Fire point detection method, device and equipment of remote sensing image and storage medium

Abstract

The application discloses a fire point detection method, device and equipment of remote sensing image and a storage medium, relates to the technical field of remote sensing and fire monitoring, and completes the specific optimization of the data representation level through feature standardization and quantization, carries out fire point identification and determination with the aid of the classifier trained in advance, can fully excavate the feature difference between the fire point and the background area, effectively avoids the identification deviation caused by the environmental light, atmospheric interference and ground color band, and improves the accuracy of the fire point target detection in the remote sensing image. The method comprises the following steps: determining an expert feature set in a candidate area of an input remote sensing image, wherein the expert feature set is obtained by standardizing the multi-dimensional features representing the difference between the fire point and the background in the image element; performing quantization processing on the expert feature set to obtain an expert quantized feature set; inputting the expert quantized feature set into a pre-trained classifier for fire point detection to obtain the detection result of each image element in the candidate area as a fire point.

Description

Technical Field

[0001] This application relates to the fields of remote sensing technology and fire monitoring technology, and in particular to a method, apparatus, equipment and storage medium for detecting fire points in remote sensing images. Background Technology

[0002] Remote sensing fire detection technology has significant application value in fields such as forest fire prevention, environmental monitoring, and disaster early warning. With the rapid development of remote sensing technology, automatic fire detection using remote sensing images acquired from platforms such as satellites and drones has become an important means of fire monitoring.

[0003] Existing remote sensing fire detection technologies mainly include thermal infrared thresholding, multispectral rule-based methods, machine learning methods, and deep learning methods. Thermal infrared thresholding utilizes brightness temperature information in the mid-infrared or thermal infrared bands, setting fixed or dynamic thresholds to determine whether a pixel is a fire point. Multispectral rule-based methods further utilize the differential responses of fire points in near-infrared, short-wave infrared, and thermal infrared bands, using multi-band ratios, band differences, normalization exponents, and background contrast rules to identify fire points. Machine learning methods extract spectral, texture, thermal anomaly, and contextual features from remote sensing images, using classifiers such as random forests, support vector machines, and gradient boosting trees to complete fire point identification. Deep learning methods employ convolutional neural networks, attention mechanisms, transformers, and other models, automatically learning complex spatial patterns through end-to-end training, achieving high accuracy in remote sensing fire point detection and fire segmentation tasks.

[0004] Although the aforementioned fire detection methods have achieved some success, the following problems still exist: First, the dense computation across the entire image leads to low processing efficiency for large-format remote sensing images. In high-resolution remote sensing images, the vast majority of pixels belong to the background area, while real fire pixels usually account for only a very small proportion. Existing methods, especially deep learning methods, generally adopt pixel-by-pixel, dense convolution across the entire image, or full-image feature calculation methods, resulting in a large amount of computing power being consumed by irrelevant backgrounds, which is not conducive to real-time detection on spacecraft and rapid processing at the edge. Second, the false alarm rate is high against complex backgrounds. Traditional thresholding methods and some purely data-driven models lack sufficient thermophysical constraints, easily mistaking highly reflective ground features, high-temperature bare soil, industrial heat sources, and clouds. Third, the training process is difficult to adapt to resource-constrained platforms. Deep learning models rely on gradient backpropagation and large amounts of feature map storage, resulting in significantly higher memory and energy consumption during training than during inference, making it difficult to achieve autonomous training and online model updates on spaceborne computers or low-power edge devices. Fourth, existing machine learning methods lack data representation optimization for edge hardware. Existing methods based on tree models or shallow classifiers typically use floating-point continuous features directly for training and inference, resulting in limited training efficiency and energy efficiency on edge devices. In summary, existing fire detection technologies generally suffer from technical problems such as low efficiency in processing large-format remote sensing images due to intensive full-map computation, high false alarm rates in complex backgrounds, difficulty in adapting the training process to resource-constrained platforms, and lack of data representation optimization for edge hardware. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, device and storage medium for fire detection of remote sensing images. The main purpose is to solve the technical problems that existing fire detection technologies generally have, such as low efficiency in processing large-format remote sensing images due to dense full-map computation, high false alarm rate in complex backgrounds, difficulty in adapting the training process to resource-constrained platforms, and lack of data representation optimization for edge hardware.

[0006] The first aspect provides a method for detecting fire points in remote sensing imagery, the method comprising: Within the candidate region of the input remote sensing image, an expert feature set is determined. The expert feature set is obtained by standardizing the multidimensional features in the pixels that represent the difference between the fire point and the background. The expert feature set is quantized to obtain the expert quantized feature set; The expert-quantified feature set is input into a pre-trained classifier for fire detection, and the detection result of each pixel in the candidate region is obtained as a fire point.

[0007] Secondly, a fire detection device for remote sensing imagery is provided, the device comprising: The determining unit is used to determine an expert feature set within a candidate region of the input remote sensing image. The expert feature set is obtained by standardizing the multidimensional features in the pixels that characterize the difference between the fire point and the background. A quantization unit is used to quantize the expert feature set to obtain an expert quantized feature set. The detection unit is used to input the expert quantized feature set into a pre-trained classifier to detect fire points and obtain the detection result of each pixel in the candidate region as a fire point.

[0008] Thirdly, a fire detection device for remote sensing images is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, it implements the aforementioned fire detection method for remote sensing images.

[0009] Fourthly, a storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the aforementioned method for detecting fire points in remote sensing images.

[0010] By employing the above technical solution, this application provides a method, apparatus, device, and storage medium for detecting fire points in remote sensing images. Compared with existing remote sensing fire point detection technologies, this application determines an expert feature set within the candidate region of the input remote sensing image. The expert feature set is obtained by standardizing multi-dimensional features representing the differences between fire points and the background in pixels. The expert feature set is then quantized to obtain a quantized expert feature set. This quantized expert feature set is input into a pre-trained classifier for fire point detection, resulting in the detection result of each pixel within the candidate region as a fire point. The entire process relies on accurately extracting multi-dimensional features that can distinguish between fire points and the background from the candidate region of the remote sensing image and completing the standardization to construct a unified and standardized expert feature set. Quantization then simplifies the feature data to reduce redundant information and computational overhead. Using a pre-trained classifier for fire point identification and judgment, this process can fully exploit the feature differences between fire points and the background region, effectively avoid recognition biases caused by ambient light and atmospheric interference, and surface discoloration, significantly improving the accuracy of fire point target detection in remote sensing images. Meanwhile, targeted optimizations at the data representation level are achieved through feature standardization and quantization, which are tailored to the operating characteristics of edge hardware and reduce the consumption of computing resources. While ensuring the accuracy of fire detection, it also takes into account the computing speed and deployment adaptability, which can meet the practical application needs of efficient and accurate fire detection of remote sensing images in different usage scenarios.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a method for detecting fire points in remote sensing images according to an embodiment of this application; Figure 2 yes Figure 1 A flowchart illustrating a specific implementation method prior to step 101; Figure 3 yes Figure 1 A flowchart illustrating a specific implementation method of step 101; Figure 4 yes Figure 1 A flowchart illustrating a specific implementation method for step 102; Figure 5 yes Figure 1 A flowchart illustrating a specific implementation method following step 102; Figure 6 This is a schematic diagram of the structure of a fire detection device for remote sensing images in one embodiment of this application; Figure 7 This is a schematic diagram of the device structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] Existing remote sensing fire detection technologies mainly include thermal infrared thresholding, multispectral rule-based methods, machine learning methods, and deep learning methods. Thermal infrared thresholding utilizes brightness temperature information in the mid-infrared or thermal infrared bands, setting fixed or dynamic thresholds to determine whether a pixel is a fire point. Multispectral rule-based methods further utilize the differential responses of fire points in near-infrared, short-wave infrared, and thermal infrared bands, using multi-band ratios, band differences, normalization exponents, and background contrast rules to identify fire points. Machine learning methods extract spectral, texture, thermal anomaly, and contextual features from remote sensing images, using classifiers such as random forests, support vector machines, and gradient boosting trees to complete fire point identification. Deep learning methods employ convolutional neural networks, attention mechanisms, transformers, and other models, automatically learning complex spatial patterns through end-to-end training, achieving high accuracy in remote sensing fire point detection and fire segmentation tasks.

[0015] Although the aforementioned fire detection methods have achieved some success, they generally suffer from technical problems such as low efficiency in processing large-format remote sensing images due to dense computation across the entire map, high false alarm rates in complex backgrounds, difficulty in adapting the training process to resource-constrained platforms, and a lack of data representation optimization for edge hardware.

[0016] To address this problem, this embodiment provides a method for detecting fire points in remote sensing images, such as... Figure 1 As shown, it includes the following steps: 101. Within the candidate regions of the input remote sensing image, determine the set of expert features.

[0017] The expert feature set is obtained by standardizing the multidimensional features representing the differences between fire points and the background in pixels. In the input remote sensing image, the candidate region refers to the set of pixels that may contain fire points after preliminary screening. These candidate regions are usually obtained by calculating the thermal response score of each pixel. When the thermal response score of a pixel exceeds a preset threshold, the pixel is included in the candidate region. The determination of the candidate region effectively narrows the scope of subsequent feature extraction, avoids full processing of the entire remote sensing image, and significantly improves the efficiency of the algorithm.

[0018] In this embodiment, the expert feature set is a feature vector composed of features of multiple dimensions, each characterizing the difference between the fire point and the background from different perspectives. In one specific implementation, the expert feature set includes three main dimensions: thermal features, spectral features, and spatial anomaly features. Thermal features reflect the degree of temperature anomaly in the pixels, spectral features characterize the spectral response characteristics of the fire point in different bands, and spatial anomaly features describe the abrupt changes in the spatial distribution of the fire point. These three dimensions complement each other, together forming a comprehensive description of the fire point.

[0019] The specific standardization process transforms the feature values ​​of each dimension into standard feature values ​​with uniform dimensions and distribution range. One possible implementation is a zero-mean, unit-variance method. For each feature dimension, the mean and standard deviation of that dimension in the training samples are first calculated. Then, the mean is subtracted from the feature value of each pixel in that dimension, and the result is divided by the standard deviation. Standardization eliminates dimensional and numerical range differences between different feature dimensions, ensuring that each dimension has the same weight in subsequent quantization and classification processes. For example, the original values ​​of thermal features might be in the range of 0 to 100, while the original values ​​of spectral features might be in the range of 0 to 1. After standardization, the feature values ​​of both dimensions are transformed into a distribution with a mean of zero and a standard deviation of 1.

[0020] Through the above steps, the final expert feature set is a multi-dimensional vector, with each dimension standardized to have the same statistical properties. This standardized feature vector retains the differences between fire points and the background across multiple dimensions while eliminating the influence of units and numerical ranges, providing a data foundation for subsequent quantization processing and fire point detection by the classifier.

[0021] 102. Quantize the expert feature set to obtain the expert quantified feature set.

[0022] Understandably, the various mathematical operations involved in feature extraction, such as ratio calculations, normalized differences, and morphological operations, produce continuous floating-point values. Standardization also converts the original features into a floating-point distribution with a mean of zero and a standard deviation of one. Typically, floating-point values ​​require 32 bits of storage space, but in practical applications of fire detection in remote sensing imagery, computational and storage resources are usually limited. By quantizing the expert feature set, feature values ​​that originally required 32 floating-point representations can be compressed into 8-bit or 16-bit integer representations, significantly reducing memory usage and data transmission bandwidth. Furthermore, integer operations are faster and consume less power than floating-point operations, making them particularly suitable for deploying fire detection algorithms on resource-constrained edge devices such as satellites and drones.

[0023] In this embodiment, the process of quantizing the expert feature set essentially involves mapping continuous floating-point feature values ​​to discrete integer levels. Specifically, based on the actual value range and discrimination requirements of various features in the fire detection task, the floating-point features of each dimension in the expert feature set are divided into several ordered numerical intervals, either equally or unequally, according to a preset number of levels. Each interval corresponds to an integer level identifier; the original floating-point feature value is assigned the corresponding integer level value based on which interval it falls into, thus obtaining the quantized expert feature set.

[0024] In actual fire monitoring applications, the number of corresponding levels can be set according to the fluctuation range of fire-related characteristics, and the continuous characteristic values ​​reflecting temperature differences and brightness differences can be divided into multiple value ranges from low to high.

[0025] Through the quantization process described above, the originally continuously changing floating-point features are transformed into a finite number of discrete integer levels. This preserves the relative magnitude relationships and discriminative information of the features while simplifying and compressing the data representation. This quantization conversion not only significantly reduces the storage space requirements of the features but also eliminates redundant subtle numerical fluctuations within the standard features, retaining feature information that can distinguish between fire points and background.

[0026] 103. Input the expert quantized feature set into a pre-trained classifier to perform fire detection, and obtain the detection result of each pixel in the candidate region as a fire point.

[0027] In this embodiment, the classifier can be a tree-structured classifier or a deep learning classifier. Taking a tree-structured classifier as an example, a tree-structured classifier is a classification model built based on the principle of decision trees. It consists of multiple decision nodes and leaf nodes. Each decision node corresponds to a judgment condition of a feature dimension. By judging layer by layer, the expert quantified feature set is guided to the final classification result.

[0028] Specifically, when the expert-quantified feature set is input into the classifier, the classifier starts from the root node and compares the quantized value of the corresponding dimension of the pixel with a preset threshold according to the feature judgment rule set at the current node. If the quantized value meets the judgment condition, it moves down the corresponding branch to the next decision node and continues to perform feature judgment. This process continues until the expert-quantified feature set is guided to a leaf node, which represents the final classification result of that pixel.

[0029] In practical applications of fire detection, each decision node of the classifier typically sets judgment conditions for key feature dimensions of fire detection. For example, for the temperature feature dimension, the classifier might set a temperature level threshold at a certain decision node. When the temperature quantization level of a pixel is higher than the temperature level threshold, the pixel is considered to have a high probability of being a fire point, and subsequent judgments are made along the branch with high fire point probability. Through this layer-by-layer judgment mechanism, the classifier can comprehensively consider multiple quantization feature dimensions of pixels, achieving accurate fire point identification.

[0030] Ultimately, the classifier can output a corresponding judgment for each individual pixel within the candidate region, typically including classification labels such as fire point, non-fire point, or fire point confidence level. Based on this, the judgment information for all locations within the candidate region can be completely obtained, clearly distinguishing locations belonging to fire points from normal background locations, forming a comprehensive and detailed global fire detection result. This global fire detection result can be used for subsequent advanced processing such as fire point region merging and fire point intensity assessment, providing reliable technical support for fire monitoring.

[0031] The fire detection method for remote sensing images provided in this application, compared with existing remote sensing fire detection technologies, determines an expert feature set within the candidate region of the input remote sensing image. The expert feature set is obtained by standardizing multi-dimensional features representing the differences between fire points and the background in pixels. The expert feature set is then quantized to obtain a quantized expert feature set. This quantized expert feature set is input into a pre-trained classifier for fire detection, resulting in the detection result of each pixel within the candidate region as a fire point. The entire process relies on accurately extracting multi-dimensional features that distinguish fire points from the background from the candidate region of the remote sensing image and standardizing them to construct a unified and standardized expert feature set. Quantization then simplifies the feature data, reducing redundant information and computational overhead. Using a pre-trained classifier for fire point identification, this method fully exploits the feature differences between fire points and the background region, effectively avoiding recognition biases caused by ambient light and atmospheric interference, as well as surface discoloration, significantly improving the accuracy of fire target detection in remote sensing images. Meanwhile, targeted optimizations at the data representation level are achieved through feature standardization and quantization, which are tailored to the operating characteristics of edge hardware and reduce the consumption of computing resources. While ensuring the accuracy of fire detection, it also takes into account the computing speed and deployment adaptability, which can meet the practical application needs of efficient and accurate fire detection of remote sensing images in different usage scenarios.

[0032] In practical fire detection tasks, input remote sensing images typically contain massive amounts of pixel data. Performing complete fire detection processing on every single pixel in the input image would not only consume significant computational resources but also significantly extend processing time, making it difficult to meet the needs of real-time fire monitoring. Therefore, before performing refined fire detection, it is necessary to first screen the input remote sensing image to locate areas suspected of containing fires, focusing subsequent processing on these candidate areas of suspected thermal anomalies. Furthermore, such as... Figure 2 As shown, prior to step 101, the method further includes the following steps: 201. Traverse each pixel in the input remote sensing image and extract the first and second feature values ​​of the pixel in different set bands.

[0033] 202. Calculate the thermal response score of the pixel based on the consistency relationship between the first feature value and the second feature value.

[0034] 203. Select the set of pixels in the input remote sensing image whose thermal response score is greater than the set threshold to form a candidate region for suspected thermal anomalies.

[0035] In this embodiment, the input remote sensing image can be a multispectral and thermal infrared remote sensing image, which includes at least four key spectral channels: near-infrared band, shortwave infrared zone 1 band, shortwave infrared zone 2 band, and thermal infrared band.

[0036] Understandably, to ensure that the input data has accurate radiometric and spatial location information, providing a reliable data foundation for subsequent fire detection and processing, the input remote sensing image can be a remote sensing data block that has already undergone preprocessing steps such as radiometric correction, geometric registration, and invalid value removal, or it can be an image of the area to be detected cropped from the entire remote sensing image according to actual monitoring needs.

[0037] The first characteristic value is the thermal infrared band response value, and the second characteristic value is the near-infrared band reflectance. Generally speaking, the thermal infrared and near-infrared bands are two key sources of spectral information in fire detection. The thermal infrared band can sensitively capture changes in surface temperature; when a fire occurs, the fire area will exhibit obvious temperature anomalies. The near-infrared band has good response characteristics to the burning state of vegetation; burning vegetation will exhibit special reflectance characteristics in the near-infrared band. By simultaneously utilizing information from these two bands and analyzing their consistency, it is possible to effectively distinguish true fire points from ordinary high-temperature objects.

[0038] In practice, it is necessary to traverse every pixel in the input remote sensing image and extract its response value in the thermal infrared band and its reflectance in the near-infrared band. The thermal infrared response value reflects the pixel's temperature characteristics; a higher value indicates a higher temperature. The near-infrared reflectance reflects the pixel's vegetation burning characteristics; this value usually shows abnormal changes during fires. Based on these two feature values, a thermal response score is calculated for each pixel, which comprehensively considers the degree of temperature anomaly and the consistency of spectral characteristics.

[0039] The thermal response score of a pixel can be calculated using the following formula:

[0040] in, Indicates the current cell Thermal response score; Indicates the current cell The response value in the thermal infrared band; Indicates the current cell Near-infrared reflectivity; This represents a small constant that prevents division by zero.

[0041] The thermal response score calculation described above fully considers the typical characteristics of fire points. When a pixel exhibits both high-temperature characteristics and conforms to the spectral characteristics of a fire, its thermal response score will be significantly higher. Conversely, if a pixel has a high temperature but its near-infrared reflectivity does not conform to fire characteristics, its thermal response score will be relatively low. By setting a reasonable scoring threshold, the system can filter out pixels with thermal response scores exceeding the threshold, forming candidate areas for suspected thermal anomalies.

[0042] Accordingly, conditional judgments are performed on each pixel of the input remote sensing area, and the thermal response score of the pixel is calculated. With set threshold Compare them.

[0043] when When a pixel meets the preset fire point candidate conditions, the pixel is extracted and recorded in the candidate region for suspected anomalies; when If a pixel does not meet the preset fire point candidate conditions, it is classified as a background pixel, and further processing of that pixel is terminated, thus avoiding computationally expensive feature extraction and classification operations. This efficient pre-screening mechanism can eliminate the vast majority of non-fire point background pixels across the entire image, retaining only a small number of pixels with fire point potential for subsequent processing. This significantly reduces the computational burden of feature extraction and classification, greatly improving the overall efficiency of the fire point detection algorithm while ensuring complete coverage of potential fire point areas.

[0044] For example, when monitoring a forest area, the system analyzes each pixel in the remote sensing image. Pixels with normal vegetation cover in the forest have thermal infrared response values ​​within the normal range, and their near-infrared reflectance also matches the characteristics of healthy vegetation, resulting in a low thermal response score. However, for areas of active fire, the thermal infrared band detects a significant temperature increase, while the near-infrared band captures the abnormal reflectance caused by vegetation burning. The consistency of these two characteristics makes the thermal response score of this area significantly higher than the threshold, thus accurately identifying it as a candidate area. For exposed rocks under direct sunlight, although the thermal infrared band may detect a high temperature, their near-infrared reflectance does not match the characteristics of a fire, so the thermal response score will not reach the threshold, thus avoiding misjudgment.

[0045] During fire detection, although candidate regions have already eliminated most background interference through initial screening, they may still contain various types of thermal anomaly targets. To further improve the accuracy and reliability of fire detection, it is necessary to extract more comprehensive and refined multi-dimensional fire feature data from candidate regions. By comprehensively considering information from three different dimensions—thermal features, spectral features, and spatial anomaly features—a more complete and robust feature description system can be constructed, thereby effectively distinguishing genuine fire points from various interfering targets. Specifically, for example... Figure 3 As shown, step 101 above includes the following steps: 301. Extract multidimensional fire point features of candidate pixels within the candidate region of the input remote sensing image.

[0046] 302. Standardize the thermal features, the spectral features, and the spatial anomaly features respectively to obtain an expert feature set.

[0047] In this embodiment, the multidimensional fire point features include at least thermal features, spectral features, and spatial anomaly features. Thermal features primarily reflect the degree of temperature anomaly in candidate pixels, effectively capturing the high-temperature characteristics of the fire point. Spectral features characterize the spectral properties of the fire point by analyzing the response relationship of candidate pixels across different wavelengths. Spatial anomaly features focus on the degree of anomaly of candidate pixels within their local spatial environment, quantifying the difference between the pixel and the surrounding background through morphological operations and local statistical analysis.

[0048] After extracting the multidimensional fire point features, these features need to be standardized. Since thermal features, spectral features, and spatial anomaly features may have different dimensions and numerical ranges, directly using these raw features for subsequent processing may lead to some features having unreasonable weights in the classification decision. Standardization can map different features to a uniform numerical range, eliminating dimensional differences and allowing various features to be compared and merged fairly, thus providing high-quality input data for the subsequent classifier.

[0049] In the specific process of extracting multidimensional fire point features of candidate pixels, the third and fourth feature values ​​of candidate pixels in different set bands are extracted within the candidate region of the input remote sensing image; thermal response scores are used as thermal features of candidate pixels; spectral features of candidate pixels are calculated based on the ratios and normalized differences between the second feature value and the third and fourth feature values; and spatial anomaly features of candidate pixels are calculated based on the fourth feature value through morphological white hat operations and statistical analysis of the differences between local peak values ​​and background mean values.

[0050] In this embodiment, the third characteristic value is the reflectance in the shortwave infrared region I band, and the fourth characteristic value is the reflectance in the shortwave infrared region II band. These two bands have special sensitivity to the spectral response of the fire point. The shortwave infrared region I band can effectively capture the high-temperature radiation characteristics of the fire point, while the shortwave infrared region II band has a good ability to identify the combustion products of the fire point.

[0051] The aforementioned thermal characteristics directly employ a pre-calculated thermal response score, which integrates radiation intensity in the thermal infrared band with background temperature difference information, and can intuitively reflect the degree of temperature anomaly in candidate pixels.

[0052] The calculation of the aforementioned spectral characteristics primarily involves the ratio and normalized difference calculation between near-infrared and short-wave infrared reflectance. In a specific implementation, the near-infrared reflectance can be... Compared with the reflectivity of the shortwave infrared region 1 band ratio As a spectral feature; reflectivity in the shortwave infrared second region. Near-infrared band reflectivity ratio As another spectral feature, the standardized burn index is obtained by calculating the normalized difference between the reflectance in the short-wave infrared region II band and the reflectance in the near-infrared band. As another spectral feature, see the formula below. This type of operation can effectively amplify the differences in the spectral response of fire points under specific band combinations, suppress the influence of surface background and atmospheric interference, and enhance the spectral discrimination power of fire point identification.

[0053]

[0054] The aforementioned spatial anomaly features include white-hat enhancement features and peak-background comparison features.

[0055] The white-hat enhancement feature is obtained by performing a morphological white-hat transformation on the fourth feature value. This process is achieved by applying a morphological white-hat transformation to the reflectance of the short-wave infrared second band. This operation can highlight the small bright target structure of candidate pixels relative to the local background. Specifically, it can be implemented using the following formula:

[0056] in, This indicates a white-hat enhanced feature; Represents a nonlinear activation function; Indicates the current cell Reflectivity of the shortwave infrared second band; Indicates adoption The morphological opening operation is performed on the structuring element. This feature emphasizes small, bright protrusions and potential fire targets.

[0057] Peak-background comparison features are obtained by calculating the difference between the maximum value of the local window and the average value of the neighboring window for the fourth feature value. This process involves setting a local analysis window, calculating the difference between the maximum reflectance value within the window and the average reflectance value of the neighboring background region, quantifying the abruptness of candidate pixels in spatial distribution, and capturing the spatial anomalies of isolated fire points in the image. Specifically, this can be achieved using the following formula:

[0058] in, Indicates peak-background comparison features; Represents a nonlinear activation function; Indicates the current cell Reflectivity of the shortwave infrared second band; express The local maximum value of a window; express The local average value of the window. This feature captures sharp peaks relative to the surrounding background by subtracting the mean of the large-scale background from the local peaks.

[0059] Further standardization processing was performed on thermal features, spectral features, and spatial anomaly features to obtain an expert feature set. Specifically, this can be achieved through the following formula:

[0060] in, This represents a single dimension feature in a multi-dimensional feature set that needs to be standardized. and These represent the mean and standard deviation of the feature in this dimension on the training samples, respectively. This represents a small constant that prevents division by zero.

[0061] In fire detection systems, expert feature sets are typically stored as floating-point numbers. While this provides high accuracy, it also consumes significant storage space and is slow to compute, hindering the rapid processing of large-scale remote sensing data. Quantization can convert floating-point features into integer form, significantly reducing storage requirements and computational complexity while maintaining sufficient recognition accuracy. Specifically, for example... Figure 4 As shown, step 102 above includes the following steps: 401. Traverse each dimension feature in the expert feature set, use training samples to statistically determine the low quantile threshold and high quantile threshold of the dimension feature, obtain the lower clipping boundary and upper clipping boundary of the dimension feature, and form the cutoff interval of each dimension feature.

[0062] 402. Calculate the quantization scale factor based on the cutoff interval of each dimension feature.

[0063] 403. Use the truncation interval to prune the expert feature set, and perform quantization processing on the pruned expert feature set based on the quantization scale factor to obtain the expert quantized feature set.

[0064] During the quantization process, it is necessary to traverse every dimension of the expert feature set, statistically analyze the distribution of each dimension using a large number of training samples, and determine the low and high quantiles as the lower and upper clipping boundaries for that dimension, thus forming the cutoff interval for each dimension. This process ensures that the quantization range covers the vast majority of valid samples while eliminating the interference of extreme outliers. The specific clipping boundaries can be achieved using the following formula:

[0065]

[0066] in, This represents the set of features sampled from the training samples; Indicates the first Dimensional features; and They represent the first Low quantile threshold and high quantile threshold for dimensional features; Represents the quantile function; and They represent the first The lower and upper clipping boundaries of the dimensional features. Correspondingly, Indicates the first The truncation interval of dimensional features.

[0067] Next, based on the truncation interval of each feature dimension, the corresponding quantization scaling factor is calculated. The quantization scaling factor is the scaling factor that maps the feature dimension from floating-point values ​​to a signed integer lattice; its calculation needs to consider the width of the truncation interval and the target integer range. Through reasonable scaling factor design, accurate mapping of feature values ​​can be achieved within a limited integer bit width, preserving the discriminative information of the original features to the greatest extent. The specific quantization scaling factor can be implemented using the following formula:

[0068] in, Indicates the first The scaling factor that maps dimensional features from floating-point values ​​to signed integer lattices; Indicates the quantization bit width.

[0069] Finally, the expert feature set is pruned using a truncation interval, restricting feature values ​​exceeding the boundary range to within the valid interval. Then, quantization processing is performed on the pruned expert feature set based on the quantization scaling factor, yielding the final quantized expert feature set. This process converts continuous floating-point features into discrete integer features, ensuring feature validity while significantly reducing data representation complexity. The specific quantization process can be implemented using the following formula:

[0070] in, Represents the set of expert-quantified features; Indicates the integer operation; Indicates truncation operation; Indicates the first A set of expert features after dimensionality trimming; Quantization bit width The upper bound of the quantification is determined.

[0071] In practical applications, the choice of truncation interval type is crucial. Different feature dimensions may exhibit drastically different numerical distribution characteristics; some features are symmetrically distributed around zero, while others show obvious skewed distributions. In this embodiment, the truncation interval is either a symmetrical or asymmetrical truncation interval. The symmetrical truncation interval is centered around zero, meaning the absolute values ​​of the lower and upper truncation boundaries are equal. This approach is suitable for cases where the feature distribution is approximately symmetrical, and it is computationally simple and easy to implement. The asymmetrical truncation interval is determined based on the degree of skewness in the feature distribution, meaning the absolute values ​​of the lower and upper truncation boundaries are not equal. This approach better adapts to the actual distribution of features and retains more effective information. The construction of the truncation interval is based on the deployment requirements of edge devices and is used to limit the impact of outliers on the quantization scale. By reasonably setting the truncation interval, the impact of outliers on the quantization scale factor can be effectively limited, avoiding the stretching of the entire quantization range due to a few extreme values, thus ensuring that most normal samples maintain sufficient accuracy after quantization. This design allows the system to maintain high detection accuracy while meeting the performance requirements of real-time processing when running on edge devices.

[0072] In fire detection scenarios, quantization inevitably introduces information loss. If a classifier is trained directly using unquantized floating-point features and then quantized features are used during inference, the data distribution between training and inference will be inconsistent, severely impacting detection performance. By proactively introducing quantization error during training, the classifier can adapt to the distribution characteristics of quantized features in advance, ensuring stable detection performance during actual deployment. Furthermore, as... Figure 5 As shown, after step 102, the method further includes the following steps: 501. Perform dequantization on the expert quantization feature set to obtain the dequantized floating-point features.

[0073] 502. The difference between the dequantized floating-point features and the expert feature set is used as the quantization error. The quantization error is injected into the training process of the classifier so that the classifier adapts to the quantized feature distribution during inference.

[0074] In this embodiment, the dequantization process first performs a dequantization operation on the expert quantization feature set, converting it back to floating-point form to obtain the dequantized floating-point features. The specific dequantization process can be implemented using the following formula:

[0075] in, This indicates the floating-point characteristics after dequantization.

[0076] Then, by calculating the difference between the dequantized floating-point features and the original expert feature set, the quantization error of each feature dimension is precisely quantized. This error information is systematically injected into the classifier's training process, enabling the classifier to fully consider the impact of quantization during learning.

[0077] Accordingly, during the classifier training process, histogram statistics and node splitting are performed based on the expert quantized feature set; this process directly utilizes the expert quantized feature set for histogram statistics and node splitting operations. This means that each decision node of the classifier is constructed based on the distribution of quantized features, rather than on the original floating-point features. In this way, the classifier can learn the optimal split point in the quantized feature space, forming a decision tree structure that highly matches the distribution of quantized features. This training strategy ensures that the classifier can directly process quantized features during the inference phase without performance loss.

[0078] Correspondingly, during the classifier's inference process, tree traversal and node discrimination are performed based on the expert-quantized feature set, outputting the detection result of each pixel in the candidate region as a fire point. This process performs tree traversal and node discrimination operations based on the expert-quantized feature set. Since the distribution characteristics of the quantized features have been fully considered during the training phase, the classifier can efficiently and accurately detect and judge fire points for each pixel in the candidate region. This end-to-end quantized perception training and inference process not only significantly improves the system's operating efficiency but also ensures the reliability of the detection results.

[0079] In real-world fire detection scenarios, feature extraction for different bands often employs separate computational processes. Each band undergoes separate data reading, feature calculation, and result storage. This decentralized processing leads to numerous redundant memory accesses and cache misses, severely limiting the system's inference efficiency. This embodiment overcomes these problems through feature fusion and continuous cache organization.

[0080] The specific implementation process of feature fusion involves integrating multi-band computation into a single computational flow. In fire detection scenarios, the system needs to process data from multiple bands simultaneously, including shortwave infrared region 1, shortwave infrared region 2, and thermal infrared. Traditional methods extract features from each band independently, requiring the raw data to be read from memory each time. However, through feature fusion, the system completes feature calculations for all bands simultaneously during a single data read, avoiding multiple repetitive data loading operations. This fusion computation not only reduces memory bandwidth usage but also fully utilizes processor computing resources, maximizing computational efficiency.

[0081] Contiguous cache organization optimizes the storage method for candidate pixel features. After feature extraction, the system stores various features corresponding to candidate pixels, including thermal features, spectral features, and spatial anomaly features, in memory in a contiguous and aligned manner. This storage organization ensures that the processor can fully utilize the cache line prefetching mechanism when accessing feature data, significantly reducing cache misses. Aligned storage also ensures that memory accesses meet the processor's alignment requirements, avoiding performance losses caused by unaligned accesses.

[0082] In the classifier design of the fire detection system, the XGBoost tree model classifier is preferred. However, the system architecture offers good flexibility, supporting various alternative classifiers such as LightGBM, CatBoost, and Random Forest. This design allows the system to select the most suitable classifier model based on the needs of different application scenarios and the computing power of edge devices. Regardless of the classifier chosen, the format, dimension, and numerical range of the input features need to be adapted. The system will adjust the feature preprocessing parameters according to the characteristics of the selected classifier to ensure that the feature data can be correctly parsed and utilized by the classifier.

[0083] It should be noted that the implementation of the alternative solution also needs to consider the consistency of model training and deployment. During the training phase, the system uses labeled fire point sample data to train the selected classifier and optimize model parameters. During the deployment phase, the trained model is converted into a format suitable for edge devices to ensure the efficiency of the inference process. This end-to-end training and deployment process guarantees the feasibility of the alternative solution in practical applications. Through this flexible classifier replacement mechanism, the fire detection system can select the optimal classifier model based on actual application needs while ensuring the stability of core functions.

[0084] While the preferred input data source design for the fire detection system utilizes remote sensing images with four key spectral channels—near-infrared, short-wave infrared (SIR) zone I, SIR zone II, and thermal infrared—the system architecture offers high flexibility and adaptability, supporting alternative input data sources. These alternatives can employ other data sources with similar physical meanings, as long as a correspondence between near-infrared, short-wave infrared, and thermal anomaly information can be established.

[0085] For example, alternative multispectral or thermal infrared satellite data can be used to replace the original data source; external thermal data fusion can be used when thermal infrared bands are lacking; when only multispectral data is available, the detection scheme can be reduced to one based on spectral features and spatial anomaly features; and the processing window size and quantization strategy can be adjusted accordingly for data with different spatial resolutions.

[0086] It should be noted that the fire detection technique for the aforementioned remote sensing images is not limited to a single remote sensing platform, but can be applied to a variety of remote sensing data sources that can provide similar physical information, greatly improving the practicality and adaptability of the system.

[0087] Furthermore, as a specific implementation of the above method, embodiments of this application provide a fire detection device for remote sensing images, such as... Figure 6 As shown, the device includes: a determination unit 61, a quantization unit 62, and a detection unit 63.

[0088] The determining unit 61 is used to determine an expert feature set within a candidate region of the input remote sensing image. The expert feature set is obtained by standardizing the multidimensional features in the pixels that characterize the difference between the fire point and the background. Quantization unit 62 is used to quantize the expert feature set to obtain an expert quantized feature set; The detection unit 63 is used to input the expert quantized feature set into a pre-trained classifier to detect fire points and obtain the detection result of each pixel in the candidate region as a fire point.

[0089] The fire detection device for remote sensing images provided in this invention, compared with existing remote sensing fire detection technologies, determines an expert feature set within the candidate region of the input remote sensing image. The expert feature set is obtained by standardizing multi-dimensional features representing the differences between fire points and the background in pixels. The expert feature set is then quantized to obtain a quantized expert feature set. This quantized expert feature set is input into a pre-trained classifier for fire detection, resulting in the detection result of each pixel within the candidate region as a fire point. The entire process relies on accurately extracting multi-dimensional features that distinguish fire points from the background from the candidate region of the remote sensing image and standardizing them to construct a unified and standardized expert feature set. Quantization then simplifies the feature data, reducing redundant information and computational overhead. Using a pre-trained classifier for fire point identification, the device can fully exploit the feature differences between fire points and the background region, effectively avoiding recognition biases caused by ambient light and atmospheric interference, as well as surface discoloration, significantly improving the accuracy of fire target detection in remote sensing images. Meanwhile, targeted optimizations at the data representation level are achieved through feature standardization and quantization, which are tailored to the operating characteristics of edge hardware and reduce the consumption of computing resources. While ensuring the accuracy of fire detection, it also takes into account the computing speed and deployment adaptability, which can meet the practical application needs of efficient and accurate fire detection of remote sensing images in different usage scenarios.

[0090] In specific application scenarios, the device further includes: The extraction unit is used to traverse each pixel in the input remote sensing image before determining the expert feature set in the candidate region of the input remote sensing image, and extract the first feature value and the second feature value of the pixel in different set bands; the first feature value is the thermal infrared band response value, and the second feature value is the near-infrared band reflectance. The calculation unit is used to calculate the thermal response score of the pixel based on the consistency relationship between the first feature value and the second feature value; The selection unit is used to select a set of pixels in the input remote sensing image whose thermal response score is greater than a set threshold, forming a candidate region for suspected thermal anomalies.

[0091] In specific application scenarios, the determining unit is specifically used for: Within the candidate region of the input remote sensing image, multidimensional fire point features of candidate pixels are extracted. The multidimensional fire point features include at least thermal features, spectral features, and spatial anomaly features. The thermal features, spectral features, and spatial anomaly features are standardized respectively to obtain an expert feature set.

[0092] In specific application scenarios, the determining unit is further used for: Within the candidate region of the input remote sensing image, the third and fourth feature values ​​of the candidate pixels are extracted in different set bands; the third feature value is the reflectance of the shortwave infrared zone 1 band, and the fourth feature value is the reflectance of the shortwave infrared zone 2 band. The thermal response score is used as the thermal feature of the candidate pixel; The spectral features of the candidate pixels are calculated based on the ratios between the second feature value and the third and fourth feature values, respectively, and the normalized difference. Based on the fourth feature value, the spatial anomaly features of the candidate pixels are calculated by morphological white hat operation and statistical analysis of the difference between local peak values ​​and background mean values. The spatial anomaly features include white-hat enhancement features and peak-background comparison features; The white-hat enhancement feature is obtained by performing a morphological white-hat transformation on the fourth feature value; The peak-background comparison feature is obtained by calculating the difference between the local window maximum value and the neighborhood window average value of the fourth feature value.

[0093] In specific application scenarios, the quantization unit is specifically used for: Traverse each dimension feature in the expert feature set, use training samples to statistically determine the low quantile threshold and high quantile threshold of the dimension feature, obtain the lower clipping boundary and upper clipping boundary of the dimension feature, and form the cutoff interval of each dimension feature. Based on the cutoff interval of each dimension feature, a quantization scaling factor is calculated, which is the scaling factor for mapping the dimension feature from floating-point values ​​to a signed integer lattice. The expert feature set is clipped using the truncation interval, and then quantized based on the quantization scale factor to obtain the expert quantized feature set.

[0094] In specific application scenarios, the truncation interval can be a symmetrical truncation interval or an asymmetrical truncation interval; the symmetrical truncation interval is centered at zero, and the asymmetrical truncation interval is determined according to the degree of skewness of the feature distribution; wherein, the construction of the truncation interval is based on the deployment requirements of edge devices and is used to limit the impact of outliers on the quantization scale.

[0095] In specific application scenarios, the device further includes: The dequantization unit is used to dequantize the expert feature set after quantizing the expert feature set to obtain the expert quantized feature set, so as to obtain the dequantized floating-point features. An injection unit is used to inject the quantization error into the training process of the classifier, using the difference between the dequantized floating-point features and the expert feature set as the quantization error, so that the classifier adapts to the quantized feature distribution during inference. Accordingly, during the training process of the classifier, histogram statistics and node splitting are performed based on the expert quantized feature set; Accordingly, during the inference process of the classifier, tree traversal and node discrimination are performed based on the expert quantized feature set, and the detection result of each pixel in the candidate region as a fire point is output.

[0096] Based on the above-described method for detecting fire points in remote sensing images, this application also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method for detecting fire points in remote sensing images.

[0097] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0098] Based on the above-described method for detecting fire points in remote sensing images and the virtual device embodiment, in order to achieve the above objectives, this application embodiment also provides a physical device for detecting fire points in remote sensing images. Specifically, it can be a computer, smartphone, tablet computer, smartwatch, server, or network device, etc. The physical device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to implement the above-described method for detecting fire points in remote sensing images.

[0099] Optionally, the physical device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0100] In an exemplary embodiment, see Figure 7 The aforementioned physical device includes a communication bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device. The various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the fire detection method for remote sensing images described in the above embodiments.

[0101] Those skilled in the art will understand that the physical device structure for fire detection of remote sensing images provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0102] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the physical device for fire detection in the aforementioned remote sensing imagery, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware. Compared with existing methods, by applying the technical solution of this application, this application relies on the accurate extraction of multi-dimensional features that can distinguish fire points from the background from candidate regions of remote sensing images, and completes the standardized construction of a unified set of expert features. Then, through quantization processing, the feature data is simplified to reduce redundant information and computational overhead. Fire point identification and judgment are carried out using a pre-trained classifier, which can fully explore the feature differences between fire points and background areas, effectively avoid recognition biases caused by ambient light and atmospheric interference and surface discoloration, and significantly improve the accuracy of fire point target detection in remote sensing images. At the same time, through feature standardization and quantization, targeted optimization at the data representation level is achieved, conforming to the operating characteristics of edge hardware, reducing computational resource consumption, and ensuring fire point detection accuracy while also considering computational speed and deployment adaptability. This can meet the practical application needs of efficient and accurate fire point detection in remote sensing images under different usage scenarios.

[0104] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or they can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0105] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for detecting fire points in remote sensing images, characterized in that, include: Within the candidate region of the input remote sensing image, an expert feature set is determined. The expert feature set is obtained by standardizing the multidimensional features in the pixels that represent the difference between the fire point and the background. The expert feature set is quantized to obtain the expert quantized feature set; The expert-quantified feature set is input into a pre-trained classifier for fire detection, and the detection result of each pixel in the candidate region is obtained as a fire point.

2. The method according to claim 1, characterized in that, Before determining the expert feature set within candidate regions in the input remote sensing image, the method further includes: Traverse each pixel in the input remote sensing image and extract the first feature value and the second feature value of the pixel in different set bands; the first feature value is the thermal infrared band response value and the second feature value is the near-infrared band reflectance. Based on the consistency relationship between the first feature value and the second feature value, the thermal response score of the pixel is calculated; Select a set of pixels in the input remote sensing image whose thermal response score is greater than a set threshold to form a candidate region for suspected thermal anomalies.

3. The method according to claim 2, characterized in that, The step of determining the expert feature set within the candidate region of the input remote sensing image includes: Within the candidate region of the input remote sensing image, multidimensional fire point features of candidate pixels are extracted. The multidimensional fire point features include at least thermal features, spectral features, and spatial anomaly features. The thermal features, spectral features, and spatial anomaly features are standardized respectively to obtain an expert feature set.

4. The method according to claim 3, characterized in that, The step of extracting multidimensional fire point features of candidate pixels within the candidate region of the input remote sensing image includes: Within the candidate region of the input remote sensing image, the third and fourth feature values ​​of the candidate pixels are extracted in different set bands; the third feature value is the reflectance of the shortwave infrared zone 1 band, and the fourth feature value is the reflectance of the shortwave infrared zone 2 band. The thermal response score is used as the thermal feature of the candidate pixel; The spectral features of the candidate pixels are calculated based on the ratios between the second feature value and the third and fourth feature values, respectively, and the normalized difference. Based on the fourth feature value, the spatial anomaly features of the candidate pixels are calculated by morphological white hat operation and statistical analysis of the difference between local peak values ​​and background mean values. The spatial anomaly features include white-hat enhancement features and peak-background comparison features; The white-hat enhancement feature is obtained by performing a morphological white-hat transformation on the fourth feature value; The peak-background comparison feature is obtained by calculating the difference between the local window maximum value and the neighborhood window average value of the fourth feature value.

5. The method according to any one of claims 1-4, characterized in that, The process of quantizing the expert feature set to obtain a quantized expert feature set includes: Traverse each dimension feature in the expert feature set, use training samples to statistically determine the low quantile threshold and high quantile threshold of the dimension feature, obtain the lower clipping boundary and upper clipping boundary of the dimension feature, and form the cutoff interval of each dimension feature. Based on the cutoff interval of each dimension feature, a quantization scaling factor is calculated, which is the scaling factor for mapping the dimension feature from floating-point values ​​to a signed integer lattice. The expert feature set is clipped using the truncation interval, and then quantized based on the quantization scale factor to obtain the expert quantized feature set.

6. The method according to claim 5, characterized in that, The truncation interval can be a symmetrical truncation interval or an asymmetrical truncation interval; the symmetrical truncation interval is centered at zero, and the asymmetrical truncation interval is determined according to the degree of skewness of the feature distribution; wherein, the construction of the truncation interval is based on the deployment requirements of edge devices and is used to limit the impact of outliers on the quantization scale.

7. The method according to any one of claims 1-4, characterized in that, After quantizing the expert feature set to obtain the expert quantized feature set, the method further includes: The expert quantized feature set is dequantized to obtain the dequantized floating-point features; The difference between the dequantized floating-point features and the expert feature set is used as the quantization error. The quantization error is injected into the training process of the classifier so that the classifier can adapt to the quantized feature distribution during inference. Accordingly, during the training process of the classifier, histogram statistics and node splitting are performed based on the expert quantized feature set; Accordingly, during the inference process of the classifier, tree traversal and node discrimination are performed based on the expert quantized feature set, and the detection result of each pixel in the candidate region as a fire point is output.

8. A fire detection device for remote sensing images, characterized in that, include: The determining unit is used to determine an expert feature set within a candidate region of the input remote sensing image. The expert feature set is obtained by standardizing the multidimensional features in the pixels that characterize the difference between the fire point and the background. A quantization unit is used to quantize the expert feature set to obtain an expert quantized feature set. The detection unit is used to input the expert quantized feature set into a pre-trained classifier to detect fire points and obtain the detection result of each pixel in the candidate region as a fire point.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images