Patrol early warning method and system based on satellite communication unmanned aerial vehicle

Abstract

The invention belongs to the technical field of forest fire early warning, and particularly relates to a patrol early warning method and system based on a satellite communication unmanned aerial vehicle, and the method comprises the steps: enabling the unmanned aerial vehicle carrying a multi-sensor load to patrol along a preset route, and synchronously collecting thermal infrared, visible light and near-infrared multi-mode data through a unified clock protocol; identifying a candidate region based on temperature anomaly of the thermal infrared image and transmissivity anomaly of the visible light image, then extracting a space heat concentration ratio, a vegetation background imbalance index and a smoke form structure similarity of the candidate region, and constructing a structure feature vector; calculating an environment context weight by combining wind speed, time and candidate area space distribution density; and calculating the confidence coefficient of the core fire point by fusing the structural features and the context weight, and when the confidence coefficient exceeds a threshold value in a continuous time sequence, sending an early warning through the airborne satellite terminal. According to the method, the accuracy and robustness of forest fire identification in a complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of forest fire early warning technology. More specifically, this invention relates to a patrol and early warning method and system based on satellite communication unmanned aerial vehicles (UAVs). Background Technology

[0002] Forest fires pose a significant threat to ecological and environmental security. Utilizing drones equipped with multimodal sensors and satellite communication for real-time patrol and early warning of large-scale forest areas has become a crucial technological means for forest fire prevention. Existing patrol and early warning methods typically employ threshold-based detection algorithms, such as setting temperature thresholds in thermal infrared images to screen for high-temperature anomalies, or using visual features of visible light images to identify suspected smoke areas to determine the presence of a fire. Related technologies include, for example, Chinese patent document CN207199034U, which discloses an aerial forest fire monitoring and early warning system that overlays detected fire information with a geographic information-assisted decision-making system to analyze and process elements collected during fire reconnaissance.

[0003] However, the aforementioned detection methods based on traditional algorithms have certain shortcomings in practical applications. For example, relying solely on temperature thresholds or simple regional average calculations ignores the spatial distribution structure of heat sources, making it difficult to effectively distinguish between real fire points with peak heat distributions and interference sources with flat heat distributions, such as rocks or metal exposed to sunlight. Furthermore, it struggles to capture the specific geometric shapes of smoke conforming to physical diffusion laws. Moreover, analyzing anomalies in isolation at a single moment, without considering the health status of vegetation as a feature, and lacking comprehensive consideration of environmental context information such as wind speed, time of day, and spatial density of anomalies, as well as methods for multi-dimensional structural feature fusion and dynamic environmental perception, can lead to false alarms or missed alarms in complex and ever-changing forest environments, hindering the achievement of highly reliable and accurate early warning systems. Summary of the Invention

[0004] To address the aforementioned technical problem of insufficient reliability in forest fire early warning systems, this invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides a patrol and early warning method based on satellite communication unmanned aerial vehicles (UAVs), comprising: The system conducts patrols along a preset route to collect multimodal data, including thermal infrared images, visible light images, and near-infrared images. Based on the significant temperature characteristics of the thermal infrared images and the significant transmittance characteristics of the visible light images at each time point, a list of candidate anomaly regions is obtained. The spatial heat concentration of the candidate anomaly regions is calculated based on their spatial temperature distribution structure in the thermal infrared images. The vegetation background imbalance index of the candidate anomaly region is obtained based on the NDVI difference between the candidate anomaly region and its closely surrounding annular background region. The smoke morphology and structural similarity of the candidate anomaly regions are obtained based on the morphological and structural differences between the smoke area and the standard trapezoidal template. The spatial heat concentration, the vegetation background imbalance index, and the smoke morphology and structural similarity are combined to form the structural feature vector of the candidate anomaly region. Environmental spatial context weights are obtained based on wind speed data and the spatial distribution density of the candidate region at any given time. Multi-feature fusion is performed based on the context weights and the structural feature vector to calculate the core fire point confidence. An early warning decision is made based on the core fire point confidence, and the early warning information is sent via a satellite communication terminal based on the decision result.

[0006] This invention constructs a multi-dimensional structural feature vector that includes thermal distribution, vegetation health differences, and physical morphology, and combines it with environmental context weights such as wind speed, time, and spatial density for weighted fusion. This effectively distinguishes real fire points from interference sources with similar thermal or visual characteristics, improving the accuracy and robustness of forest fire identification in complex field environments, reducing false alarm rates, and achieving high-precision, robust, real-time early warning based on satellite communication.

[0007] Preferably, obtaining the candidate abnormal region list at each time point includes: For each thermal infrared image, the global average temperature and standard deviation are calculated, and an adaptive thresholding method is used to identify pixel-level thermal anomalies. Then, through connected component analysis, candidate thermal anomaly regions for each thermal infrared image are obtained. For each visible light image, a dehazing algorithm based on dark channel prior is used to calculate the transmittance map of the corresponding visible light image. The transmittance map is then thresholded to identify low transmittance regions as candidate smoke regions for the corresponding visible light image. All candidate thermal anomaly regions and candidate smoke regions are then coordinate-registered and merged to remove duplicates, forming a list of candidate anomaly regions.

[0008] Preferably, the spatial heat concentration satisfies the expression: ; In the formula, Indicates the spatial heat concentration of the target area; This represents the maximum value of the temperature set within the target region; This represents the average value of the temperature set in the target region; This represents the standard deviation of the temperature set in the target region; This represents the first preset minimum positive number.

[0009] The spatial heat concentration calculation method proposed in this invention takes into account both the deviation of the highest temperature from the average temperature and the standard deviation of the temperature distribution. It can accurately capture the peak-shaped distribution characteristics of the fire point, which is extremely hot at the center and decays at the edge. This distinguishes the fire point from the heated rocks or bare soil that are hot but have a flat distribution, thus improving the specificity of thermal anomaly identification.

[0010] Preferably, the acquisition of the vegetation background imbalance index includes: Any candidate abnormal region at any given time is taken as the target region. The average NDVI value of all pixels in the target region is recorded as the first average value. A binary mask image of the target region is obtained. A structuring element with a size proportional to the target region is set, and a morphological dilation operation is performed on the binary mask image to obtain an enlarged outer boundary mask. The binary mask image is removed from the outer boundary mask to obtain the annular background region mask of the target region. The pixel values ​​in the target region that belong to the annular background region mask are extracted, and their arithmetic mean is calculated and recorded as the second average value. The difference between the second average value and the first average value is compared with the sum of the second average value, the first average value, and a second preset minimum positive number to obtain the vegetation background imbalance index of the target region.

[0011] This invention captures the sudden damage to vegetation at the center of a fire point relative to the adjacent background by calculating the NDVI normalized difference between the candidate region and the closely surrounding annular background region. This reduces the interference caused by the overall low background NDVI due to season and soil type, and ensures that the fire point confidence is increased only when the vegetation state is abnormally out of balance with the environment.

[0012] Preferably, the similarity of the smoke morphology and structure includes: The binary mask image of the candidate smoke region of the target area is denoted as the target smoke mask image; a trapezoidal template aligned with the target smoke mask image is constructed. The similarity of the smoke morphology and structure in the target area satisfies the expression: ; In the formula, Indicates the similarity of smoke morphology and structure in the target area; Represents the target smoke mask image; Represents a trapezoidal template aligned with the target smoke mask image; This represents the vertical distance from the bottommost pixel of the target smoke mask image to the centroid. This represents the area of ​​the target smoke mask image; Intersection symbol; This represents the third preset minimum positive number.

[0013] This invention utilizes the physical laws of smoke rising and spreading to construct a trapezoidal template aligned with the target and introduces a root position penalty factor, which can characterize the shape of smoke that is wide at the top and narrow at the bottom with a clear root, enabling the algorithm to filter out interference from irregularly shaped or suspended cloud formations.

[0014] Preferably, constructing the trapezoidal template aligned with the target smoke mask image includes: Calculate the principal axis direction of the target smoke mask image, and rotate the preset standard trapezoidal template along the principal axis direction to obtain a trapezoidal template aligned with the target smoke mask image.

[0015] Preferably, the acquisition of the environmental spatial context weights satisfies the expression: ; In the formula, Indicates the environmental spatial context weights of the target region; Indicates the time corresponding to the target region; This indicates the time corresponding to the highest temperature value monitored by the ground weather station on the day of the test; Indicates the preset standard deviation; This indicates the wind speed corresponding to the target area at that time. Indicates the preset center wind speed; Indicates the number of neighboring regions of the target region; Represents the natural exponential function; This represents the normalization function.

[0016] This invention introduces environmental spatial context weights, uses Gaussian and Sigmoid functions to model the time periods when fires are likely to occur and the wind speed conditions that support combustion, and considers the spatial clustering effect of fire points. This is consistent with the natural laws of fire occurrence, that is, the probability of fire is higher under conditions of high temperature, strong wind and hot spot clustering, thereby dynamically enhancing the detection sensitivity in high-risk scenarios and suppressing random noise under low-risk conditions.

[0017] Preferably, obtaining the number of neighboring regions of the target region includes: Obtain the center coordinates of each candidate anomaly region in the candidate anomaly region list at the corresponding time in the target region, and calculate the distance between the center coordinates of any two candidate anomaly regions; count the candidate anomaly regions that are less than a preset radius threshold from the target region, and record them as the neighborhood regions of the target region.

[0018] Preferably, the core fire point confidence level satisfies the expression: ; In the formula, This indicates the confidence level of the core fire point in the target area; Indicates the environmental spatial context weights of the target region; These represent the spatial heat concentration, vegetation background imbalance index, and smoke morphological structure similarity of the target area, respectively.

[0019] Secondly, the present invention provides a patrol and early warning system based on satellite communication drones, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned patrol and early warning method based on satellite communication drones is implemented.

[0020] By adopting the above technical solution, the patrol and early warning method based on satellite communication UAV is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0021] The beneficial effects of this invention are as follows: This invention constructs a structural feature vector that includes spatial heat concentration representing heat distribution peaks, vegetation background imbalance index representing relative vegetation damage, and smoke morphological structure similarity representing the trapezoidal physical characteristics of smoke columns, thus distinguishing real fire from interference based on physical distribution patterns; at the same time, it introduces environmental context weights composed of time, wind speed, and spatial aggregation to dynamically adjust the confidence calculation; combined with a multimodal spatiotemporal synchronization and continuous time-series verification mechanism, it avoids the problem of high false alarm rates for interference sources such as hot rocks and clouds in complex environments in wide-area forest areas, and realizes high-precision, robust real-time early warning based on satellite communication. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the patrol and early warning method based on satellite communication UAVs in this invention; Figure 2 This is a schematic diagram illustrating a visible light image; Figure 3 This is a schematic diagram illustrating a thermal infrared image; Figure 4 This is a schematic diagram showing the NDVI image corresponding to a visible light image; Figure 5 This is a binary mask image schematically showing the candidate smoke region corresponding to a visible light image; Figure 6 This is a schematic diagram illustrating the alarm target corresponding to the visible light image. Detailed Implementation

[0023] This invention discloses a patrol and early warning method based on satellite communication drones, referring to... Figure 1 This includes steps S1-S4: S1: Based on the preset flight path and triggering mechanism, acquire the airborne synchronous multimodal and multispectral data stream.

[0024] It should be noted that during drone patrol missions in wide-area forest areas without terrestrial network coverage, accurate fire point identification relies on comprehensive analysis of the same ground feature across different physical dimensions, such as thermal radiation, spectral reflectance, and visible light morphology. If the data deviates in time or space, it will lead to errors in subsequent feature fusion, resulting in misjudgments. Considering that airborne multi-sensor payloads can achieve high-precision synchronous data acquisition through hardware or software triggering mechanisms, this invention employs a standardized data acquisition process to ensure the acquisition of highly synchronized, spatially accurate, multimodal data streams, providing high-quality, unambiguous raw input for all subsequent analyses.

[0025] Specifically, based on the preset flight path and triggering mechanism, the airborne synchronous multimodal and multispectral data stream is acquired, including: A drone equipped with a multi-sensor payload flies over the target patrol area along a pre-set route. It should be noted that the multi-sensor payload integrates an infrared thermal imager, a multispectral camera covering visible and near-infrared bands, a high-precision GPS / IMU module, and a satellite communication terminal. The pre-set route is planned based on historical fire point data and forest area risk levels to achieve focused coverage of high-risk areas. The target patrol area is a specific fire prevention responsibility zone designated by the forestry management department.

[0026] A unified high-precision network time protocol ensures that the airborne infrared thermal imager and multispectral camera are triggered synchronously to acquire ground data at a preset frame rate, generating thermal infrared image sequences and multispectral image sequences. For example, the preset frame rate is... Hz.

[0027] A thermal infrared image and a multispectral image taken at the same time are recorded as an image pair. The multispectral image includes both visible light and near-infrared images. It should be noted that, as... Figure 2 This is a schematic diagram of a visible light image, illustrating a visible light image of a forest fire scene, such as... Figure 3 This is a schematic diagram of a thermal infrared image, where the highlighted areas represent high-temperature anomalies.

[0028] Wind direction and speed are obtained from ground weather stations at every moment.

[0029] Thus, a synchronous multimodal data stream containing thermal radiation, multispectral reflectance, visible light vision, and precise spatiotemporal and environmental labels was acquired.

[0030] S2: Based on the significant temperature performance of the thermal infrared image and the significant transmittance performance of the visible light image at each time point, obtain a list of candidate anomalous regions at each time point; calculate the spatial heat concentration of the candidate anomalous regions based on the spatial temperature distribution structure of the candidate anomalous regions on the thermal infrared image; obtain the vegetation background imbalance index of the candidate anomalous regions based on the NDVI difference between the candidate anomalous regions and the closely surrounding annular background regions; obtain the smoke morphology and structural similarity of the candidate anomalous regions based on the morphological and structural differences between the smoke area and the standard trapezoidal template; combine the spatial heat concentration, the vegetation background imbalance index, and the smoke morphology and structural similarity to form the structural feature vector of the candidate anomalous regions.

[0031] It should be noted that traditional thermal imaging-based processes may miss early surface fires due to thermal signal obstruction by trees; simultaneously, traditional regional average calculation methods lose valuable spatial distribution information of fire points or smoke, failing to distinguish real fires from similarly shaped obstructions. Considering that real fire points are usually accompanied by thermal anomalies or smoke, and that both have unique spatial morphology, this invention employs a parallel detection strategy for thermal anomalies and smoke to avoid missed detections, and achieves deep perception of the target's internal structure by analyzing heat concentration and smoke morphology.

[0032] Specifically, based on the significant temperature characteristics of the thermal infrared images and the significant transmittance characteristics of the visible light images at each time point, a list of candidate anomalous regions is obtained for each time point, including: It should be noted that thermal infrared images directly reflect the surface temperature distribution and are sensitive to heat sources; visible light images, on the other hand, can capture the smoke morphology in the early stages of a fire. Therefore, this invention obtains candidate thermal anomaly regions based on the temperature anomaly characteristics of thermal infrared images and candidate smoke regions based on the atmospheric transmittance anomaly characteristics of visible light images.

[0033] For each thermal infrared image, the global average temperature and standard deviation are calculated. An adaptive thresholding method is used to identify pixel-level thermal anomalies. Then, connected component analysis is used to obtain candidate thermal anomaly regions for each thermal infrared image. The adaptive thresholding method classifies pixels with temperatures higher than (global mean + N × standard deviation) as anomalies, where N is an adjustable parameter. The connected component analysis is used to merge spatially adjacent anomalous pixels into independent regions.

[0034] For each visible light image, a dehazing algorithm based on dark channel prior is used to calculate the transmittance map of the corresponding visible light image. Threshold segmentation is then performed on the transmittance map to identify low transmittance regions as candidate smoke areas for the corresponding visible light image. It should be noted that the dehazing algorithm based on dark channel prior is an existing technique that utilizes the statistical regularity of dark pixels commonly found in haze-free images to estimate fog concentration. The threshold segmentation identifies regions with transmittance below a preset value as smoke candidate regions; for example, the preset value is 0.3.

[0035] All candidate thermal anomaly regions and candidate smoke regions are registered and merged to remove duplicates, forming a candidate anomaly region list. It should be noted that the coordinate registration is based on the pre-calibrated intrinsic and extrinsic parameters of the two cameras. Merging and deduplication means that if a thermal anomaly region and a smoke region highly overlap in space, they are merged into a single candidate region.

[0036] At this point, a list of candidate abnormal regions for each time point has been obtained.

[0037] It should be noted that the heat distribution of an early fire point should exhibit a peak-like structure with extremely high central heat and rapid outward decay; while the internal temperature distribution of a uniformly heated background object, such as a metal plate exposed to the sun, is relatively flat. Therefore, this invention uses the concentration of heat distribution in space as the first structural feature of the candidate anomalous region. At the same time, a real fire point will damage the vegetation in the surrounding area, causing the NDVI (normalized difference vegetation index) in the area to be lower than that of the unaffected background environment. Therefore, this invention uses the vegetation damage as the second structural feature of the candidate anomalous region. Finally, considering that smoke columns rise from the fire source and have a cone-shaped or feather-like structure with narrow roots that gradually spread upwards, while natural clouds or fog lack this distinct root structure, this invention uses the structural regularity of smoke as the third structural feature of the candidate anomalous region.

[0038] Any candidate anomaly region at any given time is taken as the target region.

[0039] Preferably, based on the spatial temperature distribution structure of the candidate anomaly region in the thermal infrared image, the spatial heat concentration of the candidate anomaly region is calculated, including: Obtain the set of temperature values ​​corresponding to the target area on the thermal infrared image, denoted as the temperature set of the target area, and obtain the maximum value, average value, and standard deviation of the temperature set of the target area.

[0040] The spatial heat concentration of the target area satisfies the expression: ; In the formula, Indicates the spatial heat concentration of the target area; This represents the maximum value of the temperature set within the target region; This represents the average value of the temperature set in the target region; This represents the standard deviation of the temperature set in the target region; This represents a first preset minimum positive number, used to avoid a denominator of zero. For example, .

[0041] In the formula, This indicates the significance of the highest temperature in the target area relative to the overall fluctuation of the area. The larger the value, the more significant the temperature peak in the target image, and thus the greater the spatial heat concentration in the target area. This represents the ratio of peak temperature to average temperature. The larger the value, the greater the spatial heat difference in the target area, and the higher the heat in the target area, and the higher the spatial heat concentration in the target area. This indicates that when the temperature distribution simultaneously satisfies both high significance and a high proportion of peak structures, the spatial heat concentration in the target area is greater.

[0042] It should be noted that fire spots can damage vegetation in the surrounding area, causing the NDVI of that area to be lower than that of the unaffected background environment. Although the NDVI value of bare soil exposed to sunlight is also low, it is consistent with the background environment, which is also bare soil. Therefore, this invention calculates the vegetation background imbalance index of the target area based on the difference in NDVI between the candidate abnormal area and the ring-shaped background area it closely surrounds.

[0043] Preferably, the vegetation background imbalance index of the candidate anomaly region is obtained based on the NDVI difference between the candidate anomaly region and the closely surrounding annular background region, including: The NDVI map of the target region is calculated using the near-infrared and red bands in the corresponding multispectral image. It should be noted that, for example... Figure 4 This is the NDVI map corresponding to the visible light image, where the red area represents vegetation imbalance.

[0044] Calculate the average NDVI value of all pixels in the target region, and denote it as the first average value.

[0045] Constructing a ring-shaped background region mask for the target region: Obtain a binary mask image of the target region; Set a structuring element with a size proportional to the target region, perform a morphological dilation operation on the binary mask image to obtain an enlarged outer boundary mask; Remove the binary mask image from the outer boundary mask through a Boolean subtraction operation to obtain the ring-shaped background region mask of the target region.

[0046] By using the annular background region mask of the target region, extract all pixel values ​​belonging to the annular background region of the target region on the NDVI map of the target region, and calculate their arithmetic mean to obtain the average NDVI value of the annular background region of the target region, which is denoted as the second average value.

[0047] It should be noted that the greater the difference in NDVI values ​​between the target area and the annular background area, the greater the vegetation background imbalance index of the target area. The vegetation background imbalance index of the target area is calculated accordingly.

[0048] The vegetation background imbalance index of the target area satisfies the expression: ; In the formula, This indicates the vegetation background imbalance index of the target area; This represents the first average value; This represents the second average value; This represents a second preset minimum positive number, used to avoid a denominator of zero. For example, .

[0049] In the formula, The normalized difference in NDVI values ​​between the target area and the annular background area was calculated. The larger the value, the more abnormal the vegetation condition of the target area is relative to its adjacent background, and the higher the probability of fire.

[0050] It should be noted that the shape of the smoke column conforms to the physical law of a trapezoid, which is wider at the top and narrower at the bottom. Its centroid is usually deviated from the geometric center and is closer to the upper part of the diffusion. Therefore, when comparing the smoke area with a standard smoke column template, the greater the difference between the smoke area and the standard smoke column template, and the greater the deviation of the centroid from the geometric center, the greater the similarity of the smoke morphology structure of the target area. Thus, this invention obtains the similarity of the smoke morphology structure of candidate abnormal areas based on the differences in shape and structure between the smoke area and the standard smoke column template.

[0051] Preferably, the similarity of the smoke morphology structure of the candidate anomaly region is obtained based on the difference between the binarized shape of the candidate anomaly region and the shape of the ideal smoke column template, including: The binary mask image of the candidate smoke area in the target region is denoted as the target smoke mask image. The principal axis direction of the target smoke mask image is calculated. A preset standard trapezoidal template is rotated along the principal axis direction to construct a trapezoidal template aligned with the target smoke mask image. It should be noted that the trapezoidal template is an idealized trapezoidal template aligned with the target smoke mask image and conforming to the characteristics of a wide upper base, a narrow lower base, and a coincident geometric center. It should be noted that... Figure 5 This is a binary mask image of the candidate smoke region corresponding to the visible light image.

[0052] Calculate the area of ​​the target smoke mask image and the vertical distance from the bottom pixel to the centroid.

[0053] The similarity of the smoke morphology and structure in the target area satisfies the expression: ; In the formula, Indicates the similarity of smoke morphology and structure in the target area; Represents the target smoke mask image; Represents a trapezoidal template aligned with the target smoke mask image; This represents the vertical distance from the bottommost pixel of the target smoke mask image to the centroid. This represents the area of ​​the target smoke mask image; Intersection symbol; This represents a third preset minimum positive number, used to avoid a denominator of zero. For example, .

[0054] In the formula, This value indicates the degree of overlap between the target smoke mask image and the trapezoidal template. The larger the value, the greater the overlap between the target smoke mask image and the trapezoidal template, and the more similar the target smoke mask image and the trapezoidal template are. In terms of smoke morphology, the target area is more similar to the smoke column. Indicates the penalty factor at the root location. Normalized by the square root of the region size to eliminate scale effects, the penalty factor is close to 1 for smoke columns with a high centroid and a well-defined root, and close to 0 for symmetrical fog patches.

[0055] Specifically, the spatial heat concentration, vegetation background imbalance index, and smoke morphological structure similarity of any candidate abnormal region are used to establish the structural feature vector of the candidate abnormal region.

[0056] At this point, the structural feature vectors of each candidate anomaly region have been obtained.

[0057] S3: Obtain the environmental spatial context weights based on the wind speed data and the spatial distribution density of the candidate area at any time; perform multi-feature fusion based on the context weights and the structural feature vector to calculate the core fire point confidence.

[0058] It should be noted that evaluating structural feature vectors in isolation ignores environmental conditions that contribute to fire occurrence, such as the increased flammability during the afternoon heat and spatial clustering effects, where multiple small fire points clustered together are more dangerous than a single isolated point. Considering that the probability of a fire is closely related to time, weather, and spatial clustering, afternoon heat, strong winds, and fire point clustering are all important indicators of increased fire risk. Therefore, this invention introduces dynamic contextual information to weight the structural feature vectors, thereby achieving more intelligent decision-making that aligns with real-world physical laws.

[0059] Specifically, based on wind speed data at any given time and the spatial distribution density of candidate areas, environmental spatial context weights are obtained, including: It should be noted that the probability of forest fires occurring is not uniformly distributed throughout the day, typically peaking in the afternoon; meanwhile, wind is a key driver of fire spread; furthermore, multiple spatially clustered anomalies are more likely to indicate actual fire conditions than isolated points. Therefore, this invention analyzes these three dimensions separately and integrates them into a unified environmental spatial context weight.

[0060] A Gaussian function is established, with its center representing the highest temperature recorded by a ground weather station for the day, and its standard deviation a preset value. It should be noted that the temperature follows a distribution pattern of "high in the middle and low at both ends," with midday temperatures higher than morning and evening temperatures. Therefore, the Gaussian function is used to reflect the probability of forest fires occurring over time. For example, the highest temperature is 30 degrees Celsius, and the standard deviation reflects the range of peak forest fire periods; for example, the preset standard deviation is 2 hours.

[0061] A sigmoid function is established, which is modeled using the center wind speed. It should be noted that the impact of wind speed on the risk of fire spread exhibits a non-linear growth relationship. At lower wind speeds, the risk increases slowly; however, once the wind speed exceeds a critical value, the risk rises rapidly and eventually tends towards a saturation point. This "S"-shaped growth curve can be fitted using the sigmoid function, and this critical value is the center wind speed.

[0062] Obtain the center coordinates of each candidate anomaly region in the candidate anomaly region list at any given time, and calculate the distance between the center coordinates of any two candidate anomaly regions. Count the candidate anomaly regions whose distance from any candidate anomaly region is less than a preset radius threshold, and denote them as the neighborhood regions of the candidate anomaly region. It should be noted that a single anomaly may be an isolated disturbance, while multiple spatially clustered anomalies are highly likely to indicate a developing fire or fire cluster. Therefore, the importance of a candidate anomaly region should increase with the number of other candidate anomaly regions in its neighboring area.

[0063] The environmental spatial context weights of the target region satisfy the expression: ; In the formula, Indicates the environmental spatial context weights of the target region; Indicates the time corresponding to the target region; This indicates the time corresponding to the highest temperature value monitored by the ground weather station on the day of the test; Indicates the preset standard deviation; This indicates the wind speed corresponding to the target area at that time. Indicates the preset center wind speed; Indicates the number of neighboring regions of the target region; Represents the natural exponential function; This represents the normalization function.

[0064] In the formula, the first term In the context of the target area, when the temperature at a given time is close to the time corresponding to the highest temperature of the day, then... The smaller, The closer it is to 1, the closer it is to the probability of a wildfire occurring over time; the second term... The Sigmoid function implements a soft threshold switching effect, depending on the wind speed. Much smaller than the center wind speed When the wind speed approaches 0, this term has a negligible contribution to the weighting; when the wind speed... Much larger When this term approaches 0, the denominator approaches 1, and the entire fraction approaches 1, thus significantly weighting the strong winds that contribute to combustion; the third term In this context, the environmental spatial context weight of the target region increases with the number of neighboring regions of the target region, indicating that the more neighboring regions the target region has, the more concentrated the fire points become.

[0065] At this point, the environmental spatial context weights of the candidate anomaly regions have been obtained.

[0066] Preferably, the core fire point confidence is calculated by fusing the context weights and the structural feature vectors, including: It should be noted that in the structural feature vector In the middle, spatial heat concentration Similarity to the morphological structure of smoke It is direct physical evidence of the fire point, while the vegetation background imbalance index It serves as corroborating evidence of the environmental impact of the fire. Therefore, direct evidence should be used as the basis for judgment, enhanced by supporting environmental evidence, and finally weighted by the overall environmental and spatial context.

[0067] The confidence level of the core fire point in the target area satisfies the expression: ; In the formula, This indicates the confidence level of the core fire point in the target area; Indicates the environmental spatial context weights of the target region; These represent the spatial heat concentration, vegetation background imbalance index, and smoke morphological structure similarity of the target area, respectively.

[0068] In the formula, The structured information of heat and smoke, the two main fire indicators, is fused with equal weights, so that a high base score can be generated when either feature is significant. The presence of significant vegetation damage amplifies the confidence level, while in non-vegetated areas such as bare soil... Approximately equal to This does not have any impact, thus effectively avoiding excessive reliance on vegetation; on this basis, the greater the environmental spatial context weight of the target area, the greater the confidence of the core fire point in the target area.

[0069] At this point, the core fire point confidence of each candidate anomaly region has been obtained.

[0070] S4: Based on the confidence level of the core fire point, make early warning decisions and send early warning information through satellite communication terminals based on the early warning decision results.

[0071] It should be noted that, in order to ensure the ultimate reliability of the early warning and to make effective use of the valuable bandwidth of satellite communication, the final early warning information will only be sent when the confidence level of the core fire point in the candidate anomaly area continuously and stably exceeds the preset threshold.

[0072] Specifically, based on the confidence level of the core fire point, an early warning decision is made, and based on the early warning decision result, early warning information is sent through a satellite communication terminal, including: It should be noted that real fire events are continuous in time. Therefore, this invention uses temporal consistency verification to filter out instantaneous high-score interference, thereby minimizing false alarms to the greatest extent.

[0073] For the center coordinates of each candidate anomaly region, if the center coordinates are within the candidate anomaly region for K consecutive time intervals, and the confidence level of the corresponding core fire point is greater than a preset alarm threshold, then the center coordinates are designated as an alarm target, where K is a preset number of confirmation frames. For example, the preset alarm threshold is 0.75, and K is 5. It should be noted that... Figure 6 This is a schematic diagram of alarm targets corresponding to a visible light image, showing multiple monitored alarm targets in the visible light image and their corresponding core fire point confidence levels.

[0074] The alarm information is sent to the ground command center via an airborne satellite communication terminal. The alarm information includes the geographical coordinates of the alarm target, the confidence level of the core fire point, the timestamp, environmental data, and the corresponding image slices.

[0075] Thus, the patrol and early warning system based on satellite communication drones was completed.

[0076] This invention also discloses a patrol and early warning system based on satellite communication drones, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the patrol and early warning method based on satellite communication drones according to this invention is implemented.

[0077] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0078] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A patrol and early warning method based on satellite communication unmanned aerial vehicles (UAVs), characterized in that, include: The patrol area is patrolled along a preset route and multimodal data is collected, including thermal infrared images, visible light images and near-infrared images. Based on the significant temperature characteristics of the thermal infrared images and the significant transmittance characteristics of the visible light images at each time point, a list of candidate anomalous regions is obtained for each time point. The spatial heat concentration of the candidate anomalous regions is calculated based on their spatial temperature distribution structure in the thermal infrared images. The vegetation background imbalance index of the candidate anomalous regions is obtained based on the NDVI difference between the candidate anomalous regions and the closely surrounding annular background regions. The morphological and structural similarity of the smoke areas in the candidate anomalous regions is obtained based on the morphological and structural differences between the smoke areas and the standard trapezoidal template. The spatial heat concentration, the vegetation background imbalance index, and the morphological and structural similarity of the smoke areas are combined to form the structural feature vector of the candidate anomalous regions. Based on wind speed data at any given time and spatial distribution density of candidate areas, environmental spatial context weights are obtained; based on the context weights and the structural feature vector, multi-feature fusion is performed to calculate the core fire point confidence. Based on the confidence level of the core fire point, an early warning decision is made, and based on the result of the early warning decision, the early warning information is sent through a satellite communication terminal.

2. The patrol and early warning method based on satellite communication UAV according to claim 1, characterized in that, The process of obtaining the candidate abnormal region list at each time point includes: For each thermal infrared image, the global average temperature and standard deviation are calculated, and an adaptive thresholding method is used to identify pixel-level thermal anomalies. Then, through connected component analysis, candidate thermal anomaly regions for each thermal infrared image are obtained. For each visible light image, a dehazing algorithm based on dark channel prior is used to calculate the transmittance map of the corresponding visible light image. The transmittance map is then thresholded to identify low transmittance regions as candidate smoke regions for the corresponding visible light image. All candidate thermal anomaly regions and candidate smoke regions are then coordinate-registered and merged to remove duplicates, forming a list of candidate anomaly regions.

3. The patrol and early warning method based on satellite communication UAVs according to claim 1, characterized in that, The spatial heat concentration satisfies the expression: ； In the formula, Indicates the spatial heat concentration of the target area; This represents the maximum value of the temperature set within the target region; This represents the average value of the temperature set in the target region; This represents the standard deviation of the temperature set in the target region; This represents the first preset minimum positive number.

4. The patrol and early warning method based on satellite communication UAV according to claim 1, characterized in that, The acquisition of the vegetation background imbalance index includes: Any candidate abnormal region at any given time is taken as the target region. The average NDVI value of all pixels in the target region is recorded as the first average value. A binary mask image of the target region is obtained. A structuring element with a size proportional to the target region is set, and a morphological dilation operation is performed on the binary mask image to obtain an enlarged outer boundary mask. The binary mask image is removed from the outer boundary mask to obtain the annular background region mask of the target region. The pixel values ​​in the target region that belong to the annular background region mask are extracted, and their arithmetic mean is calculated and recorded as the second average value. The difference between the second average value and the first average value is compared with the sum of the second average value, the first average value, and a second preset minimum positive number to obtain the vegetation background imbalance index of the target region.

5. The patrol and early warning method based on satellite communication UAV according to claim 4, characterized in that, The similarity of the smoke morphology and structure includes: The binary mask image of the candidate smoke region of the target area is denoted as the target smoke mask image; a trapezoidal template aligned with the target smoke mask image is constructed. The similarity of the smoke morphology and structure in the target area satisfies the expression: ； In the formula, Indicates the similarity of smoke morphology and structure in the target area; Represents the target smoke mask image; Represents a trapezoidal template aligned with the target smoke mask image; This represents the vertical distance from the bottommost pixel of the target smoke mask image to the centroid. This represents the area of ​​the target smoke mask image; Intersection symbol; This represents the third preset minimum positive number.

6. The patrol and early warning method based on satellite communication UAV according to claim 5, characterized in that, The construction of the trapezoidal template aligned with the target smoke mask image includes: Calculate the principal axis direction of the target smoke mask image, and rotate the preset standard trapezoidal template along the principal axis direction to obtain a trapezoidal template aligned with the target smoke mask image.

7. The patrol and early warning method based on satellite communication UAV according to claim 4, characterized in that, The obtained environmental spatial context weights satisfy the expression: ； In the formula, Indicates the environmental spatial context weights of the target region; Indicates the time corresponding to the target region; This indicates the time corresponding to the highest temperature value monitored by the ground weather station on the day of the test; Indicates the preset standard deviation; This indicates the wind speed corresponding to the target area at that time. Indicates the preset center wind speed; Indicates the number of neighboring regions of the target region; Represents the natural exponential function; This represents the normalization function.

8. The patrol and early warning method based on satellite communication UAV according to claim 7, characterized in that, Obtaining the number of neighboring regions of the target region includes: Obtain the center coordinates of each candidate anomaly region in the candidate anomaly region list at the corresponding time in the target region, and calculate the distance between the center coordinates of any two candidate anomaly regions; count the candidate anomaly regions that are less than a preset radius threshold from the target region, and record them as the neighborhood regions of the target region.

9. The patrol and early warning method based on satellite communication UAV according to claim 4, characterized in that, The confidence level of the core fire point satisfies the expression: ； In the formula, This indicates the confidence level of the core fire point in the target area; Indicates the environmental spatial context weights of the target region; These represent the spatial heat concentration, vegetation background imbalance index, and smoke morphological structure similarity of the target area, respectively.

10. A patrol and early warning system based on satellite communication unmanned aerial vehicles (UAVs), characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the patrol and early warning method based on a satellite communication UAV according to any one of claims 1-9.

Images