A method and system for multi-stage confirmation of small target fire points in forest areas

Abstract

The present application relates to the technical field of fire intelligent monitoring, and particularly provides a small target fire point multi-level confirmation method and system for forest areas, which comprises the following steps: acquiring multi-source image data required for target area monitoring of forest fire points; adopting a three-level screening mechanism to reprocess the preprocessed multi-source image data; the first level screening is used to quickly extract potential fire point candidate areas in the multi-source image data in a complex forest environment; the second level screening is used to perform spatio-temporal consistency analysis on the potential fire point candidate areas, and to determine whether they have flame dynamic change characteristics; the third level screening is used to fuse visual semantic information and external environmental information, to perform authenticity determination on the potential fire point candidate areas after spatio-temporal consistency analysis, to eliminate false detection targets, and to output high-confidence fire points; the target areas confirmed after the three-level screening mechanism are determined as real fire points, and the real fire point detection results are output in a structured manner. The present application enhances the early warning and auxiliary decision-making capabilities for forest fires.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire monitoring technology, and in particular to a method and system for multi-level confirmation of small target fire points in forest areas. Background Technology

[0002] Forest fires, as a natural disaster characterized by their sudden onset, rapid spread, and difficulty in extinguishing, seriously threaten ecological security and the safety of people's lives and property. Currently, forest fire monitoring methods mainly include satellite remote sensing, video surveillance, drone patrols, and ground sensors; among these, image and video-based fire point identification technology has become an important technical approach for early fire warning due to its high real-time performance and flexible deployment. However, existing image and video-based fire point identification methods still face the following technical shortcomings in practical applications: First, in complex forest environments, such as mountainous forest areas, densely vegetated areas, or urban-rural fringe areas, there are a large number of interference sources with visual characteristics similar to flames. Existing detection models have difficulty distinguishing them effectively, which can easily lead to false detections and reduce the reliability of the system.

[0003] Existing technologies typically employ fire detection methods based on fixed surveillance cameras or video stream image analysis, using traditional image processing methods or deep learning models to detect and identify flame areas. Examples include background modeling, frame difference methods, or convolutional neural networks and YOLO-based target detection models. However, these methods have significant limitations in complex forest environments. First, the complex forest background contains numerous interfering targets similar in color or brightness to flames, such as sunlight reflections, vehicle lights, building lighting, and water reflections, easily leading to false detections. Second, early-stage fire points are usually small targets with extremely low pixel counts in long-distance monitoring scenarios, making feature extraction by detection models difficult and prone to missed detections. Furthermore, relying on single-frame or short-term information lacks full utilization of the dynamic characteristics of flame changes, thus reducing the stability and robustness of the detection.

[0004] Second, early-stage fire points typically appear as small, low-feature targets, accounting for a low percentage of pixels in images or videos, and are significantly affected by factors such as smoke obscuring, changes in lighting, and shooting distance. Existing deep learning detection models, such as the YOLO series, lack the ability to extract features from small fire points in complex backgrounds, easily leading to missed detections or unstable recognition results.

[0005] With the development of artificial intelligence technology, deep learning-based fire detection methods have gradually become a research hotspot. Existing methods typically employ target detection models such as YOLO and Faster R-CNN, using large-scale data training to achieve flame area recognition and improve detection accuracy to some extent. However, in practical applications in complex forest areas, the following problems still exist: First, the ability to detect small target fires is limited, and feature representation is insufficient in long-distance or low-resolution scenarios; second, the models rely heavily on static image training, making it difficult to effectively model the dynamic changes of flames over time; third, they are prone to misjudgment under complex background interference, resulting in a high false alarm rate; and fourth, the models are highly dependent on computing resources, making efficient deployment and real-time operation difficult in edge devices or environments with limited communication.

[0006] Third, existing methods mostly rely on single-frame images or static spatial features for discrimination, failing to fully utilize the inherent temporal variation characteristics of flames, such as brightness fluctuations, morphological evolution, and irregular flickering dynamic information; single-frame analysis is difficult to effectively distinguish between real flames and static fire-like interference objects, which restricts further improvement in recognition accuracy.

[0007] Fourth, in the integrated air-space-ground forest fire monitoring system, forest areas often have communication blind spots and limited network bandwidth, which leads to delays or loss of images, videos and multi-source sensor data during transmission, affecting the real-time performance and stability of fire point identification algorithms and reducing the availability of the system under complex communication conditions.

[0008] Existing technologies utilize multi-source sensing devices such as satellites, drones, and ground sensors, but these systems still suffer from problems such as fragmented data links, strong communication dependencies, and insufficient intelligent analysis capabilities. In forest areas with communication blind spots or unstable networks, multi-source data transmission experiences delays or packet loss, affecting the real-time performance of fire identification. Furthermore, data processing often relies on a centralized cloud-based model, resulting in weak edge-side intelligent analysis capabilities and difficulty in achieving rapid local screening and multi-level confirmation of fire points. In addition, the lack of a unified fusion and collaboration mechanism between different data sources leads to insufficient adaptability of the overall system in complex scenarios.

[0009] Furthermore, while satellite remote sensing technology boasts advantages such as wide coverage and the ability to achieve large-scale monitoring, it suffers from drawbacks including low temporal resolution and significant susceptibility to weather conditions, making it difficult to provide real-time early warnings of fires. Although drone inspections can acquire high-resolution images, limitations in endurance, flight range, and mission scheduling frequency hinder all-weather, full-area coverage monitoring. Moreover, in terms of multi-source data fusion, existing methods typically involve simple data overlay or feature stitching, lacking a unified spatiotemporal modeling mechanism. This makes effective collaboration between data from different sources difficult, thus affecting the accuracy and consistency of fire point identification.

[0010] In summary, how to achieve high-precision identification and multi-level confirmation of small target fire points under complex forest environments and limited communication conditions, reduce false alarm and false negative rates, and improve the overall robustness of the system are technical problems that urgently need to be solved in the field of intelligent fire monitoring. Summary of the Invention

[0011] To achieve the above objectives, the present invention adopts the following technical solution: One aspect of the present invention provides a method for multi-level confirmation of small target fire points in forest areas, comprising the following steps: To acquire multi-source image data required for monitoring target areas of forest fires; The multi-source image data includes satellite remote sensing image data, UAV inspection video data, real-time video stream data collected by ground monitoring camera equipment, infrared data, and meteorological information, etc.; each data source has different resolutions, imaging angles, and lighting conditions.

[0012] A three-level screening mechanism is used to reprocess the preprocessed multi-source image data; The first-level screening is used to quickly extract potential fire point candidate regions from multi-source image data in complex forest environments; the second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate regions to determine whether they have dynamic flame change characteristics; the third-level screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire point candidate regions after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points. The target area confirmed by the three-level screening mechanism is determined to be the real fire point, and the detection results of the real fire point are output in a structured manner.

[0013] Another aspect of the present invention provides a multi-level confirmation system for small target fire points in forest areas, comprising: The image data acquisition module is used to acquire multi-source image data required for monitoring target areas of forest fires; The filtering and processing module is used to reprocess the preprocessed multi-source image data using a three-level filtering mechanism; The first-level screening is used to quickly extract potential fire point candidate regions from multi-source image data in complex forest environments; the second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate regions to determine whether they have dynamic flame change characteristics; the third-level screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire point candidate regions after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points. The real fire point determination module is used to determine whether the target area after being confirmed by the three-level screening mechanism is a real fire point, and to output the real fire point detection results in a structured manner.

[0014] This invention presents a rapid fire point candidate detection technology for complex forest environments, addressing the challenges of small target size, weak image features, and significant differences in data quality acquired by different sensing devices in the early stages of fire detection. First, it performs unified preprocessing on satellite remote sensing, UAV video, and ground monitoring images, improving input data quality through image enhancement, noise suppression, and multi-scale adaptation. A lightweight target detection model is then used to rapidly locate potential fire point regions within the images, employing a high-recall candidate extraction strategy to prioritize areas where fire points may exist, providing candidate targets for refined confirmation. This invention also includes a fire point authenticity verification technology based on spatiotemporal dynamic consistency. Addressing the limitations of existing methods that rely on single-frame images and are susceptible to influences from static light sources and reflective areas, this invention introduces temporal dimension information, performing continuous frame analysis on candidate fire point regions. By extracting brightness, motion, and morphological changes in candidate regions over time, it establishes differences between the dynamic behavior of flames and background interference targets, enabling further judgment of candidate target authenticity. This shifts the judgment process from traditional single-image assessment to assessment based on spatiotemporal behavioral features, improving the stability of fire point identification in complex environments.

[0015] This invention presents a high-reliability fire point confirmation technology based on the fusion of semantic and environmental information. Addressing the issue of misjudgment easily occurring with a single visual model in complex forest environments, it fuses visual semantic recognition results with external environmental information. By combining multi-source data such as infrared thermal anomaly information, meteorological conditions, and forest geographic information, it comprehensively evaluates candidate fire points, ensuring that the fire point confirmation process simultaneously satisfies the rationality of visual features and environmental conditions, thus improving the reliability of the final judgment. This invention also addresses edge intelligent processing technology for integrated air-space-ground forest fire monitoring, tackling communication blind spots and limited data transmission in forest areas. It deploys a three-level fire point confirmation method on edge computing nodes, enabling rapid local analysis and filtering of perceived data. By completing candidate discovery and multi-level confirmation at the data source end, only high-confidence fire point results are uploaded to the intelligent forest fire assessment platform, reducing communication pressure and improving fire response speed in complex environments.

[0016] The spatial output and intelligent analysis technology for fire point identification results of this invention not only completes the detection and confirmation of fire points, but also outputs the confirmed fire point information in a structured manner, including information such as fire point location, time, and confidence level. It also combines Beidou positioning and forest fire monitoring platform to realize the visualization of fire information and assist decision-making. Through the above technical means, a complete technical closed loop is realized from the discovery of suspected targets to the confirmation of real fire points and then to supporting fire prevention decisions. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of the multi-level confirmation method for small target fire points in forest areas provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the second-level screening principle provided in Embodiment 5 of the present invention; Figure 3 This is a schematic diagram of the third-level screening principle provided in Embodiment 6 of the present invention; Figure 4 This is a schematic diagram of the multi-level confirmation method for small target fire points in forest areas provided in Embodiment 1 of the present invention; Figure 5 This refers to the multi-level confirmation method and system for small target fire points in forest areas provided in Embodiment 8 of the present invention; Figure 6 A block diagram of the electronic device provided by the present invention; Figure 7 A block diagram of a computer-readable storage medium provided for this invention.

[0018] Reference numerals: 1. Image data acquisition module; 2. Filtering and processing module; 3. Fire point detection module; 4. Central processing unit / microprocessor / main control chip; 5. Storage medium; 6. Data bus; 7. Input / output bus / external bus / device bus; 8. Display; 9. Input / output device; 10. Computer-readable instructions; 11. Non-transitory computer-readable storage medium. Detailed Implementation

[0019] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, "connection" can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0022] In this embodiment of the invention, directional terms such as "up," "down," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0023] This invention acquires multi-source image data required for monitoring target areas of forest fires; it employs a three-level screening mechanism to reprocess the pre-processed multi-source image data; the first level of screening is used to quickly extract potential fire candidate areas from multi-source image data in complex forest environments; the second level of screening is used to perform spatiotemporal consistency analysis on potential fire candidate areas to determine whether they possess dynamic flame change characteristics; the third level of screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire candidate areas after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points; the target areas confirmed by the three-level screening mechanism are determined to be real fire points, and the real fire point detection results are output in a structured manner. To address the challenges of monitoring small target fire points in complex forest environments, such as susceptibility to interference, high false detection rates, and weak early fire point features making stable identification difficult, a three-level confirmation method for small target fire points in complex forest environments is proposed. This method constructs a hierarchical discrimination mechanism based on image preprocessing, including candidate fire point discovery, spatiotemporal consistency verification, and semantic and environmental fusion confirmation. This gradually reduces detection uncertainty, achieving high recall and high precision in fire point detection, thereby improving the accuracy, robustness, and real-time early warning capabilities of fire point identification in complex forest environments.

[0024] This invention presents a first-level screening method based on multi-source image data preprocessing and rapid candidate fire point discovery. It improves the consistency of image data from different sources by preprocessing satellite remote sensing images, UAV video data, and ground monitoring video data. Furthermore, it uses target detection methods to rapidly extract potential small target fire point areas in complex forest backgrounds, forming a set of candidate fire point areas. The invention also presents a fire point authenticity verification method based on spatiotemporal dynamic consistency. For candidate fire point areas, it utilizes continuous time series information to comprehensively analyze the motion, brightness, and morphological changes of candidate targets, establishing a dynamic verification mechanism distinct from static background interference targets, enabling further screening of suspected fire point targets. Finally, the invention presents a three-level fire point confirmation method based on the fusion of visual semantic information and environmental information. It integrates the visual semantic recognition results of candidate fire point areas with infrared thermal information, meteorological environmental information, and geospatial information. Through multi-source evidence joint judgment of candidate target authenticity, it achieves highly reliable fire point confirmation in complex forest environments. This invention presents a three-tiered, progressively converging fire point confirmation process, consisting of candidate discovery, dynamic verification, and fusion confirmation. Following a method of coarse detection, dynamic screening, and comprehensive confirmation, it gradually improves the confidence level of fire point judgment, transforming fire point identification from a single-model judgment to multi-stage evidence fusion confirmation. This enhances the accuracy and robustness of identifying small-target fire points in complex environments. The invention also addresses the edge deployment and result output mechanism of an integrated air-space-ground forest fire monitoring system. It deploys the three-tiered screening and confirmation method on edge computing nodes and integrates with BeiDou positioning, satellite communication, or terrestrial networks to output structured information such as the confirmed fire point location, time, and confidence level. This information is used for data access and fire early warning for the intelligent forest fire assessment platform.

[0025] Example 1: As Figure 1 As shown, this embodiment of the invention provides a method for multi-level confirmation of small target fire points in forest areas, comprising the following steps: Step S100: Obtain multi-source image data required for monitoring target areas of forest fires; The multi-source image data includes satellite remote sensing image data, UAV inspection video data, real-time video stream data collected by ground monitoring camera equipment, infrared data, and meteorological information, etc.; each data source has different resolutions, imaging angles, and lighting conditions. Step S200: The preprocessed multi-source image data is reprocessed using a three-level screening mechanism; The first-level screening is used to quickly extract potential fire point candidate regions from multi-source image data in complex forest environments; the second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate regions to determine whether they have dynamic flame change characteristics; the third-level screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire point candidate regions after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points. Step S300: The target area confirmed by the three-level screening mechanism is determined to be a real fire point, and the real fire point detection results are output in a structured manner.

[0026] Forest fire monitoring is a detection and early warning activity targeting combustion phenomena caused by natural or human factors in forests and forest edges. It focuses on the early detection and confirmation of small-scale fires, differing from conventional large-scale fire monitoring. Its core challenge lies in the small size of the targets, their weak features, and complex background interference. The target area is the spatial range of the forest area to be monitored, specifically defined by geographical coordinates. Depending on the monitoring task, the target area can be a single key forest area, a forest grid unit, or a detection range determined by satellite remote sensing coverage and UAV patrol routes. Its size directly affects the acquisition strategy and selection parameter configuration of multi-source image data. Multi-source image data is a collection of visual and non-visual data from different imaging platforms and sensing modalities, including at least satellite remote sensing image data, UAV inspection video data, real-time video stream data collected by ground monitoring cameras, infrared thermal imaging data, and meteorological information data. These data sources differ in resolution, imaging angle, lighting conditions, and spectral bands. By jointly utilizing these multi-source data, the information deficiencies of a single data source in small target fire detection are compensated for, providing an input basis for three-level screening. Complex forest environments are spatial scenarios within the target area characterized by diverse vegetation types, dramatic terrain undulations, variable lighting conditions, and interference from artificial facilities in urban-rural fringe areas. In such environments, numerous interference sources exist that exhibit visual similarities to flames in color, shape, and brightness, such as sunlight reflection, water surface reflection, and surface thermal anomalies. These significantly interfere with fire detection algorithms and are the main causes of false positives and false negatives for small target fires. Potential fire point candidate regions are sets of local image regions that may contain fire points, initially identified from multi-source image data after the first-level screening. These regions are represented by bounding boxes or pixel masks, characterized by a large number of regions and a high recall requirement. A certain proportion of false positives is allowed, aiming to reduce the data processing load for subsequent screenings while maintaining sensitivity to small fire points. Spatiotemporal consistency analysis, performed in the second-level screening, involves joint analysis of the same candidate region across consecutive time-series image frames. By comparing the evolution of parameters such as positional shift, area change, brightness fluctuation, and edge deformation at different times, it determines whether the target region exhibits the continuous dynamic changes characteristic of flames, rather than the occasional similarity of static interference in a single frame. The dynamic characteristics of flames are the inherent non-steady-state time-varying properties of real flame combustion, including but not limited to irregular fluctuations in brightness intensity, irregular movement of the flame front, pulsating expansion and contraction of the area, continuous deformation of the edge contour, and randomness of the flicker frequency; the core criterion for spatiotemporal consistency analysis is the essential attribute that distinguishes small target fire points from static or quasi-static interference objects.Visual semantic information refers to high-level features extracted from image data that describe the visual attributes of a target, including color distribution statistics, texture roughness, morphological structure regularity, and contrast with the surrounding background. External environmental information refers to environmental constraints independent of image data, including infrared temperature data, meteorological parameters, wind speed, humidity, temperature, terrain slope and aspect data, historical fire point records, and vegetation type distribution. These two types are integrated in the third-level screening process to achieve a comprehensive determination of the authenticity of candidate regions. Authenticity determination, based on visual semantic features and environmental constraints, involves binary classification of candidate regions that have passed the spatiotemporal consistency screening, classifying them as either real fire points or false fire point interference items. Authenticity determination employs a multi-source evidence collaborative confirmation approach; upgrading fire point identification from a single visual judgment to a comprehensive decision-making process combining environmental logical constraints, it is the final checkpoint before outputting high-confidence fire points. False positives are non-fire points that are incorrectly identified as fire points during the entire detection process. These primarily fall into two categories: first, static interference that visually resembles flames, such as reflections from leaves under sunlight, bright areas on rock surfaces, and water reflections; second, non-fire heat sources that possess certain dynamic characteristics but do not conform to the fire point change patterns, such as temporary man-made heat sources and animal activity traces. False positives are eliminated through a three-level screening process: the first level retains recall, the second level removes static interference, and the third level removes items with conflicting environmental logic. Confidence-based fire points are the fire point detection results output after the complete three-level screening process. Each output fire point is accompanied by a corresponding confidence score, which comprehensively reflects the target's feature response strength in the first-level screening, the dynamic feature matching degree in the second-level screening, and the probability output of multi-source information fusion discrimination in the third-level screening. Confidence-based fire points are divided into different levels according to scoring thresholds to prioritize fire incidents and support dynamic adjustments to the multi-level confirmation strategy.

[0027] For details on the principles described in the above embodiments, please refer to the appendix. Figure 4 This embodiment employs a three-level screening mechanism to accurately identify and suppress false alarms of small target fire points in complex forest environments. It acquires target area information through multi-source image data, providing a data foundation for subsequent processing. The first level of screening uses image processing algorithms to quickly locate potential fire point candidate areas, solving the initial detection challenge of small target fire points in forest backgrounds. The second level of screening uses a time-series analysis method to enhance the ability to distinguish fire points from interference objects through flame dynamic feature recognition. The third level of screening introduces multimodal data fusion technology, combining environmental parameters and visual features to establish a fire point authenticity discrimination model, significantly reducing the false alarm rate caused by vegetation reflection, heat source interference, etc. The final structured detection results meet the dual requirements of accuracy and timeliness for forest fire point monitoring.

[0028] Compared with existing forest fire point identification methods, this embodiment has the following advantages: It improves the detection capability of small target fire points in complex forest environments; addressing the problems of small target size, low pixel ratio, and unclear early fire characteristics in forest fire areas, it enhances the quality of input data through image preprocessing and employs a candidate fire point discovery mechanism to quickly extract potential targets, improving the ability to detect early, small fire points while ensuring detection efficiency and reducing missed detections due to excessively small target size. It also reduces false detections caused by complex background factors, improving the reliability of fire point identification; addressing sunlight reflection, water surface reflection, and other high-brightness interference targets in forest environments, it utilizes a spatiotemporal dynamic consistency screening method, leveraging motion changes, brightness fluctuations, and morphological changes in continuous image sequences to effectively eliminate static interference targets, reducing misjudgments caused by relying solely on single-frame image recognition. To improve the accuracy of fire point confirmation, multi-source information fusion is achieved. Building upon visual recognition, infrared thermal information, meteorological environmental information, and forest geospatial information are further integrated. By combining semantic features with environmental constraints, candidate fire points are comprehensively judged, transforming the fire point confirmation process from a single visual judgment to collaborative confirmation based on multi-source evidence, thus improving the stability of fire situation assessment in complex scenarios. A progressively converging fire point confirmation mechanism is constructed to enhance system robustness. A three-level screening mechanism—candidate discovery, spatiotemporal dynamic verification, and semantic environment fusion confirmation—gradually improves the reliability of the fire point judgment process from coarse to fine. Each level of screening addresses different types of errors, avoiding the problem of false alarms and false negatives that are prone to occur with single detection models in complex backgrounds, thereby improving the overall recognition robustness. Adapting to integrated air-space-ground forest fire monitoring applications and enhancing real-time processing capabilities, this system can be deployed as an edge computing node within the integrated air-space-ground forest fire monitoring system. It enables rapid analysis of satellite remote sensing, drone inspections, and ground video data. In forest environments with limited communication or unstable networks, it reduces a significant amount of invalid data transmission, improving the efficiency of fire information processing. Combined with BeiDou positioning and communication links, it facilitates rapid reporting of fire point information, providing technical support for forest fire early warning and emergency response. It enhances early fire warning and decision support capabilities by outputting structured information including fire point location, discovery time, and confidence level. This provides reliable fire data to forest fire prevention platforms, enabling rapid fire point location, risk analysis, and subsequent emergency dispatch, supporting the shift from passive response to proactive prevention of forest fires.

[0029] Example 2: Based on Example 1, the process of obtaining multi-source image data required for monitoring forest fire points in step S100 of this embodiment of the invention specifically includes the following steps: Step S101: Denoise the multi-source image data by using bilateral filtering or non-local mean denoising to suppress high-frequency interference information introduced by smoke, haze and sensor noise in the forest environment; and reduce background noise while preserving flame edge features to obtain a denoised image. Step S102: Perform illumination adaptive enhancement processing on the denoised image. Use the limited contrast adaptive histogram equalization method to enhance low-brightness areas and suppress the influence of overexposed areas to obtain an enhanced image. Step S103: Perform color space conversion and normalization on the enhanced image, convert the original RGB image to HSV or YCbCr color space to enhance the expressive power of the flame color dimension, and perform unified normalization on the pixel distribution of different data sources to meet the scale consistency requirements, forming a standardized feature tensor for the first-level screening to extract potential fire point candidate regions.

[0030] Among them, bilateral filtering or non-local mean denoising methods are denoising algorithms or bilateral filtering based on weighted filtering of pixel spatial neighborhood and gray-level similarity, respectively, and denoising algorithms or non-local mean denoising based on non-local averaging based on the similarity between image blocks. These two methods are optional denoising techniques used to replace conventional linear filtering. Their characteristic is that they can suppress noise while maintaining the edge structure in the image without blurring, and are particularly suitable for high-fidelity processing of flame edges and background vegetation boundaries in forest images. Smoke, haze, and sensor noise are unique interference factors in forest environments. Smoke is suspended particulate matter generated in the early stages of a fire or combustion in adjacent areas, appearing as locally semi-transparent blurred areas in the image; haze is a global contrast reduction phenomenon caused by water vapor and aerosols in the atmosphere; sensor noise is electrical noise generated by imaging equipment under low illumination or high gain conditions, including thermal noise, readout noise, and quantization noise. These factors together constitute the sources of image degradation that need to be suppressed in preprocessing. High-frequency interference refers to signal components in an image whose pixel values ​​change drastically in the spatial domain, corresponding to high-frequency components in the frequency domain. In forest fire monitoring scenarios, high-frequency interference mainly consists of random textures or noise generated on the image by smoke particle scattering, sensor thermal noise, light flicker, and leaf swaying. This interference can mask the weak features of small target fires and needs to be suppressed during the denoising process. Flame edge features are the morphological attributes of the spatial transition boundary between the real flame and the background area, including the gradient intensity of the edge, irregular contour shape, and the gradual color change at the edge. In small target fire scenarios, the flame size is extremely small, and its edge features occupy a very limited number of pixels in the image. This is a key criterion for distinguishing flames from interference in each stage of filtering. Therefore, the preprocessing stage must prioritize protecting this feature from excessive smoothing by filtering operations. The denoised image is the image data output after the original multi-source image data has been processed by bilateral filtering or non-local mean denoising methods. High-frequency random interference in the image has been effectively suppressed, while flame edge features and background structure information are preserved to the greatest extent, serving as input data for adaptive illumination enhancement processing. Illumination-adaptive enhancement processing is a process that dynamically adjusts the pixel grayscale mapping relationship based on the brightness distribution characteristics of local areas of the image. This processing aims to solve the problem of uneven image brightness in forest scenes caused by daylight cycles, tree canopy occlusion, and strong local reflections, ensuring that small target fire points in low-light areas have sufficient grayscale dynamic range for effective response by detection algorithms. Contrast-limited adaptive histogram equalization is an adaptive enhancement algorithm that performs histogram equalization on image blocks and limits the contrast gain within each block. By setting a contrast limiting threshold, it avoids the excessive amplification of noise in uniform areas while enhancing low-brightness areas, thus improving the distinguishability of small target fire points while controlling enhancement side effects.Low-brightness areas are localized areas in forest images where pixel grayscale values ​​are significantly lower than the global mean. These areas typically correspond to tree canopy shadows, forest floor, valleys, and areas with insufficient light during dawn and dusk. Small flames in the early stages of a fire are easily obscured by the low-contrast background in these areas, making them key areas requiring enhancement. Overexposed areas are negatively impacted by strong direct sunlight, water reflection, or highly reflective objects on the ground, causing some pixels to saturate their grayscale values. Specifically, pixel information within these areas is constrained to its maximum value, resulting in a loss of internal texture details. Adaptive lighting enhancement requires suppressing overexposed areas to prevent fire detection algorithms from misidentifying overexposed edges as bright flame areas during subsequent color space conversion. Color space conversion and normalization: The former refers to the transformation operation that converts an image from its original RGB color representation to other color coordinate systems; the latter refers to the process of linearly or non-linearly mapping pixel values ​​of each channel to a unified numerical range. The combined application of these two methods aims to unify the color expression scale of images from different data sources, eliminate feature distribution drift caused by differences in imaging platforms, and provide a consistent data input format for the first-level screening. The original RGB image is a color image data composed of three channels: red, green, and blue. It is the original color format directly output by most optical imaging devices. The original RGB image is the input form before color space conversion. There is a high correlation between the three channels, which limits its color discrimination ability when directly used for fire detection. Therefore, it needs to be converted to a color space that is more suitable for expressing flame characteristics. HSV or YCbCr color space: The HSV color space describes color information in three dimensions: hue, saturation, and lightness. The hue channel can effectively separate the flame color from the background vegetation color. The YCbCr color space describes color information in terms of lightness component and blue and red chromaticity components. The red chromaticity component has a high responsiveness to the red and hot areas of the flame. Both spaces are optional conversion targets for RGB images to enhance the discriminability of flames in the color dimension. Enhancing the expressive power of flame color dimensions involves color space transformation. This allows for a more significant quantitative description of the flame's unique color distribution, such as the clustering characteristics of orange-red and yellow regions in the chromaticity space, within the transformed coordinate system. This improves the separability of flame and non-flame pixels in the color feature space, providing more discriminative color information input for feature extraction at each stage of the selection process. Pixel distribution refers to the statistical distribution characteristics of each pixel in the image's numerical space, including the mean, variance, histogram shape, and covariance structure of each channel. Images from different data sources, such as satellites, drones, and ground monitoring equipment, exhibit different pixel distribution characteristics. Normalization is necessary to control these differences within an acceptable range for subsequent models.Scale consistency requires that images from all data sources input to the first-level screening model meet uniform standards in terms of numerical range, size specifications, and channel order. Normalization maps pixel values ​​from different sources to the same numerical range, such as [0,1] or [-1,1], ensuring consistent data distribution during training and inference phases and avoiding performance degradation in candidate region detection due to differences in data scale. The standardized feature tensor is a structured multidimensional data array output after denoising, illumination enhancement, color space conversion, and normalization. It has uniform dimensions, consistent pixel value range, and standardized color representation, and can be directly used as input data for the first-level screening model for rapid extraction of potential fire point candidate regions.

[0031] In the above embodiments, the multi-source image denoising processing of this embodiment suppresses high-frequency interference information introduced by smoke, haze, and sensor noise through bilateral filtering or non-local means algorithms. While preserving the flame edge features, it reduces background noise and provides a foundation for image data with a high signal-to-noise ratio. It effectively solves the problem of weak features of small target fire points being submerged due to high-frequency interference, ensuring that key discriminative features such as the gradient intensity and contour shape of the flame edge are preserved. The illumination adaptive enhancement adopts a limited contrast adaptive histogram equalization method to dynamically compensate for the uneven illumination problem unique to forest areas. While enhancing low-brightness areas such as tree canopy shadows and forest floor, it suppresses the influence of overexposed areas, expands the grayscale dynamic range of small target fire points in low-light environments, avoids the noise amplification or detail loss problems caused by traditional enhancement methods, and significantly improves the distinguishability of fire points from the background. Color space conversion and normalization convert RGB images to HSV or YCbCr spaces, enhancing the linear separability of flame color features. The hue channels in the HSV space effectively separate flame and vegetation colors, while the red chromaticity component in the YCbCr space enhances the responsiveness to red-hot regions. Combined with cross-data source pixel distribution normalization, it eliminates feature distribution differences caused by different imaging devices, ensuring scale consistency and providing standardized input with a unified color expression scale for the first-level screening model. The standardized feature tensor, as the final output of the entire process, integrates noise-preserving edge features, illumination-enhanced contrast features, and color space-converted enhanced chromaticity features. The standardized characteristics of the tensor in dimension, numerical range, and color expression enable the first-level screening model to efficiently extract potential fire candidate regions based on consistent discrimination criteria. This embodiment addresses key challenges in small target fire detection, such as low signal-to-noise ratio, uneven illumination, and color interference. Through the chain-like synergy of technical features at each step, it constructs a front-end data processing system that meets the needs of multi-level confirmation.

[0032] Example 3: Based on Example 1, step S100 of this embodiment of the invention can also perform multi-scale pyramid construction processing on the image. By generating image layers of different resolutions, the subsequent candidate fire point detection module can simultaneously acquire feature information of distant small target fire points and close-range large-scale flame targets, thereby improving the detection capability of small target fire points.

[0033] Gaussian pyramid downsampling is performed on the standardized feature tensor, and pixels are extracted by alternating rows and columns after Gaussian kernel convolution to generate image levels with decreasing resolution layer by layer, thus obtaining a multi-scale image pyramid. Each level of the multi-scale image pyramid is input into the feature extraction network, and feature response calculations are performed at each level under different receptive fields. Fire candidate features at each scale are extracted to form a multi-scale feature map set. The features at each level in the multi-scale feature map set are fused across scales. High-level semantic information and low-level detail information are integrated by feature overlay or attention weighting to form a multi-scale fused feature tensor, which is used for the first-level screening to jointly extract potential fire point candidate regions at different scales.

[0034] The Gaussian pyramid downsampling process involves applying Gaussian smoothing filters to each layer of the image, followed by pixel extraction at preset intervals to reduce image resolution. This process constructs a hierarchical structure from high to low resolution, ensuring that each layer corresponds to a forest scene at different spatial scales, balancing detailed information of distant, small fire targets with overall structural information of nearby, large-scale flame targets. Gaussian kernel convolution is a weighted summation operation using a Gaussian-distributed convolution kernel on image pixels and their neighborhoods. As an anti-aliasing process before downsampling, Gaussian kernel convolution suppresses spectral aliasing caused by high-frequency components such as tree canopy textures and leaf edges during resolution reduction, ensuring information fidelity at each layer after downsampling. Interlaced pixel extraction is a pixel filtering method that samples one pixel value at intervals of one or more pixels in the horizontal and vertical directions. This operation, combined with Gaussian kernel convolution, constitutes the core step of pyramid downsampling, systematically removing some pixels to progressively reduce image size and generate a sequence of image layers with decreasing resolution. The image hierarchy, with decreasing resolution, consists of multiple image versions that progressively reduce spatial resolution from the original image resolution. Higher levels retain fine spatial information of small fire targets at original or near-original resolution for near-range detection, while lower levels, after downsampling, expand the receptive field for far-range detection. All levels collectively cover the range of scale variations that fire targets may occur in forest scenes. The multi-scale image pyramid is a set of images arranged from highest to lowest resolution, formed by Gaussian kernel convolution and interleaved extraction from the same image. The pyramid structure provides multi-resolution input for the first-level screening, enabling candidate fire point detection to simultaneously respond to fire targets of different physical sizes in the image, particularly addressing the difficulty of detecting small fire targets at long distances in forest areas due to their extremely low pixel ratio. The feature extraction network is a trained deep neural network structure that can compute and output feature map representations from image inputs. Taking images from each level of the multi-scale image pyramid as input, the feature extraction network transforms pixel-level information into high-dimensional feature representations through basic operations such as convolution, activation, and pooling, providing subsequent candidate fire point detection modules with the ability to locate potential fire areas at different scales. Different receptive fields represent the spatial range of the input image corresponding to a single neuron on the feature map in the feature extraction network. After different levels of the multi-scale image pyramid are input into the feature extraction network, the physical receptive fields corresponding to the same convolutional layer at each level differ. The receptive field of the low-resolution level covers a larger actual spatial range of the forest area, while the receptive field of the high-resolution level focuses on a smaller local area. The two complement each other to achieve multi-scale perception of fire targets.Feature response calculation involves inputting image data into a feature extraction network and then calculating the activation values ​​at corresponding locations on the feature maps of each layer through forward propagation. Feature response calculation is performed independently for each level of the pyramid. High-response regions reflect the spatial locations of potential fire candidate targets at different scales, providing a basis for judging the saliency of targets at each level for subsequent cross-scale fusion. Fire candidate features are sets of feature map responses calculated by the feature extraction network at different image scale levels, related to fire targets in terms of color, texture, shape, and spatial distribution. Fire candidate features at each scale focus on fire targets of different sizes. Higher-level features are sensitive to small flame edges and color abrupt changes, while lower-level features are sensitive to the contextual information of the surrounding environment. Cross-scale fusion processing involves integrating the feature map sets obtained from feature extraction at each level of the multi-scale image pyramid while maintaining their spatial correspondence. This process combines high-level detailed features with low-level contextual features, enabling the fused feature representation to simultaneously possess the ability to finely distinguish small target fire points and the ability to globally understand the forest scene background. Feature overlay refers to directly adding or stitching feature maps at different scales element-wise at the same spatial resolution; attention weighting refers to adaptively modulating feature responses at different scales using learned weight coefficients. These two methods are optional implementations of cross-scale fusion. The former is simple and efficient, while the latter can adaptively emphasize the contribution of the most advantageous scale features in the current forest scene. The specific choice can be flexibly configured according to the computing resources of the deployed equipment. High-level semantic information and low-level detail information: High-level semantic information refers to the feature responses output by the deeper layers of the feature extraction network, which has a large receptive field and forest scene category discrimination power, reflecting abstract information such as the presence of fire targets and their rough location, fire situation, etc.; low-level detail information refers to the feature responses output by the shallower layers of the feature extraction network, which has high spatial resolution and accurate edge localization capabilities, reflecting specific information such as the precise contour, boundary clarity, and color distribution of small target fire points. The two are integrated through cross-scale fusion processing, so that the output candidate regions of the first-level screening have both accurate spatial localization accuracy and reliable category judgment criteria.

[0035] In the above embodiments, this embodiment uses multi-scale pyramid construction to sequentially complete the transformation process from single-scale feature tensors to multi-scale image hierarchy construction, and then to multi-scale feature extraction and fusion, so that the first-level screening can simultaneously detect distant small target fire points and close-range large-scale flame targets at different spatial resolutions.

[0036] Example 4: Based on Example 1, the first-level screening process in step S200 provided in this embodiment of the invention specifically includes the following steps: Step S201: Extract candidate fire point regions from the preprocessed standardized feature tensor, use a lightweight target detection model to perform dense prediction on the image, and implement a relaxed detection strategy by setting a confidence threshold lower than the conventional detection threshold to prioritize the recall integrity of potential fire point targets, thereby obtaining an initial candidate region set. Step S202: Spatial constraint filtering is performed on the initial candidate region set to retain regions that meet the basic flame visual attribute discrimination conditions. The discrimination conditions include pixel brightness threshold determination, red-orange channel response intensity determination, and abnormal brightness change degree of local region relative to background determination. Non-target regions that do not have obvious flame optical characteristics are eliminated to form a candidate fire point region set.

[0037] The relaxed detection strategy involves setting the confidence threshold lower than the standard threshold in conventional deployment scenarios during the object detection model's inference process, thereby enabling the model to output more potential detection results. The relaxed detection strategy uses a low confidence threshold, ensuring the model outputs all regions in the image with weak visual flame response features, without directly rejecting them due to low confidence. The purpose of the relaxed detection strategy is to maximize the initial recall rate of small target fire points, avoiding missed detections due to weak initial fire point features and low confidence responses, while allowing a certain proportion of false positives in the output results, which are then gradually removed through second and third-level filtering. Pixel brightness threshold determination refers to a filtering operation that compares the brightness components of each pixel in a candidate region of an image. Specifically, after color space conversion, the value of each pixel in the brightness channel is extracted and compared with a preset brightness threshold. Only connected regions composed of pixels whose brightness values ​​exceed the threshold are retained, while candidate regions with low overall brightness or that do not conform to the characteristics of flame emission are eliminated. The brightness threshold is dynamically adjusted according to the overall lighting conditions of the forest scene to adapt the determination results to factors such as day-night changes and shadow occlusion. The purpose of the determination is to utilize the physical characteristics of flame emission to quickly eliminate low-brightness non-fire background areas such as vegetation, soil, and tree trunks in forest images. Red-orange channel response intensity determination refers to a filtering operation that quantitatively evaluates and compares the response intensity of candidate regions in the red and orange spectral bands with thresholds. Specifically, based on the converted color space, the response value of each pixel in the candidate region in the red channel, as well as the ratio of red to green or red to blue channels, are calculated. These response values ​​are compared with a preset red-orange response threshold, and only candidate regions with red channel responses significantly higher than those in the green and blue channels are retained. The determination criteria are set to a red channel value greater than that of the green channel and greater than that of the blue channel, and a red-to-green or red-to-blue channel ratio greater than a preset ratio threshold. The purpose of this determination is to utilize the spectral distribution characteristics of flames in the visible light band, which are predominantly red-orange, to eliminate bright interfering objects in the forest environment that do not have a red-orange advantage.

[0038] Candidate fire point region extraction refers to the process of inputting the preprocessed standardized feature tensor into the target detection model, and locating and outputting the location information of local regions that may contain fire points in the image space through the model's forward inference calculation. As the initial operation of the first-level screening, candidate fire point region extraction's core task is to achieve rapid coarse localization of potential small target fire points, outputting a set of bounding box coordinates or a set of pixel masks, providing initial input for subsequent spatial constraint filtering. Lightweight target detection models refer to target detection neural network structures with fewer network layers, lower parameter count, and lower computational complexity, including but not limited to lightweight versions of the YOLO series and detection structures built based on lightweight backbone networks. The purpose of using lightweight target detection models is to adapt to the limited computing resources of forest edge computing nodes, completing the initial localization of potential fire points in the image while ensuring the detection speed meets real-time requirements. The detection accuracy is compensated for by subsequent relaxed detection strategies and spatial constraint filtering. Dense prediction refers to the object detection model's prediction method in image space, which uses a sliding window approach or an anchor point mechanism to determine the existence of targets and perform boundary regression at each spatial location. Dense prediction comprehensively covers all spatial regions in the image that may contain fire targets, without relying on pre-set specific region priors, thus avoiding missed detections due to the high randomness of small target fire location. The standard detection threshold refers to the confidence score threshold used by the object detection model in a standard deployment scenario to determine the validity of the detection results. The standard detection threshold is the confidence judgment standard under the model's default or standard configuration. The first-level screening implements a lenient detection strategy by lowering the confidence threshold below the standard detection threshold, allowing weak fire responses that would otherwise be excluded due to low confidence to be retained, thereby maximizing the initial recall rate. Spatial constraint filtering refers to a post-processing operation that removes regions from the initial candidate region set that do not meet the basic visual characteristics of a flame, based on the inherent physical properties of the flame target in image space. Spatial constraint filtering does not rely on temporal information or external environmental data; it only utilizes the static features of the flame in a single frame image in spatial dimensions such as brightness, color, and contrast for rapid screening. Its purpose is to significantly reduce the number of candidate regions with extremely low computational cost, alleviating the processing burden of subsequent spatiotemporal consistency analysis. Basic flame visual attribute discrimination conditions refer to a set of quantitative criteria used to determine whether a candidate region possesses the most basic visual representation of a real flame in a static image. These criteria specifically include pixel brightness threshold judgment, red-orange channel response intensity judgment, and judgment of the degree of abnormal brightness change in a local region relative to the background. These discrimination conditions are set based on the physical luminescence characteristics and spectral distribution features of the flame. All three are applied in parallel and must be satisfied simultaneously. This is used to quickly remove falsely detected regions that do not possess the basic optical characteristics of a flame from the candidate set, retaining regions with the basic visual attributes of a flame for the second-level screening and dynamic feature analysis.The determination of abnormal brightness abrupt changes in local areas relative to the background refers to a filtering operation that quantifies and compares the contrast difference between the candidate area and the surrounding forest background in brightness space. Specifically, a background ring is formed by expanding a preset pixel range outward from the boundary of the candidate area. The average brightness of the pixels inside the candidate area and the average brightness of the pixels in the background ring are calculated separately. The difference between the two is taken as the brightness abrupt change amplitude value, and this value is compared with a preset abrupt change threshold. Only candidate areas with brightness abrupt change amplitude exceeding the threshold are retained. The determination condition is set to require the candidate area to have a significant brightness jump feature relative to the surrounding background. The purpose of the determination is to use the significant difference in brightness between the flame and the forest background to eliminate false detection areas in the image with little difference in brightness from the surrounding background, and retain the real fire point candidate areas with local high brightness abrupt change characteristics.

[0039] In the above embodiments, the first-level screening of this embodiment achieves maximum initial recall of small target fire points in forest areas under limited computing power by combining a lightweight target detection model with a relaxed detection strategy. The relaxed detection strategy with a lower confidence threshold effectively captures weak fire point features and solves the problem of missed detection of initial fire points caused by traditional threshold settings. The multi-dimensional visual attribute discrimination mechanism in spatial constraint filtering, including dynamic threshold determination of pixel brightness, determination of red-orange channel response intensity, and analysis of local brightness mutation, works synergistically to eliminate typical interference objects in forest areas. Brightness determination is adaptively adjusted according to scene lighting to exclude low-brightness vegetation, soil, and other backgrounds; the red-orange channel ratio threshold filters bright objects with non-flame spectral features; and the analysis of local brightness mutation eliminates false detection areas that lack optical contrast with the surrounding environment by quantifying the brightness gradient difference between the candidate area and the background ring. The multi-condition parallel screening mechanism of static spatial features achieves efficient compression of the number of candidate areas at the single-frame processing level, reducing the computational load.

[0040] This embodiment establishes a feature enhancement detection chain for small target fire points in forest areas: lenient detection ensures initial capture of weak targets, while multi-dimensional spatial constraints ensure strict matching of static optical features. The combination of a lightweight model structure and adaptive threshold determination satisfies the real-time requirements of edge devices while compensating for the inherent limitations of model accuracy through multi-level feature verification. Addressing key challenges in forest fire point detection, such as weak small target features, complex background interference, and limited edge computing power, a progressive filtering mechanism that balances recall and precision is established, providing high-quality initial input for spatiotemporal consistency analysis.

[0041] Example 5: Figure 2 As shown, based on Example 1, the second-level screening process in step S200 provided in this embodiment of the invention specifically includes the following steps: Step S203: For each candidate region in the candidate fire point region set output by the first-level screening, perform target association and tracking operations between consecutive time series image frames. Calculate the positional overlap and feature similarity of candidate regions in adjacent frames. Temporally associate the detection results of the same target at different times to construct short temporal image sequences corresponding to each candidate region, and obtain the associated candidate temporal sequence set. Step S204: For each sequence in each candidate temporal sequence set, extract the brightness change features, motion features, and morphological change features of the candidate region in the time dimension. The brightness change features are quantified by calculating the temporal fluctuation amplitude of the average pixel brightness within the candidate region in each frame of the sequence. The motion features are extracted by calculating the pixel-level optical flow field distribution of the candidate region between adjacent frames in the sequence. The morphological change features are quantified by calculating the deformation degree of the contour of the candidate region between adjacent frames in the sequence. Combine the three features into a dynamic feature descriptor for the candidate region to obtain a candidate temporal sequence set with dynamic feature descriptors. Step S205: Based on the brightness change features, motion features, and morphological change features of each sequence in the candidate time series sequence set with dynamic feature descriptors, calculate the dynamic consistency score corresponding to each feature, and perform a weighted summation of the brightness score, motion score, and morphological score to obtain the comprehensive spatiotemporal consistency score of each candidate region, and generate a candidate sequence set with a comprehensive score; wherein the comprehensive score is used to characterize the comprehensive confidence level of the candidate region having flame dynamic change features. Step S206: Compare the comprehensive score of each candidate region in the candidate sequence set with the comprehensive score with the preset dynamic judgment threshold. When the comprehensive score of the candidate region reaches or exceeds the threshold, the candidate region is marked as having passed the spatiotemporal consistency verification and retained as a real fire point candidate. Otherwise, the candidate region is judged as a static interference target and removed, forming a set of verified candidate fire point regions and outputting it to the third-level screening. The removed candidate regions will not enter the subsequent process.

[0042] The target association and tracking operation refers to the process of establishing the correspondence between the same physical target among the candidate regions in the candidate fire point region set output by the first-level screening in continuous time series image frames. Specifically, for continuous frame images in forest area monitoring video stream, the correspondence between each candidate target at different times is determined by matching the candidate regions located in the previous frame with the candidate regions detected in the current frame, thereby realizing continuous tracking of the same candidate fire point target in the time dimension. The purpose of the operation is to provide time-aligned candidate region data for dynamic feature analysis and avoid the failure of dynamic feature extraction due to the correspondence error between target frames. Location overlap and feature similarity are two quantitative indicators used to determine whether two candidate regions in adjacent frames belong to the same target. Location overlap is quantified by calculating the intersection-union ratio (IUU) of the bounding boxes of candidate regions between adjacent frames, which is the ratio of the intersection area to the union area of ​​two regions. A higher ratio indicates a more consistent spatial location. Feature similarity is quantified by calculating the distance between candidate regions in adjacent frames based on color histograms, texture features, or depth features. A smaller distance indicates a more consistent visual feature. The two are used together to form a dual criterion for target association. Location overlap is used to handle cases where the target moves slowly, while feature similarity is used to handle cases where the target's appearance changes slightly due to changes in forest lighting or partial occlusion. The detection result refers to the candidate fire point region detection data output after the first-level screening processes each frame of image, including the bounding box coordinates, pixel mask, detection confidence score, and corresponding frame number and timestamp information of the candidate region. The detection result provides the basic data source for target association and tracking operations in the second-level screening. The detection results of each frame form a historical trajectory sequence of the same target through temporal correlation. Temporal correlation refers to the operation of establishing a connection relationship between the detection results of the same candidate target at different times according to the chronological order. Specifically, based on the positional overlap and feature similarity determination of the target correlation and tracking operation, the detection results in adjacent frames that meet the matching conditions are connected, thereby forming the evolution trajectory of each candidate target over time. The output of temporal correlation is a list of detection results corresponding to each candidate target, ordered by time. The list constitutes the basic data unit for subsequent dynamic feature extraction. Short temporal image sequence refers to the sequence of image frames within a continuous time window obtained after temporal correlation for a single candidate target. The sequence is composed of local image blocks corresponding to the candidate region in the continuous image frames arranged in chronological order. The time window length of the short temporal image sequence is preset according to the flickering frequency of the forest flame and the video frame rate. The window length must be sufficient to capture the complete dynamic change cycle of the flame, but should not be too long to avoid additional interference introduced by changes in the forest background.Brightness variation features, motion features, and morphological variation features refer to three types of features used to characterize the dynamic attributes of candidate regions in the time dimension. Brightness variation features reflect the degree of fluctuation of the average pixel brightness within the candidate region in the time series, and are used to describe the flickering attribute of the flame. Motion features reflect the displacement distribution of pixels within the candidate region between adjacent frames, and are used to describe the non-rigid motion attribute of the flame. Morphological variation features reflect the degree of deformation of the outline shape of the candidate region between adjacent frames, and are used to describe the irregular expansion and contraction attribute of the flame front. The three together constitute a complete flame dynamic feature description system, covering the dynamic change essence of the flame from three dimensions: luminosity, kinematics, and geometry. The calculation of the temporal fluctuation amplitude of the average pixel brightness within each candidate region in the sequence involves quantizing the average brightness value of all pixels within each candidate region in a short temporal image sequence to form a sequence of brightness average values ​​changing over time. The fluctuation amplitude of the sequence is then calculated to quantify the brightness change characteristics. The fluctuation amplitude is calculated by any one of the following methods: the sum of the absolute values ​​of the differences in the average brightness values ​​between adjacent frames in the sequence, the standard deviation of the sequence, or the difference between the peak and valley values ​​of the sequence. The quantization result is compared with the preset brightness change significance condition. The brightness fluctuation amplitude of the real flame is significantly higher than that of the static interference target. Extracting the pixel-level optical flow field distribution of candidate regions between adjacent frames in a computational sequence refers to calculating the motion vector of each pixel in the candidate region of adjacent frames in a short temporal image sequence from the previous frame to the current frame, forming a dense motion vector field covering the entire candidate region, and then performing statistical analysis on the vector field to extract motion features. The optical flow field distribution reflects the motion direction and speed of each pixel within the candidate region. In a real flame, the instability of combustion causes the motion direction and size of each pixel to exhibit a chaotic and disordered distribution pattern, i.e., non-rigid motion features, while the interior of a static interference target exhibits a uniform or near-zero optical flow field distribution. Quantification is achieved by calculating the degree of deformation of candidate region contours between adjacent frames in a sequence. This involves extracting the edge contour lines of candidate regions in each frame of a short temporal image sequence and quantifying morphological change features by comparing the shape differences of contour lines between adjacent frames. The degree of contour deformation is quantified by calculating any one of the following methods: Hausdorff distance between adjacent frames, rate of change of contour area, or rate of change of contour perimeter. The contour of a real flame undergoes continuous and irregular deformation due to the instability of the combustion process, resulting in significant morphological changes, while the contour of a static interference target remains basically unchanged for a short period of time. The dynamic feature descriptor of the candidate region refers to a multi-dimensional feature vector representation formed by combining the quantized values ​​of brightness change features, motion features, and morphological change features according to a preset dimension. As a unified quantitative expression of the dynamic attributes of the candidate region, the dynamic feature descriptor provides structured feature input for the spatiotemporal consistency comprehensive scoring, and its number of dimensions matches the number of quantization indicators of the three selected feature types.The dynamic consistency score refers to the quantitative score calculated on the degree to which each candidate region represents the dynamic attributes of a real flame in three dimensions: brightness change characteristics, motion characteristics, and morphological change characteristics. The score for each dimension is obtained by normalizing the quantitative value of the corresponding feature. The higher the score, the closer the candidate region is to the dynamic characteristics of a real flame in each dimension. The dynamic consistency score is used for subsequent weighted summation to calculate the comprehensive score, realizing the fusion evaluation of multi-dimensional dynamic features. The comprehensive confidence level refers to the credibility of the candidate region, represented by the comprehensive score, in possessing the dynamic change characteristics of a real flame. Its value range is normalized to a preset interval. The higher the value, the higher the reliability of the candidate region being judged as a real fire point after spatiotemporal consistency analysis. The comprehensive confidence level provides a priority ranking basis for input candidate regions in the third-level screening. Candidate regions with higher comprehensive confidence levels can obtain higher authenticity judgment weights in the third-level screening. The comprehensive score is a weighted total score obtained by multiplying the scores for brightness change features, motion features, and morphological change features by their respective preset weight coefficients and then summing them. The formula is: Comprehensive score = Brightness score multiplied by the first weight, Motion score multiplied by the second weight, and Morphological score multiplied by the third weight. Each weight coefficient is pre-calibrated based on the importance of flame dynamic change features in different forest area scenarios. The comprehensive score is the sole quantitative basis for determining whether a candidate area passes the spatiotemporal consistency verification. Its score comprehensively reflects the overall degree of matching between the candidate area and the actual flame dynamic attributes in the time dimension. The dynamic judgment threshold is a preset critical score value used to determine whether a candidate area possesses flame dynamic change features. When the comprehensive score of a candidate area reaches or exceeds this threshold, the area is determined to have passed the spatiotemporal consistency verification; when the comprehensive score is below the threshold, the area is determined to not possess flame dynamic change features. The setting of the dynamic judgment threshold needs to balance the different tolerance requirements for false negative and false positive rates in forest fire point monitoring and is pre-configured according to specific deployment scenarios and business needs. Spatiotemporal consistency verification refers to the process of comprehensively judging whether a candidate region possesses the characteristics of dynamic flame change based on the consistency of its dynamic attributes in three dimensions: brightness change, motion characteristics, and morphological changes in a continuous time series. The verification method is to compare the comprehensive score of the candidate region with the dynamic judgment threshold. Candidate regions that pass the verification are considered to possess the characteristics of dynamic flame change and are retained as real fire point candidates to enter the third level of screening. Candidate regions that fail the verification are judged to lack the characteristics of dynamic flame change and are eliminated as static interference targets.

[0043] In the above embodiments, the multi-level screening process of this embodiment significantly improves the accuracy of small target fire point detection in forest areas through spatiotemporal feature fusion and dynamic consistency verification mechanisms. The target association and tracking module effectively solves the problem of instantaneous interference caused by leaf swaying or small animal activity in complex forest scenes by calculating inter-frame position overlap and feature similarity; it can accurately distinguish the continuous motion trajectory of real fire points from randomly occurring noise signals, providing a reliable data foundation for time series analysis. The dynamic feature descriptor construction module adopts a three-dimensional feature extraction strategy of brightness fluctuation, optical flow field distribution, and contour deformation to comprehensively capture the unique dynamic attributes of flames; the temporal fluctuation quantization of the average brightness can effectively filter steady-state high-brightness areas such as sunspots; pixel-level optical flow field analysis can identify the unique disordered diffusion motion mode of flames; contour deformation detection eliminates interference from artificial light sources with fixed shapes; the complementary design of the three features significantly enhances the algorithm's ability to characterize the essential dynamic characteristics of flames. The spatiotemporal consistency scoring mechanism achieves quantitative evaluation of flame characteristics by weighted fusion of scores from three types of dynamic features. It considers both the non-steady-state characteristics of flame brightness changing with airflow and the morphological characteristics of continuously fluctuating flame boundaries; it also incorporates the natural diffusion characteristics of flame location, forming a multi-dimensional dynamic verification standard. This comprehensive evaluation method has higher anti-interference capabilities than single-feature discrimination. The setting and application of dynamic judgment thresholds constitute a key link in adaptive screening. Through the adjustability of preset thresholds, the system can automatically balance the conflicting requirements of detection sensitivity and false alarm rate according to different forest environment characteristics, such as vegetation density and light conditions; ensuring a stable detection rate of real fire points while controlling the false judgment probability of common interferences such as leaf reflection and metallic flashes within acceptable limits.

[0044] In summary, the output of this embodiment is a set of candidate fire points with high confidence. Its technical value lies in realizing the dual confirmation of small target fire points in the spatial dimension or single frame detection and the temporal dimension or multi-frame verification. The spatiotemporal joint analysis method is particularly suitable for solving the technical problem of small target fire points and complex backgrounds in early forest fire monitoring, and provides reliable data support for accurate alarm and fire situation assessment.

[0045] Example 6: As Figure 3 As shown, based on Example 1, the third-level screening process in step S200 provided in this embodiment of the invention specifically includes the following steps: Step S207: The candidate fire points that have passed the second-level screening are cropped and normalized in size, and the local image patches corresponding to each candidate region are extracted. The normalized local image patches are input into the pre-trained lightweight flame semantic recognition model. Forward inference calculation is performed through convolutional neural network or lightweight visual Transformer structure to extract the high-level visual semantic features of the flame under the conditions of illumination change, scale change and local occlusion, and output the semantic probability of each candidate region belonging to the real flame category, so as to obtain a set of candidate regions with semantic probability annotation. Step S208: For each candidate region in the candidate region set with semantic probability annotation, simultaneously acquire its corresponding external environmental information, including infrared thermal imaging data, meteorological environmental data, and forest geospatial information. Infrared thermal imaging data is used to extract the infrared radiation intensity and local thermal anomaly features of the candidate region and calculate the thermal anomaly score. Meteorological environmental data is used to acquire the ambient temperature, air humidity, and wind speed parameters of the candidate region and calculate the environmental condition score. Forest geospatial information is used to acquire the vegetation type, terrain slope, and historical fire risk level of the candidate region and calculate the regional risk score. The three scores are combined to form the environmental feature vector of each candidate region, resulting in a candidate region set with environmental feature vectors. Step S209: The semantic probability, thermal anomaly score, environmental condition score, and regional risk score of each candidate region in the candidate region set with semantic probability annotation and environmental feature vector are weighted and fused according to preset weight coefficients. The fusion calculation method is to multiply the semantic probability by the first weight, the thermal anomaly score by the second weight, the environmental condition score by the third weight, and the regional risk score by the fourth weight, and then sum them to obtain the comprehensive authenticity score of each candidate region, and generate a candidate region set with comprehensive authenticity score. Step S210: Compare the comprehensive authenticity score of each candidate region in the candidate region set with the comprehensive authenticity score with the preset fire point determination threshold. When the comprehensive authenticity score reaches or exceeds the threshold, the candidate region is confirmed as a real fire point; otherwise, it is determined as a non-fire point target and is removed to form a confirmed set of real fire points.

[0046] Image cropping and size normalization refers to the process of extracting corresponding local image patches from the original image based on the bounding box coordinates or pixel mask positions of the candidate fire point region in the original image, and then uniformly scaling the cropped image patches to a preset standard size space. This process is used to convert candidate regions of different scales and aspect ratios into a uniform size specification to meet the requirement of the lightweight flame semantic recognition model for consistency of input tensor dimensions, and to eliminate the influence of spatial scale differences on the semantic discrimination results of candidate regions. A local image patch refers to a local sub-image cropped from the original image according to the bounding box coordinates of the candidate fire point region, containing only the candidate target and a limited surrounding background range. The local image patch is the direct input object of the lightweight flame semantic recognition model. Compared with inputting the entire image into the model, the local image patch significantly reduces the computational load required by the model, while removing large-scale forest background information unrelated to the candidate target, allowing the model to concentrate computational resources on feature discrimination of the candidate region. The lightweight flame semantic recognition model refers to a deep learning model based on a lightweight neural network structure, used for binary classification of input image patches into flame and non-flame or fire and non-flame scenes. The model includes at least an input layer, a feature extraction backbone network, and a classification output layer: the input layer receives normalized local image patches as input tensors; the feature extraction backbone network consists of several convolutional layers, activation layers, and pooling layers or attention mechanism modules stacked together, used to extract visual feature responses from the input image patches from the bottom layer to the top layer; the classification output layer consists of global pooling layers and fully connected layers, used to map high-level features to a preset category space and output the probability distribution corresponding to each category; the number of parameters and floating-point operations of the feature extraction backbone network are controlled below a preset threshold so that the model can be deployed on forest edge computing nodes to meet the computing power constraints of real-time inference; the output of the model is the semantic probability of the candidate region belonging to the flame category, which serves as the core input of the visual semantic dimension in the third-level screening. Forward inference computation refers to the process of using normalized local image patches as input data, starting from the input layer of the lightweight flame semantic recognition model, and then passing through each convolutional layer, activation layer, pooling layer, and fully connected layer in a parametric calculation, passing the activation response layer by layer, and finally obtaining the classification probability value at the output layer. Forward inference computation does not involve updating or adjusting model parameters. It is a fixed-parameter inference process in the deployment phase, and its output result is directly used as semantic probability for subsequent fusion computation.High-level visual semantic features refer to the highly abstract and class-discriminative feature map responses extracted by the deep network near the output layer in the lightweight flame semantic recognition model. Compared to the low-level visual features such as edges, colors, and textures extracted by shallow networks, high-level visual semantic features can characterize the relatively stable global morphological patterns and contextual information of flame targets in the complex background of forest areas, such as the overall morphological structure of the flame area and the semantic contrast relationship between the fire point and the background. High-level visual semantic features serve as the internal basis for the model's classification output, and their quality directly affects the reliability of semantic probabilities. Semantic probability refers to the predicted probability value of a candidate region belonging to the real flame category, output by the classification output layer after forward inference calculation by the lightweight flame semantic recognition model. The value of semantic probability is normalized to the range of zero to one. The higher the value, the greater the model's confidence in identifying the candidate region as a real flame from a visual semantic perspective. As an independent input in the weighted fusion calculation, semantic probability participates in the calculation of the comprehensive authenticity score along with the environmental information scores. Infrared thermal imaging data, meteorological environmental data, and forest geospatial information refer to three types of external environmental information obtained from multi-source image data. Infrared thermal imaging data reflects the temperature distribution information of candidate areas. By extracting the infrared radiation intensity value corresponding to the candidate area and calculating its thermal anomaly amplitude relative to the surrounding forest background, the degree of thermodynamic anomaly of the candidate area is quantitatively characterized. Meteorological environmental data includes the ambient temperature, air humidity, and wind speed parameters of the candidate area. These parameters are used to assess the degree of favorability of current meteorological conditions for fire occurrence and spread; the more favorable the meteorological conditions are for fire development, the higher the environmental condition score. Forest geospatial information includes the vegetation type classification, terrain slope value, and historical fire risk level records of the candidate area. This information is used to assess the fire risk level of the candidate area in the spatial dimension; the higher the risk level, the higher the regional risk score. All three types of information are external evidence sources independent of visual image data and, together with semantic probability, constitute the input basis for multi-source fusion discrimination. The comprehensive authenticity score refers to the comprehensive quantitative score obtained by weighted fusion of the semantic probability, thermal anomaly score, environmental condition score, and regional risk score of the candidate area. The fusion calculation method is to multiply the semantic probability by the first preset weight, the thermal anomaly score by the second preset weight, the environmental condition score by the third preset weight, and the regional risk score by the fourth preset weight, and then sum them up. The sum of each preset weight is one, and the specific value of each weight is pre-calibrated according to the confidence level of each dimension of evidence in different scenarios in the forest area. The comprehensive authenticity score is the result of the collaborative integration of visual semantic discrimination and multi-source environmental constraint evidence. Its score comprehensively reflects the overall credibility of the candidate area being identified as a real fire point with the support of multi-dimensional evidence.The fire point determination threshold refers to a preset comprehensive authenticity score threshold used to determine whether a candidate area should be confirmed as a real fire point. When the comprehensive authenticity score of a candidate area reaches or exceeds this threshold, the conclusion that the area is a real fire point is determined by the comprehensive evidence from multiple sources, and the conclusion is confirmed and output. When the comprehensive authenticity score is lower than this threshold, it indicates that the multi-source evidence does not sufficiently support the fire point determination conclusion, and the candidate area is eliminated as a false detection target. The specific value of the fire point determination threshold is preset and configured according to the system's tolerance requirements for false alarm rate and false negative rate, and can be dynamically adjusted in different forest monitoring scenarios. The geographic coordinate system refers to a reference coordinate system used to characterize the spatial position of the Earth's surface. It defines the precise location of ground points in the form of latitude and longitude coordinates. By jointly spatially calculating the pixel coordinates of the candidate fire point in the original image with the geographical location, imaging parameters, and terrain elevation data of the image acquisition device, a mapping relationship between image pixel space and geographic space is established. The confirmed real fire point image position is converted into latitude and longitude coordinates in the geographic coordinate system to support the precise location and spatial visualization of the fire point.

[0047] In the above embodiments, the third-level screening process of this embodiment achieves the final confirmation and standardized output of small target fire points in forest areas through multimodal feature fusion and structured decision-making mechanisms. The forward inference calculation of the semantic recognition model adopts a convolutional network with strong illumination robustness or a lightweight visual Transformer structure, which effectively solves the problem of loss of small target flame features caused by changes in illumination conditions or partial occlusion in forest scenes; the extraction and classification probability output of high-level visual semantic features ensure that the model can identify the texture, color and morphological characteristics of flames in complex backgrounds, and enhance the algorithm's ability to distinguish simulated flame interference. The introduction of external environmental information constitutes a multi-source data verification system. The thermal anomaly score provided by infrared thermal imaging data quantifies the unique radiation characteristics of flames, making up for the detection blind spot of the visible light band under smoke obscuration; the environmental condition score constructed by meteorological environmental data establishes a correlation model between the possibility of flame spread and wind speed and humidity; the risk score generated by geospatial information reflects the fire spread trend under different vegetation types and terrain conditions; the combined use of multi-dimensional environmental feature vectors enables the system to distinguish heat source targets with similar visual features but not flames. The weighted fusion mechanism integrates semantic probability and three types of environmental scores through preset coefficients to form a comprehensive authenticity assessment system. It considers both the visual semantic features of the flame itself and the influence of environmental factors on the probability of fire point existence, making the final score physically interpretable. The configurable nature of different weights allows the system to adjust the contribution ratio of each factor according to specific monitoring needs, achieving flexible optimization of the detection strategy. The application of fire point determination thresholds completes the algorithm-interpretable binary decision-making process, transforming continuous scores into discrete determination results, ensuring that the output results meet the binary input requirements of the business system. The dynamic adjustment capability of the thresholds enables the system to adapt to changes in fire risk levels in different seasons or regions, maintaining detection stability. The structured output of standardized results achieves seamless integration of detection information with the emergency response system. The geographic coordinate mapping of the fire point location ensures that the spatial positioning accuracy meets the needs of rescue command. The recording of timestamps and coverage areas provides a data foundation for fire evolution analysis, and the labeling of confidence levels assists in prioritization during manual review, ensuring the integrity and availability of the information chain from algorithm output to business applications.

[0048] In summary, this embodiment establishes a forest fire confirmation system based on multimodal data collaborative analysis, effectively addressing the limitations of single-sensor or single-spatiotemporal-dimensional detection methods in small-target, low-contrast scenarios. Through cross-validation of visual semantics, infrared radiation, environmental parameters, and geographic information, the system can minimize the risk of false alarms and missed alarms caused by complex environmental interference while ensuring real-time performance, thus meeting the core requirement of accurate early-stage forest fire monitoring.

[0049] Example 7: Based on Example 1, the process of structurally outputting the actual fire point detection results in step S300 of this embodiment of the invention specifically includes the following steps: Step S301: Based on the confirmed set of real fire points, extract the spatial location information, discovery time information and confidence level information of each real fire point, and combine them with the positioning information and attitude information of the image acquisition device to map the pixel coordinates of the real fire points in the image coordinate system to the latitude and longitude coordinates in the geographic coordinate system to obtain the structured information of the fire points with geographic calibration. Step S302: Input the structured information of fire points with geographic labels into the intelligent forest fire prevention assessment platform. The platform performs visual labeling of fire points and combines forest geospatial data and meteorological data to assess the risk level of fire point locations, generating fire point assessment results that include fire point location labels, risk level assessment results, and alarm level information. Step S303: Based on the current status of the forest area communication links, when the communication conditions are detected to meet the preset transmission requirements, the structured information of the fire point and the fire point assessment results are output through the conventional network link; when the communication conditions are detected to be restricted, the key data fields in the structured information of the fire point and the fire point assessment results are compressed and encoded, and uploaded to the command center through the satellite internet communication link or the Beidou short message method to realize the rapid transmission of fire information in complex forest area environments.

[0050] Among them, spatial location information, discovery time information, and confidence level information refer to the three types of core attribute data that need to be output after the real fire point is confirmed through three levels of screening. Spatial location information is the geographical coordinate data of the fire point within the target area, including longitude, latitude, and elevation values. Discovery time information is the acquisition timestamp of the image frame where the fire point is located, accurate to the second or millisecond level. Confidence level information is a qualitative level based on the interval of the comprehensive authenticity score, including at least three levels: high, medium, and low, used to characterize the reliability of the fire point detection result. The three types of information together constitute the basic information elements of the fire point detection result, providing data support for the display and disposal decision of the forest fire prevention platform. Image acquisition equipment refers to the hardware imaging terminal that acquires multi-source image data, including but not limited to ground monitoring cameras, visible light and infrared imaging payloads carried by UAVs, satellite remote sensing payloads, etc. Image acquisition equipment is the reference benchmark for establishing the mapping relationship between the fire point image coordinates and the actual geographical location. The equipment type and parameters directly affect the spatial mapping accuracy from pixel coordinates to geographic coordinates. Location information refers to the spatial geographical location of the image acquisition device at the time of image acquisition, expressed as longitude, latitude, and elevation values ​​in a geographic coordinate system. Attitude information refers to the spatial orientation angle parameters of the image acquisition device at the time of image acquisition, including azimuth, pitch, and roll angles. For satellite and UAV platforms, it also includes orbital parameters or heading and attitude data. Location and attitude information are key input parameters for constructing a spatial mapping model between image pixel coordinates and geographic coordinates, and their accuracy directly determines the accuracy of fire point geocalibration. Pixel coordinates in the image coordinate system refer to the spatial location representation of the fire point candidate region in the original image, usually expressed as the horizontal and vertical coordinate values ​​of the center pixel of the bounding box or the extreme coordinates of each boundary pixel of the fire point target region. Pixel coordinates in the image coordinate system are the original reference values ​​for mapping the spatial location information of the fire point from image space to geographic space, and can only be converted into usable geographic coordinates after spatial mapping calculations. Latitude and longitude coordinates in the geographic coordinate system refer to the spatial location values ​​of fire points on the Earth's surface represented by a latitude and longitude grid. They are expressed in degrees and include degrees, minutes, and seconds, or in decimal form. Elevation values ​​can be added to characterize the three-dimensional spatial location. Latitude and longitude coordinates are the standard location expression output after converting fire point detection results from image processing space to actual geographic space, providing forest fire prevention platforms with location data that can be directly used for map annotation and spatial analysis. Structured fire point information refers to a standardized information data body formed by organizing the actual fire point detection results after three levels of screening and confirmation according to a preset data format. It includes at least the fire point latitude and longitude coordinates field, the discovery timestamp field, the confidence score field, and the confidence level field. The data structure is in key-value pair or table row format. Structured fire point information is a standardized data input for the forest fire prevention platform to parse, display, and store; it is also the basic data unit used for compression encoding or priority transmission during communication.The intelligent forest fire prevention assessment platform refers to a data processing and decision support system platform deployed in forest management centers or the cloud. It receives, parses, displays, and analyzes structured information about fire points, and features a map visualization engine, risk assessment calculation module, and alarm push service. As the final aggregation and display point for fire point detection results, the platform transforms the multi-level filtered and confirmed fire point detection results into decision-making information that fire prevention managers can intuitively understand and respond to quickly. Visual annotation processing refers to mapping the latitude and longitude coordinates of the received structured fire point information to the corresponding locations on the electronic map on the map interface of the intelligent forest fire prevention assessment platform, and displaying them intuitively in the form of preset graphic markers. Visual annotation processing also includes overlaying the fire point location with the forest geographic layer, making the spatial relationship between the fire point and surrounding vegetation types, road networks, and water system distribution clear at a glance, assisting managers in quickly grasping the fire point location and situation. Risk level assessment refers to the process by which the intelligent forest fire prevention analysis platform, based on fire location information and combined with multi-source data such as vegetation type distribution, terrain slope, historical fire frequency, and current meteorological conditions in candidate fire point areas, calculates and quantifies the potential fire spread risk caused by a fire point using a preset set of assessment rules or a risk scoring model. Risk level assessment is a derivative information generated after fire point confirmation, and its results provide a basis for setting alarm levels and formulating emergency dispatch strategies. Forest area communication link status refers to the connectivity and bandwidth availability of the data transmission channel between the current forest area monitoring area and the command center, including various types such as conventional terrestrial communication network links, satellite internet links, and BeiDou short message communication links. The forest area communication link status is the basis for selecting different output methods in step S303, directly determining the timeliness and data volume limitations of fire point information transmission. Preset transmission requirements refer to pre-defined evaluation criteria used to determine whether the current communication link in the forest area meets the conditions for transmitting structured fire information. These criteria include at least the following: good link connectivity, available bandwidth not lower than the preset lower bandwidth threshold, and bit error rate not higher than the preset upper threshold. The preset transmission requirements serve as the technical basis for the output mode branch selection in step S303. When these requirements are met, a complete output via a conventional network link is used; otherwise, the restricted transmission mode is switched to. Compression coding refers to the operation of compressing the key data fields in the structured fire information and fire analysis results according to a preset coding protocol in forest area scenarios with limited communication conditions. This includes precision truncation coding of coordinate data, differential compression of timestamps, selective field extraction of structured fields, and shrinking and reassembling of data packets. The purpose of compression coding is to transmit the core fields of fire information with the minimum amount of data under low bandwidth or short message communication conditions, ensuring that fire intelligence can still be transmitted to the command center even under limited communication conditions.The command center refers to the superior management center or emergency dispatch agency that receives forest fire monitoring information. It has functions such as personnel on duty, comprehensive information display, and emergency dispatch instructions issuance. The command center is the final receiving end of structured fire information and is responsible for assessing the fire situation based on the received fire information and initiating corresponding emergency response procedures.

[0051] In the above embodiments, the structured output process of this embodiment achieves efficient flow and accurate application of forest fire detection results through multi-source information fusion and adaptive transmission strategies. The geographic coordinate mapping module combines image coordinates with device positioning attitude parameters to achieve precise conversion of pixel-level fire location to latitude and longitude coordinates; utilizing the spatiotemporal synchronization characteristics of device GNSS location data and IMU attitude information, it solves the spatial positioning error problem of mobile observation platforms such as drones or lookout towers under complex terrain conditions, ensuring that the accuracy of fire point geographic coordinate output meets the positioning requirements of forest grid management. The risk assessment function of the fire prevention assessment platform integrates fire point confidence and forest geospatial data to construct a multi-dimensional fire risk analysis model; it performs spatial overlay analysis of fire point locations with vegetation type distribution, historical fire point density, and real-time meteorological parameters to quantify the fire spread risk coefficient in different areas; the alarm level generation mechanism based on dynamic weight calculation can automatically adjust the warning threshold according to the environmental characteristics around the fire point, achieving accurate push of graded alarms. The adaptive communication transmission mechanism solves the transmission bottleneck problem of insufficient network coverage in remote forest areas through link status detection and data compression encoding. The conventional network transmission mode ensures the complete upload of all data under high-bandwidth environments, while the compression coding strategy targets low-bandwidth channels such as satellite or BeiDou short messages. It employs key field extraction and differential coding techniques to reduce data transmission volume while maintaining the integrity of core fire information. The dual-mode transmission design ensures the reachability of fire information under different network conditions. The optimized selection strategy for key fields focuses on core metadata such as fire latitude and longitude, discovery time, and confidence level, improving transmission efficiency by eliminating redundant information. Based on information entropy theory, data fields are prioritized to ensure that parameters most relevant to emergency response are retained during low-bandwidth transmission, maintaining the integrity of the minimum dataset required for command and decision-making.

[0052] In summary, this embodiment establishes a seamless system from fire detection to command and control applications. Through a closed-loop technology system of spatial calibration, risk assessment, and reliable transmission, it solves the pain points of information lag and inaccurate positioning in traditional forest fire monitoring. Precise geographic coordinate output provides a spatial benchmark for dispatching firefighting forces, dynamic risk assessment models assist in forming differentiated response strategies, and adaptive transmission mechanisms ensure that fire information can be delivered in a timely manner in various communication environments, forming a forest fire early warning technology chain covering the entire process of monitoring, analysis, and transmission.

[0053] Example 8: As Figure 6As shown, based on Examples 1-7, the multi-level confirmation system for small target fire points in forest areas provided by this invention includes: Image data acquisition module 1 is used to acquire multi-source image data required for monitoring target areas of forest fires; The filtering and processing module 2 is used to reprocess the preprocessed multi-source image data using a three-level filtering mechanism; The first-level screening is used to quickly extract potential fire point candidate regions from multi-source image data in complex forest environments; the second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate regions to determine whether they have dynamic flame change characteristics; the third-level screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire point candidate regions after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points. The real fire point determination module 3 is used to determine the target area after confirmation by the three-level screening mechanism as a real fire point, and to output the real fire point detection results in a structured manner.

[0054] In the above embodiments, the image data acquisition module provides multi-source heterogeneous remote sensing data support, ensuring that fire point monitoring in the target area has complete spatial resolution and spectral coverage. The technical effect of the screening and processing module is reflected in the hierarchical optimization of the three-level screening mechanism: the first-level screening achieves rapid coarse screening of potential fire point candidate areas through efficient feature extraction algorithms, improving the recall rate of initial detection; the second-level screening uses time series analysis and spatial continuity verification to eliminate transient noise interference, enhancing the robustness of dynamic fire point identification; the third-level screening uses multimodal data fusion technology combined with semantic segmentation and environmental parameter modeling to significantly reduce the false detection rate caused by thermal anomaly reflection or non-fire point heat sources. The fire point accuracy judgment module outputs structured fire point detection results, completing high-confidence fire event localization and meeting the requirements of accuracy, reliability, and automated processing for small target fire point monitoring in forest areas.

[0055] Specifically, this invention relates to the intelligent identification of small target fire points in complex backgrounds by combining satellite remote sensing, UAV inspection and ground perception data in an integrated air-space-ground forest fire monitoring system. It can be applied to scenarios such as forest fire prevention patrol, early fire warning, emergency command and dispatch and intelligent forest fire assessment.

[0056] Figure 3 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0057] The electronic device may include a central processing unit / microprocessor / main control chip 4; and a storage medium 5 coupled to the central processing unit / microprocessor / main control chip 4 and storing computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by the processor.

[0058] The central processing unit / microprocessor / main control chip 4 may include, but is not limited to, one or more processors or microprocessors.

[0059] Storage medium 5 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0060] In addition, the electronic device may include (but is not limited to) a data bus 6, an input / output bus / external bus / device bus 7, a display 8, and input / output devices 9 (e.g., keyboard, mouse, speaker, etc.).

[0061] The central processing unit / microprocessor / main control chip 4 can communicate with external devices (8, 9, etc.) via wired or wireless networks (not shown) through the input / output bus / external bus / device bus 7.

[0062] The storage medium 5 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when the central processing unit / microprocessor / main control chip 4 is running.

[0063] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0064] Figure 4 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0065] like Figure 4As shown, the non-transitory computer-readable storage medium 11 stores instructions, such as computer-readable instructions 10. When the computer-readable instructions 10 are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions 10 stored on the non-transitory computer-readable storage medium 11, the various methods described above can be performed.

[0066] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0067] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0068] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0069] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods of the various embodiments of this invention through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0070] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multi-level confirmation of small target fire points in forest areas, characterized in that, Includes the following steps: To acquire multi-source image data required for monitoring target areas of forest fires; A three-level screening mechanism is used to reprocess the preprocessed multi-source image data; The first-level screening is used to quickly extract potential fire point candidate areas from multi-source image data in complex forest environments; The second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate areas to determine whether they have dynamic flame change characteristics. The third-level screening is used to integrate visual semantic information and external environment information to determine the authenticity of potential fire point candidate areas after spatiotemporal consistency analysis, eliminate false detection targets and output high-confidence fire points. The target area confirmed by the three-level screening mechanism is determined to be the real fire point, and the detection results of the real fire point are output in a structured manner.

2. The multi-level confirmation method for small target fire points in forest areas as described in claim 1, characterized in that, The process of acquiring multi-source image data for target areas in forest fire monitoring includes the following steps: Denoising processing is performed on multi-source image data to suppress high-frequency interference information introduced by smoke, haze and sensor noise in the forest environment; and background noise is reduced while preserving flame edge features to obtain a denoised image; The denoised image is then subjected to illumination-adaptive enhancement processing, which enhances low-brightness areas and suppresses the effects of overexposed areas to obtain an enhanced image. The enhanced image undergoes color space transformation and normalization to improve the expressive power of the flame's color dimension. Pixel distributions from different data sources are also uniformly normalized to meet scale consistency requirements, resulting in a standardized feature tensor.

3. The multi-level confirmation method for small target fire points in forest areas as described in claim 1, characterized in that, The first-level screening process includes the following steps: Candidate fire point regions are extracted from the preprocessed standardized feature tensor. A lightweight target detection model is used to make dense predictions on the image. A relaxed detection strategy is implemented by setting a confidence threshold lower than the conventional detection threshold to obtain an initial set of candidate regions. Spatial constraint filtering is applied to the initial candidate region set to retain regions that meet the basic flame visual attribute discrimination conditions. The discrimination criteria include pixel brightness threshold determination, red-orange channel response intensity determination, and abnormal brightness change degree of local area relative to the background determination. Non-target areas that do not have obvious flame optical characteristics are eliminated to form a set of candidate fire point areas.

4. The multi-level confirmation method for small target fire points in forest areas as described in claim 1, characterized in that, The second-level screening process includes the following steps: For each candidate region in the candidate fire point region set output by the first-level screening, target association and tracking operations are performed between consecutive time series image frames by calculating the positional overlap and feature similarity of candidate regions in adjacent frames. The detection results of the same target at different times are temporally correlated to construct short temporal image sequences corresponding to each candidate region, thus obtaining a set of correlated candidate temporal sequences. For each sequence in each candidate time series set, extract the brightness change features, motion features, and morphological change features of the candidate region in the time dimension; combine the three features to form the dynamic feature descriptor of the candidate region, and obtain the candidate time series set with dynamic feature descriptor; Based on the brightness change features, motion features, and morphological change features of each sequence in the candidate time series sequence set with dynamic feature descriptors, the dynamic consistency score corresponding to each feature is calculated, and the brightness score, motion score, and morphological score are weighted and summed to obtain the spatiotemporal consistency comprehensive score of each candidate region, thus generating a candidate sequence set with a comprehensive score. The comprehensive score of each candidate region in the candidate sequence set with comprehensive score is compared with the preset dynamic judgment threshold. When the comprehensive score of the candidate region reaches or exceeds the threshold, the candidate region is marked as having passed the spatiotemporal consistency verification and retained as a real fire point candidate. Otherwise, the candidate region is judged as a static interference target and removed, forming a set of verified candidate fire point regions and outputting it to the third-level screening.

5. The multi-level confirmation method for small target fire points in forest areas as described in claim 4, characterized in that, Brightness variation characteristics are quantified by calculating the temporal fluctuation amplitude of the average pixel brightness within the candidate region of each frame in the sequence. Motion features are extracted by calculating the pixel-level optical flow field distribution of candidate regions between adjacent frames in the sequence; Morphological change features are quantified by calculating the degree of deformation of the contours of candidate regions between adjacent frames in the sequence.

6. The multi-level confirmation method for small target fire points in forest areas as described in claim 1, characterized in that, The third-level screening process includes the following steps: Image cropping and size normalization are performed on the candidate fire points that have passed the second-level screening to extract the local image patches corresponding to each candidate region. The normalized local image patches are input into a pre-trained lightweight flame semantic recognition model to perform forward inference calculation, extract high-level visual semantic features of flames under illumination changes, scale changes and local occlusion conditions, and output the semantic probability of each candidate region belonging to the real flame category, thus obtaining a set of candidate regions with semantic probability annotations. For each candidate region in the candidate region set with semantic probability annotation, its corresponding external environment information is acquired simultaneously. The three scores of the external environment information are combined to form the environmental feature vector of each candidate region, thus obtaining the candidate region set with environmental feature vector. The semantic probability, thermal anomaly score, environmental condition score, and regional risk score of each candidate region in the candidate region set with semantic probability annotation and environmental feature vector are weighted and fused according to preset weight coefficients to obtain the comprehensive authenticity score of each candidate region, and a candidate region set with comprehensive authenticity score is generated. The comprehensive authenticity score of each candidate region in the candidate region set with a comprehensive authenticity score is compared with a preset fire point determination threshold. When the comprehensive authenticity score reaches or exceeds the threshold, the candidate region is confirmed as a real fire point; otherwise, it is determined as a non-fire point target and is removed, thus forming a confirmed set of real fire points.

7. The multi-level confirmation method for small target fire points in forest areas as described in claim 6, characterized in that, External environmental information includes infrared thermal imaging data, meteorological environmental data, and forest area geospatial information; Infrared thermal imaging data is used to extract the infrared radiation intensity and local thermal anomaly features of the candidate region and to calculate the thermal anomaly score. Meteorological and environmental data are used to obtain the ambient temperature, air humidity, and wind speed parameters of the candidate area and to calculate the environmental condition score. Forest geospatial information is used to obtain the vegetation type, terrain slope and historical fire risk level of the candidate area and to calculate the regional risk score.

8. The multi-level confirmation method for small target fire points in forest areas as described in claim 6, characterized in that, The fusion calculation method involves multiplying the semantic probability by the first weight, the thermal anomaly score by the second weight, the environmental condition score by the third weight, and the regional risk score by the fourth weight, and then summing them up.

9. The multi-level confirmation method for small target fire points in forest areas as described in claim 1, characterized in that, The process of outputting structured results from actual fire point detection includes the following steps: Based on the confirmed set of real fire points, the spatial location information, discovery time information, and confidence level information of each real fire point are extracted. Combined with the positioning and attitude information of the image acquisition device, the pixel coordinates of the real fire points in the image coordinate system are mapped to the latitude and longitude coordinates in the geographic coordinate system to obtain the structured information of fire points with geographic calibration. The structured information of fire points with geographic labels is input into the intelligent forest fire prevention assessment platform. The platform performs visual labeling of fire points and combines forest geospatial data and meteorological data to assess the risk level of fire point locations, generating fire point assessment results that include fire point location labels, risk level assessment results, and alarm level information. The structured information on fire points and the results of fire point analysis are uploaded to the command center.

10. A multi-level confirmation system for small target fire points in forest areas, used to implement the multi-level confirmation method for small target fire points in forest areas as described in any one of claims 1 to 9, characterized in that, include: The image data acquisition module is used to acquire multi-source image data required for monitoring target areas of forest fires; The filtering and processing module is used to reprocess the preprocessed multi-source image data using a three-level filtering mechanism; The first-level screening is used to quickly extract potential fire point candidate regions from multi-source image data in complex forest environments; the second-level screening is used to perform spatiotemporal consistency analysis on potential fire point candidate regions to determine whether they have dynamic flame change characteristics; the third-level screening is used to fuse visual semantic information and external environmental information to determine the authenticity of potential fire point candidate regions after spatiotemporal consistency analysis, eliminate false detection targets, and output high-confidence fire points. The real fire point determination module is used to determine whether the target area after being confirmed by the three-level screening mechanism is a real fire point, and to output the real fire point detection results in a structured manner.

Images