[0035] Prior boxes are assigned to different detection layers of the target detection model according to their size: large-scale prior boxes are assigned to the target detection model for detecting shallow feature maps of large targets, medium-scale prior boxes are assigned to the target detection model for detecting medium-scale feature maps of medium targets, and small-scale prior boxes are assigned to the target detection model for detecting deep feature maps of small targets.

[0036] Preferably, the preprocessing of the dataset includes masking, which includes:

[0037] Calculate at least two land cover indices for the satellite image, the land cover indices including the normalized vegetation index and the normalized building index;

[0038] The system identifies vegetation areas based on the normalized vegetation index and generates a first suppression mask for the vegetation areas; it identifies building areas based on the normalized building index and generates a second suppression mask for the building areas; it sets suppression intensity weights for the first suppression mask and the second suppression mask respectively, and then fuses the first suppression mask and the second suppression mask to obtain a comprehensive background suppression mask.

[0039] The integrated background suppression mask is used to process satellite images to suppress interference from background areas.

[0040] Preferably, the target detection model includes a YOLOv5 model, wherein the loss function of the YOLOv5 model includes a generalized intersection-union loss function, and the weights of the class loss are the lowest-level weights;

[0041] In the global C3 module of the YOLOv5 model, the initial values ​​of the feature extraction weights are all equal.

[0042] The YOLOv5 model includes a feature selection mechanism, which is used to adjust the feature extraction weights of the C3 module in different detection layers of the YOLOv5 model.

[0043] When the input is a satellite image lacking color information or a satellite image of the living environment, the feature extraction weight of the shallow C3 module in the backbone network of the YOLOv5 model is increased, so that the YOLOv5 model can enhance its ability to extract edge, shape, texture features and spatial arrangement patterns.

[0044] When the input is a color image or a natural environment image, the feature extraction weight of the C3 module in the multi-scale feature fusion layer of the YOLOv5 model's neck network is increased, which enhances the YOLOv5 model's ability to extract spectral features and reflectance properties.

[0045] Preferably, the C3 module includes a C3_p structure, and the p_C3 Block in the C3_p structure includes a first feature extraction branch and a second feature extraction branch. The outputs of the first feature extraction branch and the second feature extraction branch are added together to form the output of the p_C3 Block.

[0046] The first feature extraction branch includes a first convolutional layer, a second convolutional layer, and a third convolutional layer connected in series, with the second convolutional layer connected in parallel with the second feature extraction branch.

[0047] Preferably, the training process of the target detection model includes a warm-up phase, a decay phase, and a fine-tuning phase;

[0048] The learning rate during the warm-up phase starts from an initial learning rate and gradually increases to a maximum learning rate in a non-linear manner.

[0049] The learning rate during the decay phase is decayed according to the following formula:

[0050] ;

[0051] In the formula, The learning rate during the decay phase; Minimum learning rate; The maximum learning rate; To learn the relative number of times the learning rate decays during the current number of iterations in the decay phase; This represents the total number of learning iterations during the decay phase. This represents the current number of learning iterations during the decay phase.

[0052] The learning rate during the fine-tuning phase is the minimum learning rate.

[0053] Preferably, the wildfire satellite monitoring method further includes: constructing a fire point evolution causal network based on the real fire point pixels and their time series data, and performing causal inference and prediction of fire point spread based on the fire point evolution causal network;

[0054] The nodes of the fire point evolution causal network include outcome variables and causal variables, and there are dynamic causal paths based on causal relationships between outcome variables and causal variables.

[0055] Preferably, the causal inference and prediction of fire spread based on the fire point evolution causal network includes:

[0056] Based on the trained fire point evolution causal network, and using time series data, the causal inference and prediction of the fire point spread range at future moments are performed.

[0057] The training method for the fire point evolution causal network includes:

[0058] Based on the real fire points and their time-series data, the dynamic causal paths and causal strengths between the nodes of the fire point evolution causal network are learned, and a trained fire point evolution causal network is obtained.

[0059] The time-series data includes fire point state variables and environment-driven variables.

[0060] On the other hand, the present invention provides a wildfire satellite monitoring system, comprising:

[0061] The acquisition module is used to acquire satellite images of the target area;

[0062] The interference source localization module is used to identify and locate target pixels containing target objects in the satellite image using a pre-trained target detection model. The target objects include interference sources, which include non-fire point high reflectivity artificial ground objects and / or fixed heat-generating objects.

[0063] A potential fire point location module is used to monitor and locate potential fire point pixels in the satellite image;

[0064] An interference source filtering module is used to spatially overlay the target object pixel and the potential fire point pixel. In response to any potential fire point pixel overlapping or partially overlapping with the target object pixel, the potential fire point pixel is determined to be a false fire point pixel; otherwise, the potential fire point pixel is determined to be a real fire point pixel.

[0065] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0066] The wildfire satellite monitoring method and system of the present invention pre-identifies ground interference sources in a large-scale batch, such as man-made objects that highly reflect sunlight, such as photovoltaic panels, and / or fixed heat-generating objects such as thermal plants. After comparing the initial fire point with its location, the interference objects are eliminated, thereby reducing false alarms caused by ground interference sources from the root and thus improving the overall accuracy of fire point monitoring.

[0067] Furthermore, the proactive elimination method of the present invention, which first identifies interference and then determines the authenticity of potential fire points, is highly targeted, has clear steps, and is easy to integrate into the existing satellite fire point monitoring process. It only requires the addition of a pre-interference identification and spatial overlay analysis step, and the steps are simple, the improvement cost is low, and the feasibility is strong. Attached Figure Description

[0068] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0069] Figure 1 This is a flowchart illustrating a wildfire satellite monitoring method according to an embodiment of the present invention.

[0070] Figure 2 This is a schematic block diagram of the C3 module of the YOLOv5 model in an embodiment of the present invention.

[0071] Figure 3 This is a schematic block diagram of the C3_p structure of the YOLOv5 model in an embodiment of the present invention.

[0072] Figure 4 This is a diagram illustrating the target object recognition effect in the geographical environment in an embodiment of the present invention.

[0073] Figure 5 This is a schematic diagram of a dynamic causal path in an embodiment of the present invention.

[0074] Figure 6 This is a probability diagram of the future fire spread range in an embodiment of the present invention.

[0075] Figure 7 This is a schematic block diagram of a wildfire satellite monitoring system in an embodiment of the present invention. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0077] False alarms from satellite monitoring of wildfires originate from various sources, including but not limited to industrial heat sources, bare rock or building surfaces reflecting sunlight. In recent years, with the rapid development of the photovoltaic power generation industry, the construction of large-scale photovoltaic power plants, and the widespread adoption of distributed photovoltaic panels, new and prominent sources of false alarms have emerged: highly reflective man-made features and / or stationary heat sources at fire points. Photovoltaic panels, especially crystalline silicon modules, have strong reflective properties to sunlight at specific illumination angles. Their radiation signals in the thermal infrared band can easily be confused with the high-temperature signals of actual fire points, leading to frequent false alarms from satellite monitoring systems. Particularly in areas with concentrated photovoltaic installations, such false alarms severely interfere with the accuracy and efficiency of wildfire monitoring.

[0078] To improve the accuracy of wildfire satellite monitoring, this invention introduces a wildfire satellite monitoring method and system.

[0079] Example 1: This example introduces a method for satellite monitoring of wildfires.

[0080] refer to Figure 1 The wildfire satellite monitoring method in this embodiment includes steps S1 to S4.

[0081] Step S1: Obtain satellite images of the target area to be analyzed.

[0082] Step S2: Using a pre-trained target detection model, identify and locate target pixels containing target objects in the satellite image to be analyzed.

[0083] In application, the target object includes interference sources, which include non-fire-point highly reflective artificial ground features and / or fixed heat-generating objects.

[0084] In practical applications, non-fire-point highly reflective artificial features include interfering features made of glass or polished materials. Fixed heat sources are thermal plants.

[0085] In practice, the obstructions made of glass or polished materials can be photovoltaic panels.

[0086] Step S3: Monitor and locate potential fire pixels in the satellite image to be analyzed.

[0087] In application, existing methods for identifying fire points in satellite images can be used, which will not be elaborated here. The identified fire points are used as potential fire points in this embodiment, and the pixels occupied by the potential fire points are used as potential fire point pixels in this embodiment.

[0088] In practical applications, the coordinates of each potential fire point pixel are determined based on the coordinates of each pixel in the satellite image.

[0089] Step S4: Spatially overlay the target object pixel and the potential fire point pixel. In response to any potential fire point pixel overlapping or partially overlapping with the target object pixel, determine that the potential fire point pixel is a false fire point pixel; otherwise, determine that the potential fire point pixel is a real fire point pixel.

[0090] When applied, potential fire point pixels that overlap with the pixels occupied by the target object are all treated as pseudo fire point pixels, and the remaining potential fire point pixels are treated as real fire point pixels.

[0091] In summary, the wildfire satellite monitoring method of this embodiment reduces false alarms caused by interference sources by identifying and eliminating interference from non-fire-point high-reflectivity man-made objects and / or fixed heat-generating objects. This fundamentally improves the overall accuracy of fire point monitoring, with experiments showing an improvement of approximately 10 percentage points. Furthermore, the proactive elimination method of this embodiment, which identifies interference first and then determines the authenticity of potential fire points, is highly targeted, has clear steps, and is easily integrated into existing satellite fire point monitoring processes. It only requires adding a pre-interference identification and spatial overlay analysis step, making the process simple, cost-effective, and highly feasible.

[0092] Example 2: This example introduces a target detection model. The target detection model can be the same as the target detection model described in Example 1.

[0093] The target detection model in this embodiment includes the YOLOv5 model.

[0094] In application, the C3 module of the YOLOv5 model increases the network's depth and receptive field, improving its feature extraction capabilities. Furthermore, the role of the C3 module differs in different parts: in the YOLOv5 model's backbone network, the C3 module halves the size of the feature maps to increase the network's receptive field while reducing computational cost, allowing the network to focus more on global object information, thus improving feature extraction performance. In the YOLOv5 model's neck network, which includes the upsampling module, downsampling module, concat (abbreviated as cat), and the C3 module, the C3 module does not change the feature map size, maintaining the spatial resolution of the feature maps and better preserving local object information. Simultaneously, it further extracts features, increasing the network's depth and receptive field.

[0095] Furthermore, this embodiment of the YOLOv5 model employs a global equal-weight strategy, meaning that the initial values ​​of the feature extraction weights in the global C3 module are all equal. This ensures uniform suppression of all residual connection paths in the network, reducing the direct impact of low-level features such as color and texture extracted by shallow networks on the final recognition result through shortcut connections. This forces the YOLOv5 model to rely more on high-level semantic features extracted by deeper networks to identify interference sources, rather than shallow color and texture. This helps the YOLOv5 model distinguish between features that are similar in shallow features but different in semantic structure, such as photovoltaic arrays, thermal plants, and wildfires, thereby improving its ability to filter interference sources from the feature extraction perspective.

[0096] It is worth noting that the initial values ​​of the feature extraction weights in the C3 module are generally small, for example, less than 1. Preferably, the initial values ​​of the feature extraction weights in the C3 module can be between 0.1 and 0.5.

[0097] Furthermore, the YOLOv5 model includes a feature selection mechanism used to adjust the feature extraction weights of the C3 module in different detection layers of the YOLOv5 model. Specifically: See [link to relevant documentation]. Figure 2 During the training process of the YOLOv5 model, the feature extraction weights of the C3 module in different detection layers of the YOLOv5 model are adjusted according to the environmental scene of the input image.

[0098] (1) When the input is a satellite image lacking color information or a satellite image of the living environment, increase the feature extraction weight of the shallow C3 module in the backbone network of the YOLOv5 model to enhance the YOLOv5 model's ability to extract edge, shape, texture features and spatial arrangement patterns.

[0099] (2) When the input is a color image or a natural environment image, increase the feature extraction weight of the C3 module in the multi-scale feature fusion layer of the YOLOv5 model's neck network to enhance the YOLOv5 model's ability to extract spectral features and reflectance properties.

[0100] In practical applications, each detection layer's C3 module in this embodiment employs a parallel structure, replacing traditional residual connections with a Convolution-BatchNorm-SiLU (CBS) structure containing convolutional kernels of different sizes. This enhances the YOLOv5 model's ability to capture details, enabling it to better handle complex satellite image content. Furthermore, the C3 module in this embodiment includes a C3_p structure. The p_C3 Block within the C3_p structure uses a CBS structure containing convolutions in parallel with the classic Bottleneck to extract features. The feature maps are then summed, which not only preserves the integrity of the original features but also extracts richer detail information through the combination of convolutional kernels of different sizes.

[0101] Specifically, in this embodiment, the p_C3 Block in the C3_p structure includes a first feature extraction branch and a second feature extraction branch connected in parallel.

[0102] The first feature extraction branch includes a cascaded first convolutional layer, such as a p_Conv1_1 structure; a second convolutional layer, such as a p_Conv3_1 structure; and a third convolutional layer, such as a p_Conv1_1 structure. For example: First, a 1×1 first convolutional layer is used to reduce the number of channels in the feature map, thereby reducing the computational burden on subsequent layers. Next, a 3×3 second convolutional layer is applied to the compressed feature map to extract more complex features, such as feature information within local regions. Finally, a 1×1 third convolutional layer is used to increase the number of channels in the feature map, restoring its dimensionality. By setting a series of convolutional layers, the first feature extraction branch effectively reduces the dimensionality of the feature map, thereby reducing computational cost while maintaining the network's feature extraction capability. This allows the network to continue learning features in a higher-dimensional space, providing sufficient information for subsequent network layers.

[0103] The second feature extraction branch includes a first CBS structure, such as a CBS1_1 structure, and a fourth convolutional layer, such as a p_Conv3_1 structure, connected in series. The fourth convolutional layer includes a second CBS structure, such as a CBS3_1 structure, and a third CBS structure, such as a CBS5_1 structure, connected in parallel.

[0104] The first and second feature extraction branches, operating in parallel, extract features from different perspectives. The outputs of both branches are then summed to form the output of p_C3Block. This allows the YOLOv5 model to capture a wider range of feature information, including both complex local features extracted through 3×3 convolutions and detailed features extracted through the CBS structure. By summing these feature maps before further processing, the YOLOv5 model can better integrate feature information from different levels, resulting in higher accuracy and robustness in object detection tasks.

[0105] In the specific implementation process, the head network of the YOLOv5 model in this embodiment is used as the output out of the model, and the loss function of the YOLOv5 model includes the Generalized Intersection over Union (GIoU) ​​loss function.

[0106] Compared to the CIoU (Complete Intersection over Union) loss function, GIoU considers the intersection and union of the predicted and ground truth boxes, and also introduces the area of ​​the minimum bounding box between the predicted and ground truth boxes. This allows for better handling of cases where bounding boxes do not overlap, providing a more comprehensive evaluation of the localization accuracy of the predicted boxes. Furthermore, CIoU not only considers overlapping areas of bounding boxes but also their positional relationships through center point distance and shape through aspect ratio, making CIoU more comprehensive and effective in optimizing bounding box localization. Specifically, GIoU includes penalty factors for center point distance, aspect ratio, and containment to penalize incorrect aspect ratio predictions.

[0107] Furthermore, in this embodiment, the weight of the category loss in the loss function is set to the lowest level, thereby making the bounding box loss and confidence loss play a more important role in the overall loss.

[0108] For detection box loss, use The method loss measures the difference between the predicted box and the ground truth box, taking into account the intersection and union of the predicted and ground truth boxes. It also introduces the area of ​​the minimum bounding box between the predicted and ground truth boxes to more comprehensively evaluate the localization accuracy of the predicted boxes.

[0109] For the evaluation of both confidence loss and classification loss, the Binary Cross-Entropy (BCE) loss method is used to measure the difference between the true and predicted values. The accuracy of the prediction is evaluated by calculating the cross-entropy between the true and predicted values. Specifically, for confidence loss, the BCE loss method is used to evaluate the probability that the predicted bounding box belongs to the interference source category, i.e., the confidence of the YOLOv5 model in detecting interference sources. For classification loss, the BCE loss method is used to evaluate the accuracy of the model's prediction of the interference source category.

[0110] Example 3: This example introduces a training method for an object detection model. The object detection model can be either the object detection model described in Example 1 or the object detection model described in Example 2.

[0111] The training method for the target detection model in this embodiment includes steps S21 to S23.

[0112] Step S21: Construct the dataset.

[0113] In application, to verify the accuracy of the target detection model under ideal conditions and simulate the performance of photovoltaic panel detection in real-world scenarios, the model was tested. For example, in complex urban backgrounds, the model needs to distinguish the reflective properties of photovoltaic panels from those of glass curtain walls, and in natural scenes, it needs to overcome the problem of missing local features caused by vegetation shading. The dataset in this embodiment includes training images and validation images, both of which consist of satellite images under diverse geographical environments, lighting conditions, and climate characteristics.

[0114] In practical applications, the training set contains 971 high-resolution satellite images of photovoltaic panels (640×640 pixels). These images were collected under diverse geographical and environmental conditions to enhance the generalization ability of the object detection model. Specifically, the training set covers different regions such as cities, rural areas, deserts, and coastal areas; includes various lighting conditions such as strong light, cloudy days, and dusk, as well as climatic features such as snowy and foggy days; and encompasses simple backgrounds such as solid colors or single objects, as well as complex backgrounds such as vegetation obstruction, building clusters, and shadow interference. The size, angle, age, and arrangement density of the photovoltaic panels all vary significantly.

[0115] The validation set contains 66 high-resolution 640×640 pixel satellite images of photovoltaic panels. These images comprehensively cover diverse scenes to test the object detection model's recognition capabilities under varying background complexity and image quality. Furthermore, the validation set includes solid-color backgrounds, complex urban backgrounds such as high-rise buildings and streets, natural scenes such as forests and farmland, and mixed backgrounds such as the coexistence of photovoltaic panels and solar water heaters. To address interference factors, occlusions such as birds and foliage, weather effects such as raindrops and snow cover, and imaging noise such as motion blur and low resolution are introduced to test robustness under extreme conditions. The distribution of different scene types in the validation set is consistent with the training set to avoid evaluation bias; for example, complex backgrounds account for 30%, simple backgrounds for 20%, and natural environments for 50%. All bounding box annotations are double-checked by professional annotators to ensure pixel-level accuracy of photovoltaic panel locations and exclude blurry or disputed samples.

[0116] Step S22: Preprocess the dataset to identify and label targets in the satellite image.

[0117] When applied, the preprocessing of the dataset includes labeling bounding boxes, data augmentation, setting prior boxes, and masking.

[0118] Step S23: Train the target detection model by taking the dataset as input and the preprocessed dataset as output.

[0119] In some embodiments, the annotation of the bounding box includes steps S2211 to S2212.

[0120] Step S2211: Calculate the distribution density and arrangement characteristics of the target objects in the satellite image.

[0121] In application, the distribution density of targets in satellite images can be determined by the number of targets per unit area and / or the average spacing between targets. The arrangement characteristics of targets in satellite images can be determined by analyzing whether they form a regular array and the average spacing between them.

[0122] Step S2212: Based on the distribution density and arrangement characteristics, use the region annotation mode or the independent annotation mode to annotate the bounding box.

[0123] The region labeling mode includes: when the target objects are densely and continuously distributed, a single bounding box is used to cover the continuous distribution area of ​​the target objects; when the target objects are densely arranged in an array, the minimum bounding rectangle is used to cover the array arrangement area of ​​the target objects.

[0124] The independent annotation mode includes: when the target objects are distributed independently or arranged in an independent array, each target object is annotated with a bounding box, and the proportion of the background area within each bounding box is minimized.

[0125] In practical applications, aspect ratio constraints are applied to the marked bounding box, limiting the aspect ratio of the bounding box to a preset range. Preferably, the preset range is 1:1 to 1:2.

[0126] In some embodiments, the data enhancement includes steps S2221 to S2224.

[0127] Step S2221: Based on the distribution density of the target object, the satellite image is cropped preferentially in the distribution area of ​​the target object, so that the obtained cropped image contains at least one target object.

[0128] Step S2222: Perform color balance processing on each cropped image to be stitched together to simulate color distribution under different lighting conditions.

[0129] Step S2223: Multiple cropped images are stitched together according to a preset layout to generate a composite image, and pixel fusion transition processing is performed on the stitching edges of adjacent cropped images to eliminate stitching marks.

[0130] When applied, the default layout is a 2-row, 2-column grid layout.

[0131] Step S2224: Perform enhancement operations on the synthesized image and adjust the synthesized image to the input size required by the target detection model.

[0132] In application, the enhancement operation includes at least one of photometric adjustment, geometric transformation, and spatial deformation. Specifically, photometric adjustment is used to adjust the brightness, contrast, or saturation of the synthesized image; geometric transformation is used to rotate, scale, or translate the synthesized image; and spatial deformation is used to non-rigidly distort the synthesized image.

[0133] Furthermore, the data augmentation may also include: acquiring training batch statistics in real time; increasing the application probability or intensity of geometric transformation operations in the early stage of training to improve the robustness of the model to the target shape and position; and increasing the application probability or intensity of photometric adjustment operations in the later stage of training to enhance the model's adaptability to changes in illumination.

[0134] Furthermore, during image cropping and stitching, the corresponding image space transformation matrix is ​​calculated in real time. Based on the image space transformation matrix, the geometric information of the bounding boxes corresponding to the original training images is updated synchronously to ensure that the bounding boxes are aligned with the transformed image content. Simultaneously, bounding boxes smaller than a preset size threshold caused by image space transformation are identified and filtered to avoid introducing invalid or noisy annotations. The preset size threshold can be obtained statistically, for example, by calculating the area of ​​all target bounding boxes and selecting their 5% to 10% quantiles as the preset size threshold.

[0135] In some embodiments, the setting of the prior box includes steps S2231 to S2233.

[0136] Step S2231: Based on the geometric information of the bounding box, a clustering algorithm is used to analyze the bounding box to obtain several cluster centers.

[0137] When applied, the geometric information includes at least the width and height of the bounding box or the width and height of the target object.

[0138] In practical applications, the geometric information includes the aspect ratio of the bounding box, the logarithmic transformation value of the aspect ratio of the bounding box, and the relative position of the center point of the bounding box in the image; the aspect ratio of the target object, the logarithmic transformation value of the aspect ratio of the target object, and the relative position of the center point of the target object in the image; and the ratio of the area of ​​the target object to the area of ​​its bounding box.

[0139] Step S2232: Use the bounding box size corresponding to the cluster center as the prior box size of the target detection model.

[0140] In application, a multidimensional feature space is constructed, including the target object's width, height, and logarithmic transformation value of its aspect ratio. An improved K-means++ clustering algorithm is then used to perform cluster analysis within this multidimensional feature space. Preferably, the improved K-means++ clustering algorithm uses Mahalanobis distance as the distance metric.

[0141] In practical applications, the multidimensional feature space shall include at least: the relative position of the center point of the target object in the image, and / or the ratio of the effective area of ​​the target object to the area of ​​its bounding box.

[0142] Step S2233: Assign prior boxes to different detection layers of the target detection model according to their size: assign large-scale prior boxes to the target detection model for detecting shallow feature maps of large targets, assign medium-scale prior boxes to the target detection model for detecting medium-sized targets in medium-level feature maps, and assign small-scale prior boxes to the target detection model for detecting deep feature maps of small targets.

[0143] When applying the method, the cluster centers are divided into three categories in descending order of size, corresponding to large-scale prior boxes, medium-scale prior boxes, and small-scale prior boxes, respectively.

[0144] In practical applications, large targets refer to complete photovoltaic arrays or large photovoltaic power plant areas, which have a wide coverage and complete structure in satellite images; medium targets refer to medium-sized photovoltaic groups or parts of photovoltaic arrays; small targets refer to single photovoltaic panels or small-scale photovoltaic groups, which occupy only a few pixels in satellite images and are easily affected by background interference.

[0145] Furthermore, when the training reaches a preset period, the matching degree between the current prior bounding box and the ground truth bounding box is evaluated. If the matching degree is lower than a preset threshold, the corresponding steps of cluster analysis are re-executed based on the current training data to update the prior bounding box size. The preset period can be 20% to 30% of the total training rounds; the preset threshold can be determined based on the average overlap between the prior bounding box and the ground truth bounding box, such as 80% to 90% of the average Intersection over Union (IoU) between the prior bounding box and the ground truth bounding box.

[0146] In some embodiments, the masking process includes:

[0147] Step S22, the preprocessing of the dataset, includes masking, which comprises:

[0148] Step S2241: Calculate at least two land cover indices for the satellite image.

[0149] When applied, the land cover index includes the normalized vegetation index and the normalized building index.

[0150] Step S2242: Generate a comprehensive background suppression mask based on the ground feature index.

[0151] In application, the normalized vegetation index is used to identify vegetation areas and generate a first suppression mask for the vegetation areas; the normalized building index is used to identify building areas and generate a second suppression mask for the building areas; suppression intensity weights are set for the first suppression mask and the second suppression mask respectively, and the first suppression mask and the second suppression mask are fused to obtain a comprehensive background suppression mask.

[0152] Step S2243: Process the satellite image using the integrated background suppression mask to suppress interference from the background region.

[0153] Example 4: Based on any one of Examples 1 to 3, this example introduces a training method for an object detection model. The object detection model can be the object detection model described in Example 1 or the object detection model described in Example 2.

[0154] The training process of the target detection model in this embodiment includes a warm-up phase, a decay phase, and a fine-tuning phase.

[0155] In this embodiment, the learning rate during the warm-up phase starts from an initial learning rate and gradually increases to the maximum learning rate in a non-linear manner.

[0156] In application, if the learning rate is too small, the convergence speed will be very slow. Experiments have shown that the initial learning rate during the warm-up phase should not be less than [a certain value]. Specifically, the initial learning rate during the warm-up phase includes the following formula:

[0157] ;

[0158] In the formula, This is the initial learning rate for the warm-up phase; This is the ratio of the average learning rate to the maximum learning rate during the warm-up phase. This represents the maximum learning rate.

[0159] In practical applications, a higher learning rate in the early stages of training helps the model quickly capture the main features of the data. However, if the learning rate is too large, the model may oscillate around the minimum value or even diverge. To ensure training stability, the learning rate during the warm-up phase includes the following formula:

[0160] ;

[0161] In the formula, This represents the current learning rate during the warm-up phase. The learning rate of the previous learning session during the warm-up phase, when When the learning rate is during the first warm-up phase, for ; This represents the current number of times you have studied during the warm-up phase. This refers to the total number of learning sessions during the warm-up phase, which is generally 1 to 3 times.

[0162] The learning rate in the decay phase of this embodiment is decayed according to the following formula:

[0163] ;

[0164] In the formula, The learning rate during the decay phase; Minimum learning rate; The maximum learning rate; To learn the relative number of times the learning rate decays during the current number of iterations in the decay phase; This represents the total number of learning iterations during the decay phase. This represents the current number of learning iterations during the decay phase.

[0165] In this embodiment, the learning rate during the fine-tuning phase is the minimum learning rate.

[0166] When applied, the total number of learning iterations during the fine-tuning phase includes the following formula:

[0167] ;

[0168] In the formula, This represents the total number of learning iterations during the fine-tuning phase. Total number of learning sessions; This is the ratio of the average learning rate to the maximum learning rate during the fine-tuning phase.

[0169] In practical applications, the total number of learning iterations during the fine-tuning phase ranges from 1 to 15.

[0170] In some embodiments, the impact of the learning rate on the initial stage of learning rate decay is verified by using three phases of learning rate: warm-up, decay, and fine-tuning. It should be noted that the number of learning iterations in each of the three phases refers to the number of training iterations.

[0171] The ratio of the average learning rate to the maximum learning rate during the warm-up phase The ratio of the average learning rate to the maximum learning rate during the fine-tuning phase. When both are 0.05, the number of learning iterations in the warm-up and fine-tuning phases reaches its maximum, at 3 and 15 respectively. At this point, reducing... The mAP was reduced to one-tenth of its original value. In the first 50 training sessions, the mAP decreased by only 0.6%, indicating that the initial learning rate during the warm-up phase had little impact on the early stages of the experiment.

[0172] During the decay phase, the learning rate is primarily affected by the total time spent in the decay period. In other words, given a fixed maximum and minimum learning rate, even with a sufficient number of training iterations, the total decay time has the most significant impact on performance in the early stages of the decay phase. It is 0.002. When the value is 0.05, the number of learning iterations in the warm-up phase is 2, and the number of learning iterations in the fine-tuning phase is 15. The mAP decreases by 2.8%, indicating that the decrease in the total number of decay phases due to the reduction in the number of warm-up iterations or the insufficient number of learning levels during the warm-up phase have a certain impact on the experiment.

[0173] when and When both are 0.005, the warm-up phase learning count is 3, the fine-tuning phase learning count is 5, and the mAP decreases by 6.8%. When the total number of learning counts in the fine-tuning phase is less, the total number of decay counts is greater. Under the same number of training iterations, The larger the value, The bigger it is, but each time The smaller the change, the greater the impact of the learning rate decay on the experiment in the early stages of the decay phase. Its relative size. This also verifies the effect of raising the price. Even after increasing the learning rate by 5 times and reducing the number of warm-up phases by one, there is still a decrease of 0.009, indicating that even a significant increase in the learning rate within a certain range cannot compensate for the impact of the reduced learning rate decay on the experiment.

[0174] This embodiment compares the performance of different YOLOv5 models in photovoltaic (PV) panel detection tasks with the following parameters: traditional YOLOv5, YOLOv5 with globally progressive feature extraction weights in the C3 module, YOLOv5 with locally equal feature extraction weights in the C3 module, and YOLOv5 with globally equal feature extraction weights in the C3 module. Specifically, recall, precision, the harmonic mean of recall and precision (F1-Score), and mean-mean-average precision (mAP) are used to evaluate the performance of the YOLOv5 models. The IoU threshold for mAP is 0.5. Recall reflects the proportion of PV panels correctly detected by the YOLOv5 model out of actual PV panel samples. Precision measures the proportion of samples predicted as PV panels by the YOLOv5 model that are actually PV panels. The F1-Score is the harmonic mean of recall and precision, comprehensively reflecting the overall performance of the YOLOv5 model in PV panel detection tasks.

[0175] Table 1 Performance parameters of different YOLOv5 models

[0176] .

[0177] As shown in Table 1, regardless of whether it's the YOLOv5 model with globally progressive feature extraction weights in module C3, the YOLOv5 model with locally equal weights in module C3, or the YOLOv5 model with globally equal feature extraction weights in module C3, the overall performance of the YOLOv5 model improves to varying degrees after reducing the initial value of the feature extraction weights. This indicates that by reducing the weights of the original features, i.e., the initial value of the feature extraction weights, the YOLOv5 model can focus more on extracting effective features, thus exhibiting stronger performance in the photovoltaic panel detection task. The YOLOv5 model with globally equal feature extraction weights in module C3 shows the best performance across all metrics, indicating that this weight configuration strategy not only helps the YOLOv5 model learn more critical feature information during training but also helps the YOLOv5 model maintain good performance stability when processing complex data. Its performance improvement verifies the importance of appropriately reducing the initial value of the feature extraction weights during training.

[0178] In some embodiments, the impact of color images and black-and-white images on photovoltaic panel recognition performance at different training stages was systematically compared. During the 500-iteration validation phase, the YOLOv5 model with globally equal feature extraction weights in the C3 module performed relatively stably on color images, effectively capturing the color features of photovoltaic panels, such as the reflective properties of the dark blue panel and metal frame, as well as their contrast differences with the surrounding environment. However, when the input image was converted to grayscale, although the YOLOv5 model could still detect some photovoltaic panel targets, the false detection rate increased significantly. These objects were mainly concentrated in two types of areas: firstly, linear structures on building roofs, such as drainage channels or building material joints; and secondly, interference sources with regular geometric shapes, such as glass curtain walls, car roofs, or thermal plants. These misclassifications indicate that, in the absence of color information, the YOLOv5 model over-relies on edge and texture features, leading to the misidentification of certain background objects with similar spatial features as photovoltaic panels.

[0179] As the number of training iterations increased to 1000, a significant improvement in the recognition performance of the YOLOv5 model on black and white images was observed. The YOLOv5 model of this invention not only reduced the number of false alarms, but more importantly, reduced the diversity of false alarm types. Specifically, the common early misidentification of farmland grids or road markings as photovoltaic panels was greatly reduced, retaining only those falsely detected targets that are extremely similar to real photovoltaic panels in shape and reflectivity, such as large solar water heater arrays. It is evident that by globally equalizing the initial values ​​of the feature extraction weights in the C3 module, the YOLOv5 model no longer overemphasizes local high-contrast features, but instead learns to comprehensively evaluate multi-dimensional features such as the target's size, arrangement, and relative positional relationship with the surrounding environment. In contrast, the traditional YOLOv5 model exhibits significant limitations under the same conditions: the detection results not only include targets with spectral characteristics similar to photovoltaic panels, but also generate a large number of false alarm boxes caused by image noise or texture interference, reflecting the inadequacy of traditional detection algorithms in feature decoupling capabilities.

[0180] It is worth noting that while the YOLOv5 model with locally equal weights for feature extraction in the C3 module alleviates the false detection problem to some extent, it still exhibits broken or overlapping detection boxes when processing large-area distributed photovoltaic arrays. In contrast, the YOLOv5 model with globally progressive feature extraction weights in the C3 module demonstrates better scale adaptability, maintaining sensitivity for detecting small targets at a distance while effectively suppressing background interference. The key to the performance improvement of the YOLOv5 model lies in its ability to adjust the contribution of features at different levels. Shallow networks emphasize edge and shape information extraction, while deeper networks enhance the understanding of target structural relationships. The adaptive feature fusion mechanism enables the model to accurately identify photovoltaic panels even without color information assistance, by analyzing the geometric topological features and spatial distribution patterns of the target.

[0181] In some embodiments, comparative experiments with 500 and 1000 iterations are conducted to analyze the differences in photovoltaic panel identification and optimization paths of the YOLOv5 model in residential and natural environments.

[0182] During the 500-iteration phase, the YOLOv5 model of this invention exhibited a significant feature dependency in residential environments. Due to the often dense buildings and severe target occlusion in urban remote sensing images, coupled with the loss of stereoscopic depth caused by low resolution, the YOLOv5 model primarily relied on the two-dimensional shape features of the targets for discrimination. Many objects with regular geometric shapes, such as black rooftops, ventilation ducts, and air conditioner units, were misidentified as photovoltaic panels due to their gray-black tones and rectangular outlines. Particularly noteworthy is that in high-rise building areas, due to perspective distortion and shadow interference, the YOLOv5 model's accuracy in locating small photovoltaic panels significantly decreased, with a false detection rate reaching 32.7%. In natural environment tests, the YOLOv5 model displayed different error patterns: dark areas with linear characteristics, such as gaps in shadows in dense vegetation areas and cracks in exposed rock layers, were frequently misidentified. Analysis of the feature activation map revealed that the YOLOv5 model over-relyed on local contrast features, incorrectly associating strong gradient responses at the boundary between light and dark areas as photovoltaic panel edge features.

[0183] When the number of iterations was increased to 1000, the recognition performance of the YOLOv5 model achieved a qualitative leap. In residential environment testing, the number of false detections decreased by 68%, especially the misclassification rate of easily confused targets such as balcony railings and ventilation shafts was significantly reduced. Through visualization analysis, it was found that the YOLOv5 model of this invention no longer analyzes the shape features of individual targets in isolation, but has learned to make comprehensive judgments by combining contextual information. For example, the YOLOv5 model can recognize the unique regular arrangement patterns of photovoltaic panel arrays, thereby effectively distinguishing randomly distributed similar objects. The improvement in adaptability to natural environments is even more significant. The YOLOv5 model has established a more complete understanding of light and shadow: on the one hand, it enhances the recognition of the surface texture of real photovoltaic panels through multi-scale feature fusion technology; on the other hand, it introduces illumination invariance features, enabling the YOLOv5 model to accurately distinguish between shaded areas and real photovoltaic panels.

[0184] Comparing the performance of the benchmark model YOLOv5 reveals systematic flaws in traditional algorithms under complex environments. YOLOv5 generates numerous discrete false detection boxes in residential environment tests, reflecting its insufficient feature aggregation capabilities. In natural scenes, it exhibits excessive sensitivity to lighting conditions, with false detection rates nearly three times higher during dawn and dusk than at midday. The YOLOv5 model of this invention achieves more robust detection performance by introducing a dynamic feature selection mechanism: in residential areas, it enhances spatial attention weights to focus on capturing the arrangement patterns of photovoltaic panels; in natural areas, it strengthens spectral feature analysis to accurately identify the unique reflective properties of photovoltaic panels.

[0185] In some embodiments, reference Figure 4 The satellite monitoring method in this embodiment identifies and locates photovoltaic panels in satellite images of different geographical environments. These environments can be mountainous, rural, desert, or plains. If any potential fire point pixel is located within the area of ​​a photovoltaic panel, it is determined to have high reflectivity interference and is identified as a false fire point pixel, thus improving the accuracy of wildfire satellite monitoring by 10 percentage points.

[0186] In some embodiments, in response to any potential fire point pixel being located within the area of ​​a thermal plant, it is determined that there is heat source interference, and the potential fire point pixel is determined to be a false fire point pixel.

[0187] Example 5: This example introduces a method for satellite monitoring of wildfires.

[0188] Mainstream volcano satellite prediction models mostly use statistical models, particle algorithms, or deep learning black box models based on historical data. They lack an explanation of the underlying physical mechanisms of fire spread, the prediction process is not transparent, and the causal relationship is unclear, making it difficult for disaster prevention and control personnel to fully trust them and formulate accurate emergency strategies.

[0189] To address the issues of weak interpretability and low accuracy in fire spread prediction using existing wildfire monitoring technologies, this invention introduces a satellite monitoring method and system for wildfires, constructing a causal-driven monitoring framework. Specifically, by building a causal network for fire evolution and clarifying the causal relationships between nodes, inferences are made based on causal relationships rather than simple statistical correlations, ensuring that the prediction process conforms to physical laws; thus fundamentally improving the accuracy and reliability of fire prediction.

[0190] Specifically, based on any one of Embodiments 1 to 4, the wildfire satellite monitoring method of this embodiment further includes step S5. Step S5: Based on the real fire point pixels and their time-series data, a fire point evolution causal network is constructed, and causal inference and prediction of fire point spread are performed based on the fire point evolution causal network. The nodes of the fire point evolution causal network include outcome variables and causal variables, and the outcome variables and causal variables include dynamic causal paths based on causal relationships.

[0191] Furthermore, this embodiment includes steps S51 to S52.

[0192] Step S51: Train the fire point evolution causal network.

[0193] When applied, the training method of the fire point evolution causal network includes steps S511 to S512.

[0194] Step S511: Based on the real fire point pixels and their time series data, construct a fire point evolution causal network.

[0195] In application, time-series data are listed as nodes in the fire point evolution causal network, one by one. The time-series data includes fire point state variables and environmental driving variables. The fire point state variables include fire point location and fire point spread state; the environmental driving variables include time-series meteorological factors, topographic factors, and vegetation state factors. Preferably, the fire point spread state and environmental driving variables are listed as nodes in the fire point evolution causal network, one by one.

[0196] Fire spread status is a quantitative indicator of fire intensity and behavior, which can include fire radiant power, burned area, spread rate, and spread direction. Specifically, fire radiant power is the intensity of energy released by the fire; burned area is the cumulative burning area; spread rate is the distance the fire line advances per unit time; and spread direction is the main direction of fire spread.

[0197] Meteorological factors are weather conditions that change over time, and can include wind speed, wind direction, temperature, humidity, and precipitation. Among these, wind speed and wind direction directly affect the direction and rate of spread; temperature and humidity can affect the dryness of combustibles and combustion efficiency; and precipitation can directly suppress ignition points or reduce the flammability of combustibles.

[0198] Although topographic factors such as altitude, slope, and aspect do not change over time, they act as static nodes in the fire point evolution causal network, continuously connecting with the fire point location node at each time step, thus dynamically influencing the fire point's behavior. For example, fire spreads faster when the fire point is located on an uphill slope. When the fire point at time t moves to a steeper slope, the slope node influences the spread rate at time t+1 through causal relationships. Therefore, the role of topographic factors in the fire point evolution causal network is dynamically activated as the fire point moves.

[0199] Vegetation status factors include vegetation indices and land use types. The vegetation index, such as the Normalized Difference Vegetation Index (NDVI), represents vegetation cover and growth status; areas with abundant vegetation are flammable. Land use types can be forest, grassland, farmland, urban areas, or water bodies; different types of areas exhibit significant differences in ignition probability and combustion characteristics.

[0200] In practical applications, the nodes of the fire point evolution causal network can also include satellite observation data and fire point status.

[0201] When the fire point status is 1, it indicates that the current pixel has a thermal anomaly caused by active combustion and is therefore identified as a fire point. When the fire point status is 0, it indicates that the current pixel has no active combustion; in this case, the satellite observation data is determined by the environmental background and it is identified as a non-fire point. The environmental background can include solar radiation, surface material, etc.

[0202] Step S512: Based on the real fire point pixels and their time series data, learn the dynamic causal paths and causal strengths between the nodes of the fire point evolution causal network, thereby obtaining the trained fire point evolution causal network.

[0203] When applying, refer to Figure 5 A dynamic causal path can be a single directed edge or an indirect causal chain composed of multiple connected directed edges. Directed edges represent the causal relationships between nodes. Furthermore, the weight of a dynamic causal path is its causal strength, which quantifies the magnitude and direction of the influence of the causal variable on the outcome variable.

[0204] In practical applications, step S512 of this embodiment includes steps S5121 to S5123. Further, step S512 includes steps S5124 to S5126.

[0205] Step S5121: Based on domain knowledge, configure a set of potential causal variables for each outcome variable in the fire point evolution causal network. Here, both the potential causal variables and the outcome variables in the set of potential causal variables are nodes in the fire point evolution causal network.

[0206] In application, domain knowledge refers to a priori causal relationships derived from fields such as physics, forestry, and meteorology. For example, rising temperatures and decreasing humidity directly lead to vegetation drying, thus increasing the probability of fire. Land use type determines the amount of combustible material, affecting the fire ignition state. Wind speed disperses smoke above the fire ignition point, thus affecting visible light reflectance. The solar altitude angle affects visible light reflectance, but has no direct causal relationship with the fire ignition state.

[0207] To reduce computational load and focus on key driving forces, avoiding traversing all variable combinations (node ​​combinations) in the fire evolution causal network, this embodiment, based on domain knowledge, configures a corresponding set of potential causal variables for each outcome variable, such as spread rate, spread direction, and fire state. For example: for the outcome variable spread rate, the configured set of potential causal variables includes wind speed, humidity, and slope; for the outcome variable fire radiant power, the configured set of potential causal variables includes burned area, temperature, and vegetation index; for the outcome variable fire state, the configured set of potential causal variables includes air temperature, humidity, vegetation index, and land use type; and for the outcome variable reflectivity, the configured set of potential causal variables includes wind speed.

[0208] Step S5122: Based on the Granger causality test, determine the causal variables and the dynamic causal path from the causal variables to the outcome variables according to the set of outcome variables and their potential causal variables, and add the dynamic causal path as one or more directed edges to the fire point evolution causal network.

[0209] In application, the past information of each potential causal variable in the set of potential causal variables is examined one by one to see if it can cause the present information of the corresponding outcome variable. If it can, then there is a causal relationship between the potential causal variable and the outcome variable. That is, the potential causal variable is a cause of the outcome variable, and there is a dynamic causal path between the causal variable and the outcome variable. For example: the wind speed at t-1 points to the spread direction at t, indicating that the wind speed at t-1 affects the spread direction at t; the humidity at t-2 points to the fire point radiation power at t, indicating that the humidity at t-2 has an inhibitory effect on the fire point intensity at t; the fire point area at t-1 points to the fire point area at t, indicating the time inertia of the fire itself; the air temperature and humidity at t-2 point to the vegetation index at t-1 point, and the vegetation index at t-1 point to the fire point state at t, indicating the physical and airborne inertia of the fire.

[0210] In practical applications, a first model is constructed that assumes past information of the potential causal variable cannot affect the present information of the corresponding outcome variable, and a second model is constructed that assumes past information of the potential causal variable can affect the present information of the corresponding outcome variable. The first model uses only past information of the outcome variable to predict its present information; the second model uses both past information of the outcome variable and past information of the potential causal variable to predict its present information.

[0211] Furthermore, an F-test is used to compare the errors of the first and second models. If the reduction in error of the second model is statistically significant (i.e., p-value less than 0.05), then the potential causal variable is considered a causal variable, meaning there exists a dynamic causal path from the causal variable t-1 to the outcome variable t.

[0212] ;

[0213] In the formula, F represents the error between the first model and the second model, and RSS... R The sum of squared residuals for the first model, RSS F Let df be the sum of squared residuals of the second model. R For the degrees of freedom of the first model, df F For the degrees of freedom of the second model.

[0214] Step S5123: Based on the identification of dynamic causal paths, for each outcome variable, construct a regression model with the historical variables of the cause variables of the outcome variable as independent variables, and quantify the causal strength of the corresponding dynamic causal path according to the regression coefficients of the regression model.

[0215] In application, the causal strength of each dynamic causal path is its weighted weight, used to quantify the magnitude and direction of the influence of the causal variable on the outcome variable. That is, when the causal variable changes by one unit, the resulting variable exhibits the average change. For example, on a dynamic causal path from the wind speed at t-1 to the spread rate at t, if the causal strength is quantified as +0.8, it means that, after controlling for other factors, for every 1 m / s increase in wind speed at t-1, the spread rate at t will increase by an average of 0.8 km / h.

[0216] In practical applications, the regression model is a linear regression model. The least squares method is used to fit the regression model to determine the regression coefficients, and the regression coefficients are used as the causal strength.

[0217] Step S52: Based on the trained fire point evolution causal network, use time series data to perform causal inference and prediction of the fire point spread range at future moments.

[0218] When applied, step S52 includes steps S521 and S522 to reduce the fire spread prediction error by approximately 30%.

[0219] Step S521: Train the fire spread prediction model. Step S521 includes: Step S5211.

[0220] Step S5211: Based on the causal relationships learned by the fire point evolution causal network and historical data, train the fire point spread prediction model.

[0221] In application, the causal relationship includes dynamic causal paths and their causal strength. The historical data includes time-series data of fire point state variables and environmental driving variables collected at fixed time intervals throughout the entire lifecycle of multiple historical fire events.

[0222] In practical applications, input features are weighted according to causal strength to enhance the influence of key causal variables. For example, if the causal strength of wind direction towards spread rate is high, the weight of wind speed in the fire point evolution causal network can be increased. Next, using historical data, the parameters of the fire point evolution causal network are trained with the goal of minimizing the error between the predicted and actual fire point extents. The error between the predicted and actual fire point extents can be measured using cross-entropy loss or Dice loss, which will not be elaborated upon here.

[0223] In practical implementation, the fire spread prediction model's inputs include current fire state variables, current environmental driving variables, and future weather forecast data. The output is a prediction of the fire spread range at future times. The future weather forecast data may include wind speed, wind direction, temperature, humidity, and precipitation.

[0224] Step S5212: Input the current fire point state variables, current environmental driving variables, and future weather forecast data under different scenarios into the trained fire point spread prediction model to obtain the fire point spread range prediction results under various weather conditions.

[0225] When applying the data, the same set of current state data, such as the current fire point state variable and the current environment driving variable, are paired with future weather forecast data under different scenarios and then input into the fire point spread prediction model for forward inference.

[0226] Future weather forecast data under different scenarios are obtained from meteorological departments. For example: normal weather, windy weather, drought without rain, and weather with sudden changes in wind direction. Among them, normal weather is the weather with the highest frequency in historical weather forecast data; the wind speed in windy weather is the 90th percentile value in history; the precipitation in drought without rain is 0; and the wind direction in weather with sudden changes in wind direction is towards residential areas.

[0227] In practical applications, refer to Figure 6The predicted fire spread range includes: a probability map of the future fire spread range corresponding to each weather scenario, where each pixel has the probability of being a fire point at a future time. A probability of [0, 0.2) indicates a low-risk area; a probability of [0.2, 0.5) indicates a medium-risk area; a probability of [0.5, 0.8) indicates a high-low risk area; and a probability of [0.8, 1.0] indicates an extremely high-low risk area.

[0228] Furthermore, the fire spread range prediction results may also include: a future binarized range map corresponding to each weather scenario, where the future binarized range map includes the binarization of the thermal power state of each pixel at a future time. For pixels with a probability greater than or equal to 0.5, the binarization is 1; for pixels with a probability less than 0.5, the binarization is 0.

[0229] Example 6: This example introduces a wildfire satellite monitoring system.

[0230] refer to Figure 7 The wildfire satellite monitoring system in this embodiment includes: an acquisition module, an interference source location module, a potential fire point location module, and an interference source filtering module.

[0231] The acquisition module is used to acquire satellite images of the target area; the interference source localization module is used to identify and locate target pixels containing target objects in the satellite images using a pre-trained target detection model; the potential fire point localization module is used to monitor and locate potential fire point pixels in the satellite images; the interference source filtering module is used to spatially overlay the target pixels and the potential fire point pixels, and in response to any potential fire point pixel overlapping or partially overlapping with the target pixel, the potential fire point pixel is determined to be a false fire point pixel, otherwise the potential fire point pixel is determined to be a real fire point pixel.

[0232] The target detection model can be either the target detection model described in Example 1 or the target detection model described in Example 2.

[0233] In addition, the target object includes interference sources, which include non-fire-point highly reflective artificial ground features and / or fixed heat-generating objects.

[0234] Furthermore, the wildfire satellite monitoring system of this embodiment may also include a prediction module.

[0235] The prediction module is used to construct a fire point evolution causal network based on the real fire point pixels and their time series data, and to perform causal inference and prediction of fire point spread based on the fire point evolution causal network.

[0236] For the specific implementation of the above functional modules, refer to the specific steps of the above-mentioned wildfire satellite monitoring method.

[0237] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for satellite monitoring of wildfires, characterized in that, Includes the following steps: Acquire satellite images of the target area to be analyzed; Using a pre-trained target detection model, target pixels containing target objects in the satellite image to be analyzed are identified and located. The target objects include interference sources, which include non-fire point high reflectivity artificial ground features and / or fixed heat-generating objects. Monitor and locate potential fire pixels in the satellite image to be analyzed; The target object pixel and the potential fire point pixel are spatially superimposed. In response to any potential fire point pixel overlapping or partially overlapping with the target object pixel, the potential fire point pixel is determined to be a false fire point pixel; otherwise, the potential fire point pixel is determined to be a real fire point pixel.

2. The wildfire satellite monitoring method according to claim 1, characterized in that, The training method for the object detection model includes: Construct a dataset that includes training images and validation images, both of which consist of satellite images under diverse geographical environments, lighting conditions, and climate characteristics. The dataset is preprocessed to identify and label targets in the satellite images; The target detection model is trained by taking the dataset as input and the preprocessed dataset as output. The preprocessing of the dataset includes labeled bounding boxes, which include: Calculate the distribution density and arrangement characteristics of the targets in the satellite image; Based on the distribution density and arrangement characteristics, the bounding box is labeled using either the region labeling mode or the independent labeling mode. The region labeling modes include: When the target objects are densely and continuously distributed, a single bounding box is used to cover the continuous distribution area of ​​the target objects; When the target objects are arranged in a dense array, the array area of ​​the target objects is covered by the minimum bounding rectangle. The independent annotation modes include: When the target objects are distributed independently or arranged in an independent array, each target object is labeled with a bounding box, and the proportion of the background area within each bounding box is minimized.

3. The wildfire satellite monitoring method according to claim 2, characterized in that, The preprocessing of the dataset includes data augmentation, which includes: Based on the distribution density of the target object, the satellite image is cropped preferentially in the distribution area of ​​the target object, so that the cropped image contains at least one target object. Each cropped image to be stitched is subjected to color balance processing to simulate color distribution under different lighting conditions; Multiple cropped images are stitched together according to a preset layout to generate a composite image, and pixel fusion transition processing is performed on the stitching edges of adjacent cropped images to eliminate stitching marks. An enhancement operation is performed on the synthesized image, and the synthesized image is adjusted to the input size required by the target detection model; The enhancement operation includes at least one of photometric adjustment, geometric transformation, and spatial deformation.

4. The wildfire satellite monitoring method according to claim 2, characterized in that, The preprocessing of the dataset includes setting prior boxes, which include: Based on the geometric information of the bounding box, a clustering algorithm is used to analyze the bounding box and obtain several cluster centers; The bounding box size corresponding to the cluster center is used as the prior box size of the target detection model; Prior boxes are assigned to different detection layers of the target detection model according to their size: large-scale prior boxes are assigned to the target detection model for detecting shallow feature maps of large targets, medium-scale prior boxes are assigned to the target detection model for detecting medium-scale feature maps of medium targets, and small-scale prior boxes are assigned to the target detection model for detecting deep feature maps of small targets.

5. The wildfire satellite monitoring method according to claim 2, characterized in that, The preprocessing of the dataset includes masking, which comprises: Calculate at least two land cover indices for the satellite image, the land cover indices including the normalized vegetation index and the normalized building index; The system identifies vegetation areas based on the normalized vegetation index and generates a first suppression mask for the vegetation areas; it identifies building areas based on the normalized building index and generates a second suppression mask for the building areas; it sets suppression intensity weights for the first suppression mask and the second suppression mask respectively, and then fuses the first suppression mask and the second suppression mask to obtain a comprehensive background suppression mask. The integrated background suppression mask is used to process satellite images to suppress interference from background areas.

6. The wildfire satellite monitoring method according to claim 2, characterized in that, The target detection model includes the YOLOv5 model, and the loss function of the YOLOv5 model includes the generalized intersection-union loss function, and the weights of the class loss are the lowest level weights; In the global C3 module of the YOLOv5 model, the initial values ​​of the feature extraction weights are all equal. The YOLOv5 model includes a feature selection mechanism, which is used to adjust the feature extraction weights of the C3 module in different detection layers of the YOLOv5 model. When the input is a satellite image lacking color information or a satellite image of the living environment, the feature extraction weight of the shallow C3 module in the backbone network of the YOLOv5 model is increased, so that the YOLOv5 model can enhance its ability to extract edge, shape, texture features and spatial arrangement patterns. When the input is a color image or a natural environment image, the feature extraction weight of the C3 module in the multi-scale feature fusion layer of the YOLOv5 model's neck network is increased, which enhances the YOLOv5 model's ability to extract spectral features and reflectance properties. And / or the C3 module includes a C3_p structure; The p_C3 Block in the C3_p structure includes a first feature extraction branch and a second feature extraction branch connected in parallel. The outputs of the first feature extraction branch and the second feature extraction branch are added together to form the output of the p_C3 Block.

7. The wildfire satellite monitoring method according to claim 2, characterized in that, The training process of the target detection model includes a warm-up phase, a decay phase, and a fine-tuning phase. The learning rate during the warm-up phase starts from an initial learning rate and gradually increases to a maximum learning rate in a non-linear manner. The learning rate during the decay phase is decayed according to the following formula: ； In the formula, The learning rate during the decay phase; Minimum learning rate; The maximum learning rate; To learn the relative number of times the learning rate decays during the current number of iterations in the decay phase; This represents the total number of learning iterations during the decay phase. This represents the current number of learning iterations during the decay phase. The learning rate during the fine-tuning phase is the minimum learning rate.

8. The wildfire satellite monitoring method according to claim 1, characterized in that, Also includes: Based on the real fire point pixels and their time series data, a fire point evolution causal network is constructed, and the causal inference and prediction of fire point spread are performed based on the fire point evolution causal network. The nodes of the fire point evolution causal network include outcome variables and causal variables, and there are dynamic causal paths based on causal relationships between outcome variables and causal variables.

9. The wildfire satellite monitoring method according to claim 8, characterized in that, The causal inference and prediction of fire spread based on the fire point evolution causal network includes: Based on the trained fire point evolution causal network, and using time series data, the causal inference and prediction of the fire point spread range at future moments are performed. The training method for the fire point evolution causal network includes: Based on the real fire points and their time-series data, the dynamic causal paths and causal strengths between the nodes of the fire point evolution causal network are learned, and a trained fire point evolution causal network is obtained. The time-series data includes fire point state variables and environment-driven variables.

10. A wildfire satellite monitoring system, characterized in that, include: The acquisition module is used to acquire satellite images of the target area; The interference source localization module is used to identify and locate target pixels containing target objects in the satellite image using a pre-trained target detection model. The target objects include interference sources, which include non-fire point high reflectivity artificial ground objects and fixed heat-generating objects. A potential fire point location module is used to monitor and locate potential fire point pixels in the satellite image; An interference source filtering module is used to spatially overlay the target object pixel and the potential fire point pixel. In response to any potential fire point pixel overlapping or partially overlapping with the target object pixel, the potential fire point pixel is determined to be a false fire point pixel; otherwise, the potential fire point pixel is determined to be a real fire point pixel.

Images