Fire point pixel multi-model fusion identification method, device and equipment and storage medium

Abstract

The invention discloses a fire point pixel multi-model fusion identification method and device, equipment and a storage medium. The method comprises the following steps: acquiring to-be-identified remote sensing data, performing compensation preprocessing on the to-be-identified remote sensing data to obtain compensation remote sensing data, and performing terrain preprocessing on the compensation remote sensing data to obtain initial remote sensing data; inputting the initial remote sensing data into a trained basic model to obtain a corresponding basic feature vector, wherein the basic model comprises a space model, a time model and a distinguishing model; and determining a basic vector weight corresponding to the basic feature vector, calculating a fusion feature vector according to the basic feature vector and the corresponding basic vector weight, and outputting a fire point pixel identification result based on the fusion feature vector. According to the scheme, through multi-model fusion and dynamic weight adjustment, the spatial features, the time features and the background distinguishing features of the remote sensing data are captured at the same time. The method is suitable for different remote sensing image qualities and complex scenes, and high-precision and high-stability fire point pixel recognition is achieved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, device, and storage medium for fire point pixel multi-model fusion recognition. Background Technology

[0002] With the widespread application of fire detection and remote sensing technologies, satellite and aerial remote sensing data are increasingly used in areas such as forest fires, agricultural combustion, and urban environmental monitoring. The ability to acquire massive amounts of multi-source remote sensing imagery has been significantly enhanced, providing a data foundation for real-time fire detection and disaster early warning. However, fire pixel identification faces challenges related to multispectral, multi-temporal, and complex terrain conditions. Differences in scale, resolution, and time intervals among different data sources increase the difficulty of data fusion and feature extraction.

[0003] Current technologies primarily rely on single models or static fusion strategies to identify fire points, but these still have significant limitations in practical applications, failing to meet the recognition needs of complex remote sensing scenarios. Existing methods suffer from limitations in data processing, lacking the ability to integrate multi-source remote sensing data, struggling to unify data standards across different sources and time periods, resulting in inconsistent data quality; fire point samples are generally imbalanced, with scarce samples of difficult-to-identify edge fire points, limiting the model's adaptability to complex scenarios. In terms of model construction, a single model struggles to comprehensively capture the multidimensional features of fire points. Traditional CNNs have limitations in extracting large-scale spatial contextual information, LSTMs lack sufficient focus on key time nodes, and Transformers are not highly targeted in distinguishing fire points from background features, failing to simultaneously consider spectral, spatial, and temporal features, thus limiting recognition accuracy. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for multi-model fusion identification of fire point pixels, solving the problem that existing technologies often suffer from limitations in capturing the multi-dimensional features of fire points using a single model, resulting in limited identification accuracy. This method acquires and preprocesses remote sensing data, extracts feature vectors from multiple models, fuses them to generate a fused feature vector, and outputs the fire point pixel identification result based on the fused feature vector. This achieves high-precision identification of fire point pixels and adaptability to complex scenes, making it suitable for multi-source remote sensing image monitoring, forest fire early warning, agricultural combustion monitoring, and other fire point detection scenarios.

[0005] In a first aspect, the present invention provides a method for multi-model fusion recognition of fire point pixels, comprising: Acquire remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data. The initial remote sensing data is input into the trained base model to obtain the corresponding base feature vector. The base model includes a spatial model, a temporal model, and a discrimination model. Determine the base vector weights corresponding to the base feature vectors, calculate the fused feature vectors based on the base feature vectors and their corresponding base vector weights, and output the fire pixel recognition results based on the fused feature vectors.

[0006] Secondly, the present invention provides a fire point pixel multi-model fusion recognition device, comprising: The data preprocessing module is used to acquire remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data. The model input module is used to input the initial remote sensing data into the trained basic model to obtain the corresponding basic feature vector. The basic model includes a spatial model, a temporal model, and a discrimination model. The fire point recognition module is used to determine the basic vector weights corresponding to the basic feature vectors, calculate the fused feature vectors based on the basic feature vectors and their corresponding basic vector weights, and output the fire point pixel recognition results based on the fused feature vectors.

[0007] Thirdly, the present invention provides a fire point pixel multi-model fusion recognition device, comprising: One or more processors; A memory that stores one or more programs that, when executed by one or more processors, enable the one or more processors to implement the fire point pixel multi-model fusion recognition method as described in the first aspect.

[0008] Fourthly, the present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the fire point pixel multi-model fusion recognition method as described in the first aspect.

[0009] This invention constructs a fire point pixel recognition mechanism based on multi-model feature extraction and fusion, achieving efficient processing of remote sensing data, feature fusion, and accurate determination of fire point pixels. After acquiring the remote sensing data to be identified, it is first preprocessed to obtain initial remote sensing data, which is then input into a trained basic model to obtain corresponding basic feature vectors. After obtaining the basic feature vectors, the weights of each feature vector need to be determined, and a fused feature vector is calculated based on the basic feature vectors and their corresponding weights to enhance the expressive power of multi-dimensional features. Subsequently, the fire point pixel recognition result is output based on the fused feature vector, achieving high-precision recognition of fire point pixels in remote sensing images and adaptability to complex scenes. This scheme, through the collaborative design of feature extraction and fusion, significantly improves the accuracy and robustness of fire point recognition, and is suitable for multi-source remote sensing image monitoring, forest fire early warning, agricultural combustion monitoring, and other fire point detection scenarios. Attached Figure Description

[0010] Figure 1 This is a flowchart of a multi-model fusion recognition method for fire point pixels provided in an embodiment of the present invention; Figure 2 This is a flowchart of a remote sensing data fusion method provided in an embodiment of the present invention; Figure 3 This is a flowchart of a remote sensing data adjustment method provided in an embodiment of the present invention; Figure 4 This is a flowchart of a feature vector weight calculation method provided in an embodiment of the present invention; Figure 5 This is a flowchart of a fire point pixel recognition method provided in an embodiment of the present invention; Figure 6 This is a flowchart of a training sample augmentation method provided in an embodiment of the present invention; Figure 7 This is a flowchart of a model reinforcement training method provided in an embodiment of the present invention; Figure 8 This is a flowchart of a fire point pixel multi-model training method provided in an embodiment of the present invention; Figure 9 This is a structural block diagram of a fire point pixel multi-model fusion recognition device provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a fire point pixel multi-model fusion recognition device provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining the present invention and not for limiting the present invention. It should also be noted that, for ease of description, only the parts relevant to the present invention are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. A process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0012] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] Currently, with the development of remote sensing technology and the continuous expansion of fire point monitoring applications, an increasing number of environmental monitoring, forest fire prevention, and agricultural management systems require real-time fire point pixel identification from multi-source remote sensing data. Although these systems serve different monitoring targets, they share many common needs in data processing, model building, and decision-making mechanisms. These typically manifest as efficient unification of data fusion, feature extraction, and model output; integrated processing of multi-source remote sensing images; spatial and temporal feature extraction; differentiation between fire points and background; and collaborative optimization of model decision-making mechanisms. However, achieving high-precision, low-false-alarm fire point identification while maintaining the independence and flexibility of the monitoring system has become a core technical challenge in remote sensing fire point monitoring.

[0014] In current technological practices, fire point pixel recognition mainly relies on single models or static fusion strategies, but these methods all have significant limitations. The ability to integrate and process multi-source remote sensing data is insufficient, making it difficult to unify data standards from different sources and times, resulting in inconsistent data quality. Fire point samples are generally imbalanced, especially with a scarcity of difficult-to-identify samples such as edge fire points, limiting the model's adaptability to complex scenes. A single model struggles to comprehensively capture the multi-dimensional features of fire points. Traditional CNNs have limitations in extracting large-scale spatial contextual information, LSTMs lack focus on key time nodes, and Transformers are not specific enough in distinguishing between fire points and background features, failing to consider spectral, spatial, and temporal features, thus limiting recognition accuracy. Existing methods mostly employ static weight allocation, unable to dynamically adjust weights based on the performance of sub-models in different scenarios and scene characteristics, making the fusion results difficult to adapt to complex environments. Therefore, existing remote sensing fire point recognition methods either fail to comprehensively capture fire point features and have limited recognition accuracy, or lack adaptability in complex scenes and have a high false alarm rate. There is an urgent need for a more robust, efficient, and adaptable fire point pixel recognition technology solution that can handle multi-source data and complex scenes.

[0015] To address the problems of fragmented data processing, insufficient feature extraction, and limited model fusion accuracy in existing remote sensing fire pixel identification methods, this embodiment provides a multi-model fusion identification method for fire pixels. By introducing a multi-model feature extraction and fusion mechanism, a fire pixel identification system with unified data preprocessing, dynamic feature weight allocation, and high-precision identification capabilities is constructed, enabling rapid processing of remote sensing images, efficient calculation of feature fusion, and accurate identification of fire pixels. This method uses the remote sensing data to be identified as input. First, the data is preprocessed to obtain initial remote sensing data, which is then input into a trained basic model to obtain corresponding basic feature vectors. After obtaining the basic feature vectors, the weights of each feature vector are determined, and a fusion feature vector is calculated based on the basic feature vectors and their corresponding weights to enhance the expressive power of multi-dimensional fire pixel features. Subsequently, the fire pixel identification result is output based on the fusion feature vector, achieving high-precision identification of fire pixels in remote sensing images and adaptability to complex scenes. This scheme effectively improves the accuracy and real-time performance of fire pixel identification through a collaborative mechanism of data preprocessing, basic feature extraction, and fusion feature calculation. Meanwhile, this method has good scalability and can be widely applied to multi-source remote sensing image monitoring, forest fire early warning, agricultural combustion monitoring and other fire point detection scenarios, providing key technical support for achieving efficient, accurate and controllable fire point pixel recognition.

[0016] The fire point pixel multi-model fusion recognition method provided in this embodiment can be executed by a fire point pixel multi-model fusion recognition device. This device can be implemented through software and / or hardware. The fire point pixel multi-model fusion recognition device can consist of two or more physical entities, or it can consist of a single physical entity. For example, the fire point pixel multi-model fusion recognition device can be a detection server used for fire point pixel recognition.

[0017] The fire point pixel multi-model fusion recognition device is equipped with at least one type of operating system, including but not limited to Android, Linux, and Windows. The fire point pixel multi-model fusion recognition device can install at least one application based on the operating system. This application can be a built-in application of the operating system or an application downloaded from a third-party device or server. In this embodiment, the fire point pixel multi-model fusion recognition device has at least one application capable of executing the fire point pixel multi-model fusion recognition method.

[0018] For ease of understanding, this embodiment uses the operation and maintenance server as the main body for executing the fire point pixel multi-model fusion recognition method as an example for description.

[0019] Figure 1 A flowchart of a fire point pixel multi-model fusion recognition method provided by an embodiment of the present invention is given. (Reference) Figure 1The fire point pixel multi-model fusion recognition method specifically includes: S110. Obtain the remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data.

[0020] In some embodiments, remote sensing data to be identified is first acquired. This data refers to raw images containing information about surface targets, collected by remote sensing sensors mounted on satellites, drones, or aerial platforms. After acquiring the remote sensing data, it undergoes compensation preprocessing to obtain compensated remote sensing data. This compensation preprocessing involves radiometric correction, geometric correction, and noise suppression to address factors such as sensor noise, atmospheric scattering, and illumination differences. The compensated remote sensing data is a remote sensing image that more accurately reflects the true spectral characteristics and spatial distribution of ground objects. After obtaining the compensated remote sensing data, it undergoes further topographic preprocessing to obtain initial remote sensing data. This topographic preprocessing involves using a digital elevation model to perform topographic correction, slope correction, and occlusion removal on the image. The initial remote sensing data is a remote sensing image that has eliminated the effects of topographic undulations and has a unified spatial reference, which can be used as input for subsequent feature extraction and target identification.

[0021] In one embodiment, the method for determining compensation preprocessing can be: correcting atmospheric effects based on a radiative transfer model and combining this with a geometric registration method to eliminate image distortion.

[0022] In one embodiment, the method for determining terrain preprocessing can be: using a digital elevation model for geometric correction and reducing the impact of terrain shadows through slope and azimuth correction.

[0023] Optionally, Figure 2 A flowchart of a remote sensing data fusion method provided by an embodiment of the present invention is given. (Reference) Figure 2 The remote sensing data fusion method specifically includes: S1101. Obtain the total number of pixels in the remote sensing data to be identified, and detect the number of cloud pixels in the remote sensing data to be identified.

[0024] For example, the total number of pixels in the remote sensing data to be identified is first obtained, and the number of cloud pixels in the remote sensing data to be identified is detected. The remote sensing data to be identified refers to the original image or signal data obtained by satellite, drone or other remote sensing equipment. The total number of pixels refers to the total number of all pixels that make up the remote sensing data, which is used to characterize the spatial resolution or scale of the data. The number of cloud pixels refers to the number of pixels in the remote sensing data that are covered by clouds, which is used to assess the availability of data or to perform cloud occlusion correction.

[0025] In one embodiment, the total number of pixels can be obtained by calculating the total number of pixels based on the row and column pixel dimensions of the remote sensing data.

[0026] In one embodiment, the number of cloud pixels can be detected by analyzing the spectral or brightness characteristics of each pixel through a cloud detection algorithm or threshold judgment, and counting the number of pixels that are determined to be covered by clouds.

[0027] S1102. Calculate the cloud coverage rate based on the total number of pixels and the number of cloud pixels. If the cloud coverage rate is greater than a preset coverage threshold, obtain the microwave remote sensing data corresponding to the remote sensing data to be identified.

[0028] For example, cloud coverage rate is calculated based on the total number of pixels and the number of cloud pixels. When the cloud coverage rate is greater than a preset coverage threshold, microwave remote sensing data corresponding to the remote sensing data to be identified is acquired. Here, the total number of pixels refers to the total number of pixels that make up the remote sensing data, the number of cloud pixels refers to the number of pixels covered by clouds, the cloud coverage rate refers to the ratio of the number of cloud pixels to the total number of pixels, which is used to characterize the degree of cloud occlusion in the remote sensing data, the coverage threshold refers to a pre-set critical value used to judge the availability of data, and the microwave remote sensing data refers to the data acquired by microwave remote sensing equipment, which is used to replace optical remote sensing data for analysis when cloud occlusion is severe.

[0029] In one embodiment, cloud coverage can be calculated by dividing the number of cloud pixels by the total number of pixels to obtain a percentage of cloud coverage.

[0030] In one embodiment, microwave remote sensing data can be obtained by acquiring corresponding microwave observation data from a microwave remote sensing data source or database based on the geographical area or acquisition time of the remote sensing data to be identified.

[0031] S1103. Obtain compensated remote sensing data based on the remote sensing data to be identified and the microwave remote sensing data.

[0032] For example, compensated remote sensing data is obtained based on the remote sensing data to be identified and microwave remote sensing data. The remote sensing data to be identified refers to the surface image data collected by optical remote sensing sensors, and the microwave remote sensing data refers to the remote sensing data with the ability to penetrate clouds and the ground surface obtained by microwave sensors such as synthetic aperture radar. The compensated remote sensing data refers to the remote sensing image that can effectively eliminate cloud cover, atmospheric interference and blind spots after fusing optical remote sensing data and microwave remote sensing data, and can be used as input for subsequent terrain preprocessing and feature extraction.

[0033] In one embodiment, the method for determining the compensated remote sensing data may be: using a multi-source data fusion algorithm to complementarily fuse the hyperspectral details of optical remote sensing images with the penetration information of microwave remote sensing images, and output compensated remote sensing data that combines spatial resolution and spectral characteristics.

[0034] Optionally, Figure 3A flowchart of a remote sensing data adjustment method provided by an embodiment of the present invention is given. (Reference) Figure 3 The specific methods for adjusting remote sensing data include: S1104. Locate the area to be identified corresponding to the compensated remote sensing data, and obtain the regional elevation information corresponding to the area to be identified.

[0035] For example, the region to be identified is located corresponding to the compensated remote sensing data, and the regional elevation information corresponding to the region to be identified is obtained. Here, the compensated remote sensing data refers to the remote sensing image with high-fidelity characteristics obtained after the fusion of optical remote sensing data and microwave remote sensing data, the region to be identified refers to the geographic spatial range in the remote sensing image that needs to be identified and feature extracted, and the regional elevation information refers to the geographic spatial data that characterizes the topographic relief, surface height distribution and landform features of the region to be identified.

[0036] In one embodiment, the method for determining regional elevation information may be: calling a digital elevation model or digital surface model, combining the latitude and longitude boundary information of the area to be identified, extracting elevation data that matches the spatial resolution of the compensated remote sensing data, and using it for subsequent terrain correction and feature restoration.

[0037] In one embodiment, the method for locating the area to be identified can be: determining the corresponding target geographic area through georegistration information or latitude and longitude coordinate mapping of remote sensing data.

[0038] S1105. Determine the regional slope and regional aspect of the region to be identified based on the regional elevation information.

[0039] For example, the regional slope and regional aspect of the area to be identified are determined based on the regional elevation information. The regional elevation information refers to the topographic height data of the area to be identified, the regional slope refers to the angle of inclination of the topographic surface relative to the horizontal plane, which is used to characterize the degree of topographic undulation, the regional aspect refers to the direction of the topographic surface slope, which is used to indicate the slope orientation, and the area to be identified refers to the target observation geographic area in the remote sensing data.

[0040] In one embodiment, the slope of a region can be determined by using regional elevation data to calculate the rate of change of elevation within the neighborhood, thereby obtaining the distribution of terrain tilt angles.

[0041] In one embodiment, the regional slope aspect can be determined by calculating the slope orientation through regional elevation gradient analysis to generate slope aspect information for subsequent spatial feature analysis.

[0042] S1106. Adjust the spectral characteristics in the compensated remote sensing data according to the slope and aspect of the region to obtain the initial remote sensing data.

[0043] For example, initial remote sensing data is obtained by adjusting and compensating for the spectral characteristics in the regional slope and aspect of the remote sensing data. Here, regional slope refers to the geometric parameter characterizing the rate of change of surface elevation in a two-dimensional plane, used to describe the degree of topographic relief; regional aspect refers to the orientation parameter characterizing the surface slope relative to the geographic coordinate orientation, used to reflect the relative relationship between the topography and the direction of solar incidence; spectral characteristics refer to the attribute information of reflectance, radiance, and spectral energy distribution in different bands in the compensated remote sensing data; and initial remote sensing data refers to remote sensing image data with consistent spectral characteristics after topographic effect correction.

[0044] In one embodiment, the method for adjusting spectral characteristics can be as follows: a terrain correction model is constructed based on regional slope and regional aspect, and the solar incidence angle, surface reflection angle and terrain parameters are coupled and corrected. The reflectance of each band in the compensated remote sensing data is normalized, thereby eliminating the influence of terrain shadow and aspect differences on spectral characteristics and obtaining initial remote sensing data.

[0045] S120. Input the initial remote sensing data into the trained basic model to obtain the corresponding basic feature vector. The basic model includes a spatial model, a temporal model, and a distinguishing model.

[0046] In some embodiments, initial remote sensing data is input into a trained base model to obtain corresponding base feature vectors. Here, initial remote sensing data refers to preprocessed remote sensing data to be identified, used to characterize the land cover information or environmental state of the target area. The base model refers to a machine learning or deep learning model obtained through training for feature extraction, including a spatial model, a temporal model, and a discrimination model. The spatial model refers to a model used to extract the spatial distribution features of remote sensing data, the temporal model refers to a model used to extract the time series features of remote sensing data, the discrimination model refers to a model used to distinguish different categories of land cover or target states, and the base feature vector refers to the vector representation output by the base model used to describe the spatial, temporal, and category features of the remote sensing data.

[0047] In one embodiment, the method of inputting initial remote sensing data into the base model to obtain the base feature vector can be as follows: the preprocessed remote sensing image or signal data is sequentially input into the spatial model, the temporal model and the differentiation model, and the base feature vector is generated through the feature extraction layer, which can be used for subsequent differential analysis or target recognition.

[0048] S130. Determine the basic vector weights corresponding to the basic feature vectors, calculate the fused feature vectors based on the basic feature vectors and their corresponding basic vector weights, and output the fire pixel recognition result based on the fused feature vectors.

[0049] In some embodiments, the basic vector weights corresponding to the basic feature vectors are first determined, and the fused feature vector is calculated based on the basic feature vectors and their corresponding basic vector weights. The fire pixel recognition result is output based on the fused feature vector. Here, the basic feature vector refers to the vector extracted by the trained basic model to describe the spatial, temporal and categorical features of remote sensing data, the basic vector weight refers to the weight parameters used to measure the importance of different basic feature vectors in the fusion calculation, the fused feature vector refers to the comprehensive feature representation obtained by combining the basic feature vectors through weighting or other fusion algorithms, and the fire pixel recognition result refers to the location or category information of fire pixels in the remote sensing image determined based on the fused feature vector.

[0050] In one embodiment, the method for calculating the fused feature vector can be: multiplying the basic feature vector by its corresponding basic vector weights and then performing a weighted summation or other fusion operation to generate a comprehensive feature representation.

[0051] In one embodiment, the method for outputting fire point pixel recognition results based on fused feature vectors can be: inputting the fused feature vectors into a classifier or decision model to obtain the fire point category or probability information of each pixel, and generating the final recognition result.

[0052] Optionally, Figure 4 A flowchart of a feature vector weight calculation method provided by an embodiment of the present invention is given. (Reference) Figure 4 The specific method for calculating the feature vector weights includes: S1301. Obtain the recognition confidence level corresponding to the basic feature vector and the standard remote sensing data corresponding to the basic model.

[0053] For example, the identification confidence level corresponding to the basic feature vector and the standard remote sensing data corresponding to the basic model are obtained. The basic feature vector refers to the vector extracted by the trained basic model to describe the spatial, temporal and categorical features of the remote sensing data. The identification confidence level refers to the numerical value used to characterize the reliability or prediction accuracy of the basic feature vector in the target identification task. The basic model refers to the machine learning or deep learning model obtained through training and used for feature extraction and target identification. The standard remote sensing data refers to the remote sensing image or signal data containing known target annotations used to train the basic model and is used as a reference or correction basis.

[0054] In one embodiment, the confidence level of identification can be obtained by calculating the confidence value corresponding to the basic feature vector through the output probability of the basic model or a confidence scoring mechanism.

[0055] In one embodiment, standard remote sensing data can be obtained by reading standard remote sensing data corresponding to the base model from a model training database or a pre-stored labeled dataset.

[0056] S1302. Calculate the data similarity between the remote sensing data to be identified and the standard remote sensing data.

[0057] For example, the data similarity between the remote sensing data to be identified and the standard remote sensing data is calculated. The remote sensing data to be identified refers to remote sensing image or signal data that has been preprocessed and has its basic feature vector extracted. The standard remote sensing data refers to remote sensing data used to train the basic model and containing known target annotations. The similarity refers to a numerical value or index used to measure the degree of matching between the remote sensing data to be identified and the standard remote sensing data in spatial, spectral, or temporal features, and is used to assist in evaluating the reliability or confidence of the basic feature vector.

[0058] In one embodiment, the data similarity can be calculated by obtaining a similarity index between the remote sensing data to be identified and the standard remote sensing data through Euclidean distance, cosine similarity, or other distance measurement methods between feature vectors.

[0059] S1303. Calculate the basic vector weights corresponding to the basic feature vectors based on the recognition confidence and the data similarity.

[0060] For example, the basic vector weights corresponding to the basic feature vectors are calculated based on the recognition confidence and data similarity. Here, the basic feature vector refers to the vector extracted by the trained basic model to describe the spatial, temporal, and categorical features of remote sensing data. The recognition confidence refers to the numerical value used to characterize the reliability or prediction accuracy of the basic feature vector in the target recognition task. The data similarity refers to the degree of matching between the remote sensing data to be identified and the standard remote sensing data in terms of spatial, spectral, or temporal features. The basic vector weight refers to the weight parameters used to measure the importance or contribution of different basic feature vectors in the fusion calculation.

[0061] In one embodiment, the weights of the basic vectors can be calculated by combining the identification confidence and data similarity through weighting, normalization, or other fusion algorithms to generate weight values ​​that reflect the importance of the basic feature vectors.

[0062] Optionally, Figure 5 A flowchart of a fire point pixel recognition method provided by an embodiment of the present invention is given. (Reference) Figure 5 The fire point pixel recognition method specifically includes: S1304. Calculate the fire point identification probability corresponding to each pixel in the remote sensing data to be identified using the Softmax function, wherein the Softmax function is set with a temperature coefficient.

[0063] For example, the fire point identification probability corresponding to each pixel in the remote sensing data to be identified is calculated by the Softmax function. Here, the remote sensing data to be identified refers to the remote sensing image or signal data that has been preprocessed and the fused feature vector has been extracted. A pixel refers to the basic spatial unit that constitutes the remote sensing data. The fire point identification probability refers to the numerical value representing the possibility that each pixel is a fire point. The Softmax function refers to the mathematical function used to map the fused feature vector to a probability distribution. The temperature coefficient refers to the parameter used in the Softmax calculation to adjust the smoothness of the probability distribution in order to avoid the extreme probability being too concentrated or too dispersed.

[0064] In one embodiment, the fire point recognition probability can be calculated by inputting the fused feature vector into the Softmax function and normalizing it in combination with the temperature coefficient to obtain the fire point probability value corresponding to each pixel, which is then used to generate the fire point pixel recognition result.

[0065] In one embodiment, the formula for calculating the Softmax function is as follows: in, For the fused feature vector One output value, For the number of categories, Used to adjust the concentration of the output probability distribution.

[0066] S1305. Output the fire point pixel identification result corresponding to each pixel in the remote sensing data to be identified according to the fire point identification probability.

[0067] For example, the fire point identification result is output based on the fire point identification probability of each pixel in the remote sensing data to be identified. Here, the fire point identification probability refers to the probability value of each pixel being a fire point calculated by the Softmax function, the pixel refers to the basic spatial unit that constitutes the remote sensing data, and the fire point identification result refers to the classification result of whether the pixel belongs to a fire point based on the fire point identification probability.

[0068] In one embodiment, the method for outputting the fire point pixel recognition result can be: comparing the fire point recognition probability of each pixel with a preset probability threshold; if the probability value is greater than the threshold, then the pixel is output as a fire point pixel; otherwise, it is output as a non-fire point pixel, thereby obtaining a complete set of recognition results.

[0069] Optionally, Figure 6 A flowchart of a training sample augmentation method provided by an embodiment of the present invention is given. (Reference) Figure 6 The training sample augmentation method specifically includes: S101. Obtain the remote sensing data sample corresponding to the basic model, and extract the fire point area and fire point edge area from the remote sensing data sample.

[0070] For example, remote sensing data samples corresponding to the base model are obtained, and fire point regions and fire point edge regions are extracted from the remote sensing data samples. Here, the base model refers to the training model used to characterize the spatial, temporal, and discriminative features in the fire point identification process, the remote sensing data samples refer to historical remote sensing image data used for model training or verification, the fire point region refers to a continuous region in the remote sensing data samples that exhibits high temperature or high radiation characteristics of a fire, and the fire point edge region refers to the adjacent region surrounding the fire point region where the temperature gradient or spectral features gradually transition.

[0071] In one embodiment, the method for extracting the fire point region and the fire point edge region can be: performing spectral feature analysis and threshold segmentation on remote sensing data samples, and combining spatial connectivity detection to distinguish the fire point region from its surrounding edge region, thereby providing accurate annotations for subsequent model feature extraction and training.

[0072] S102. Generate fire point augmentation samples based on the fire point region, and generate edge augmentation samples based on the fire point edge region.

[0073] For example, fire point augmentation samples are generated based on fire point regions, and edge augmentation samples are generated based on fire point edge regions. Here, the fire point region refers to a continuous region in the remote sensing data sample that exhibits high temperature or high radiation characteristics of a fire, the edge region refers to the adjacent region surrounding the fire point region and whose spectral characteristics or temperature gradients gradually transition, the fire point augmentation sample refers to the training sample generated by data augmentation, cropping or transformation of the fire point region, and the edge augmentation sample refers to the training sample generated by feature amplification, perturbation or translation of the edge region.

[0074] In one embodiment, the augmented samples can be generated by applying operations such as rotation, scaling, random noise addition, and spectral enhancement to the fire point region and the edge region respectively, so as to improve the model's ability to learn fire point features and edge transition features in different scenarios, thereby expanding the diversity of training samples.

[0075] S103. Merge the remote sensing data samples, the fire point augmentation samples, and the edge augmentation samples to obtain model training samples, and use the model training samples to train the base model.

[0076] For example, remote sensing data samples, fire point augmentation samples, and edge augmentation samples are merged to obtain model training samples, and the base model is trained using the model training samples. Here, remote sensing data samples refer to historical remote sensing image data used for model training or validation, fire point augmentation samples refer to data augmentation samples generated based on fire point areas, edge augmentation samples refer to feature extension samples generated based on fire point edge areas, model training samples refer to a comprehensive training dataset composed of samples from the above different sources, and the base model refers to a deep learning model used to characterize the spatial features, temporal features, and discriminative features in the fire point identification process.

[0077] In one embodiment, the training of the base model can be achieved by inputting model training samples into a deep neural network, using supervised learning methods to iteratively optimize the model parameters, and combining cross-entropy loss function or contrastive loss function to improve the model's ability to distinguish between fire zones and non-fire zones.

[0078] Optionally, Figure 7 A flowchart of a model reinforcement training method provided by an embodiment of the present invention is given. (Reference) Figure 7 The specific methods for enhancing the training of this model include: S104. Input the remote sensing data samples into the basic model to obtain the sample recognition results, and determine the recognition accuracy of the basic model based on the sample recognition results.

[0079] For example, remote sensing data samples are input into a base model to obtain sample recognition results. The recognition accuracy of the base model is determined based on the sample recognition results. Here, remote sensing data samples refer to historical remote sensing image data used for model validation, the base model refers to a deep learning model used to characterize the spatial, temporal, and discriminative features in the fire point recognition process, the sample recognition results refer to the classification or detection results output by the base model for the input remote sensing data samples, and the recognition accuracy refers to the degree of matching between the sample recognition results and the pre-labeled results.

[0080] In one embodiment, the recognition accuracy can be determined by comparing the sample recognition results with the corresponding manual annotation results pixel by pixel, and counting the proportion of correctly recognized pixels to the total number of pixels, thereby obtaining the recognition accuracy of the basic model in the verification stage.

[0081] S105. If the recognition accuracy is less than a preset accuracy threshold, reinforcement learning is used to reinforce the basic model.

[0082] For example, when the recognition accuracy is less than a preset accuracy threshold, reinforcement learning is used to reinforce the training of the base model. Here, the recognition accuracy refers to the degree of matching between the recognition result of the base model and the annotation result of the input remote sensing data sample. The preset accuracy threshold refers to the reference value used to determine whether the model meets the application requirements during the model training stage. Reinforcement learning refers to the training method of updating the model parameters by introducing a reward function and feedback mechanism. The base model refers to the deep learning model used to characterize the spatial features, temporal features and discriminative features in the fire point recognition process.

[0083] In one embodiment, reinforcement training can be achieved by generating a reward signal based on the difference between the sample recognition result and the annotation result, feeding the reward signal back to the base model, and iteratively optimizing the model parameters through policy gradient or value function methods to improve the model's generalization ability and stability in the fire detection task.

[0084] Optionally, Figure 8 A flowchart of a fire point pixel multi-model training method provided by an embodiment of the present invention is given. (Reference) Figure 8 The fire point pixel multi-model training method specifically includes: S201. Construct a dedicated dataset of fire point pixels and preprocess it to extract spatiotemporal joint features.

[0085] For example, a dedicated dataset of fire point pixels is constructed and preprocessed to extract spatiotemporal joint features. The dedicated dataset of fire point pixels refers to a training or validation dataset selected from remote sensing data that contains only fire point pixels and their neighboring pixels. It is used to specifically analyze fire point features. Preprocessing refers to operations such as denoising, correction, registration or normalization on the original remote sensing data. Spatiotemporal joint features refer to a comprehensive feature representation that simultaneously contains spatial and temporal information.

[0086] In one embodiment, the method for constructing a fire point pixel-specific dataset can be: filtering pixels labeled as fire points and their surrounding pixels from the initial remote sensing data to form a subset of data; In one embodiment, the spatiotemporal joint features can be extracted by performing statistical, convolutional, or sequence modeling on each pixel in combination with its neighboring pixels and multi-temporal observations to generate a feature vector that simultaneously reflects spatial distribution and temporal changes.

[0087] In one embodiment, the preprocessing method may include: collecting multi-source remote sensing images and fire point annotation data, cleaning up invalid samples, performing radiometric normalization, geometric correction, and cloud / snow masking on the images, and extracting spectral, spatial, and temporal joint features. To alleviate the problem of imbalanced sample distribution, an improved SMOTE algorithm is used to expand the dataset, and generative adversarial networks are combined to synthesize edge fire point samples, thereby improving the coverage of training samples.

[0088] In one embodiment, the preprocessing method may be as follows: Collect MODIS and Landsat satellite imagery as multi-source remote sensing data, combine it with ground observation station records and manual annotations to obtain fire point annotation data. Remove images with a signal-to-noise ratio less than 20 dB, blurry or missing annotations, and outliers with pixel values ​​exceeding a reasonable range. Convert image radiance values ​​to reflectance or brightness temperature to eliminate the influence of atmospheric and sensor differences. Use ground control points to correct the images to a unified coordinate system, ensuring that the geometric error does not exceed one pixel. Use a threshold method, for example, setting reflectance > 0.8 and temperature < 273 K to define it as snow, and use the Fmask algorithm to generate a mask to block cloud and snow areas.

[0089] In one embodiment, the sample augmentation method can be as follows: randomly select samples from the k nearest neighbors of the minor samples, generate new samples through linear interpolation to avoid excessive smoothing of edge features in the original SMOTE; use a generator to learn the distribution of fire point features, and a discriminator to distinguish between real and synthetic samples, and finally generate abnormal areas or partially occluded fire point samples to enhance the diversity of the dataset.

[0090] In one embodiment, the GDAL library can be used to read and convert multi-source data, OpenCV can be used to perform radiometric normalization and geometric correction, and TensorFlow can be used to generate GAN samples.

[0091] S202. Divide the dataset into training set, validation set and test set, input the base model group for independent training, and obtain the initial recognition results and confidence of each sub-model.

[0092] For example, the fire point pixel-specific dataset is divided into a training set, a validation set, and a test set, and then fed into the base model group for independent training to obtain the initial recognition results and confidence scores of each sub-model. Here, the fire point pixel-specific dataset refers to a subset of data containing only fire point pixels and their neighboring pixels; the training set refers to the subset of data used to optimize model parameters; the validation set refers to the subset of data used to evaluate model performance and adjust hyperparameters; the test set refers to the subset of data used to finally evaluate the model's generalization ability; the base model group refers to a collection of multiple trained base models; each sub-model refers to an individual model in the base model group; the initial recognition result refers to the classification or detection output of each sub-model on the input sample; and the confidence score refers to a numerical value used to characterize the reliability of the output results of each sub-model.

[0093] In one embodiment, the dataset can be divided into a training set, a validation set, and a test set based on time, space, or random sampling strategies. In one embodiment, the training of the base model set can be done by: inputting the training set into each sub-model, optimizing the parameters independently, evaluating the performance using the validation set, and outputting the initial recognition results and corresponding confidence scores.

[0094] In one embodiment, the dataset can be divided into training, validation, and test sets, which are then fed into the base model group for independent training to obtain the initial recognition results and corresponding confidence scores of each sub-model. The base model group includes an improved CNN, an attention-gated LSTM, and a Transformer with masked self-attention. The improved CNN introduces dilated convolutions to capture large-scale spatial contextual information; the attention-gated LSTM focuses on temperature changes at key time points; and the Transformer with masked self-attention enhances the distinction between fire points and the background.

[0095] In one embodiment, the dataset can be divided into training, validation, and test sets in a 7:2:1 ratio. The improved CNN training uses dilated convolutions with dilation rates of 2 and 4 to expand the receptive field and capture the surrounding context of the fire point. The optimizer is Adam, with an initial learning rate of 0.001, and training continues until the loss converges. The attention-gated LSTM training introduces gating units in the hidden layers, learning the weights of temperature abrupt changes to focus on key time points; its training parameters are the same as the CNN. The Transformer training with masked self-attention applies a mask to the background region, forcing the model to focus on the spectral contrast features between the fire point and the background. A masked self-attention mechanism is used to strengthen the attention weights in the fire point region; the training parameters are the same as the CNN.

[0096] In one embodiment, spectral features, spatial features, and temporal features are extracted through various sub-models. The spectral features include band reflectance, NDVI, and FIR; the spatial features include pixel gradient and gray-level co-occurrence matrix; and the temporal features include the temperature change rate of continuous images and the duration of fire points.

[0097] In one embodiment, CNN, LSTM, and Transformer sub-models can be built based on PyTorch and an integrated training interface can be provided.

[0098] S203. Construct a dynamic weight fusion layer based on the attention mechanism. Adjust the weights in real time according to the recognition accuracy of the sub-model on the validation set and the similarity between the current scene and the optimal adaptation scene of the sub-model, and generate a fused feature vector.

[0099] For example, a dynamic weight fusion layer is constructed based on the attention mechanism. The weights are adjusted in real time according to the recognition accuracy of the sub-model on the validation set and the similarity between the current scene and the optimal fit scene of the sub-model, generating a fused feature vector. Here, the attention mechanism refers to the mechanism of dynamically adjusting the influence of each sub-model by calculating the importance or contribution of each sub-model under a specific input. The dynamic weight fusion layer refers to the calculation module that combines the attention mechanism to weight and combine the outputs of each sub-model. The sub-model recognition accuracy refers to the classification or detection accuracy of each sub-model on the validation set. The optimal fit scene refers to the set of scene features that each sub-model performs best in the training or validation process. The similarity between the current scene and the optimal fit scene refers to the degree of matching between the quantitative input sample and the scene features most suitable for the sub-model. The fused feature vector refers to the comprehensive feature representation generated by weighting and combining the output features of each sub-model according to the dynamic weights.

[0100] In one embodiment, the weights can be adjusted in real time by calculating weight coefficients based on the sub-model recognition accuracy and scene similarity, and then normalizing them through an attention mechanism to generate a fusion feature vector that reflects the contribution of each sub-model.

[0101] In one embodiment, a dynamic weight fusion layer can be constructed based on an attention mechanism. The weights of each sub-model are adjusted in real time according to the recognition accuracy of the sub-model on the validation set and the similarity between the current scene and the optimal adaptation scene of the sub-model, generating a fused feature vector. The weight adjustment formula is as follows: in, For the first Sub-model weights, The accuracy of the sub-model on the validation set. The similarity between the current scene and the optimally adapted scene of this sub-model is given. and For learnable parameters, and The initial value was set to 0.5, and it was gradually optimized during training iterations.

[0102] S204. Input the fused feature vector into the decision layer, output the final recognition result through the improved Softmax function, and calculate the accuracy.

[0103] For example, the fused feature vector is input into the decision layer, and the final recognition result is output through the improved Softmax function, and the accuracy is calculated. Here, the fused feature vector refers to the comprehensive feature representation generated by fusing the outputs of each sub-model based on dynamic weights; the decision layer refers to the calculation module used to map the fused feature vector to the pixel classification or fire point recognition result; the improved Softmax function refers to the function that introduces a temperature coefficient or weighting mechanism on the basis of the traditional Softmax to optimize the probability distribution; the final recognition result refers to the fire point category or probability information of each pixel output by the decision layer; and the accuracy refers to the degree of matching between the final recognition result and manually labeled or reference data.

[0104] In one embodiment, the final recognition result and accuracy calculation can be achieved by: inputting the fused feature vector into the improved Softmax function to generate a pixel-level fire point probability distribution, comparing it with a preset probability threshold to obtain the classification result, and simultaneously comparing the recognition result with the labeled data pixel by pixel, and calculating the proportion of correctly recognized pixels to the total number of pixels to calculate the accuracy.

[0105] In one embodiment, the fused feature vector is input into the decision layer, and the final recognition result is output through an improved Softmax function to calculate the accuracy. The improved Softmax function is shown below: in, For the fused feature vector One output value, For the number of categories, Temperature coefficient used to adjust the concentration of the output probability distribution The value is set to 2 in blurred scenes and 0.5 in environments with obvious high-temperature fire points; that is, a smaller value is used for blurred scenes. For complex scenarios, take a larger value. .

[0106] In some embodiments, fire point pixel coordinates and confidence heatmaps can be output, and Shapefile and GeoTIFF formats can be exported.

[0107] S205. If the accuracy does not reach the threshold, the basic model group parameters and dynamic weight fusion layer strategy are optimized in reverse through reinforcement learning algorithm until the model converges.

[0108] For example, if the accuracy does not reach the threshold, the parameters of the base model group and the dynamic weight fusion layer strategy are optimized in reverse through a reinforcement learning algorithm until the model converges. Here, accuracy refers to the degree of matching between the final recognition result output by the decision layer and the labeled data, the threshold refers to the pre-set reference value for target recognition accuracy, reinforcement learning algorithm refers to the training method that iteratively optimizes the model parameters and fusion strategy through reward signals and feedback mechanisms, base model group refers to the model set composed of multiple sub-models, dynamic weight fusion layer strategy refers to the weight adjustment strategy that weights the outputs of each sub-model based on the attention mechanism, and model convergence means that the accuracy or reward value reaches a stable state in continuous training iterations and no longer changes significantly.

[0109] In one embodiment, the reverse optimization method can be: generating a reward signal based on the difference between the final recognition result and the labeled data, and using the policy gradient or value function method to iteratively update the parameters of the basic model group and the weights of the fusion layer, gradually improving the overall recognition accuracy until convergence.

[0110] In one embodiment, if the recognition accuracy does not reach a preset threshold, a reinforcement learning algorithm is invoked to back-optimize the parameters of the base model group and the dynamic weight fusion layer strategy until the model converges. The preset threshold can be set to 95%, and the optimization objective function is: in, To improve recognition accuracy, For false alarm rate, Balance coefficient (0 < <1), It can be set to 0.7. When the accuracy meets the requirements, the optimized parameters and strategies are saved and a report is generated.

[0111] S206. In actual recognition, if there is interference in the input image, a robust compensation mechanism will be automatically triggered.

[0112] For example, in actual recognition, if there is interference in the input image, a robust compensation mechanism is automatically triggered. Here, the input image refers to the remote sensing data actually acquired during the fire point pixel recognition process, the interference refers to the degradation of remote sensing data quality or abnormal features caused by cloud cover, noise, sensor error or environmental factors, and the robust compensation mechanism refers to an automated method used to adjust the model output or feature processing strategy to offset the influence of interference and improve the stability and accuracy of recognition.

[0113] In one embodiment, the robust compensation mechanism can be triggered by detecting cloud coverage, noise level, or feature anomaly indicators in the input image. When these exceed a preset threshold, the compensation algorithm is automatically activated to adjust the weights of the fused feature vector or sub-model to maintain the reliability of the fire pixel recognition results.

[0114] In one embodiment, cloud coverage can be calculated using the CFMAS algorithm. When cloud coverage exceeds 30%, Sentinel-1 SAR microwave remote sensing data is called to supplement fire point scattering information by utilizing its cloud penetration characteristics. When complex terrain exists, NASA SRTM DEM data is called to correct spectral deviations caused by slope and aspect, and to adjust the pixel spectral feature weights. The optimal model combination parameters of similar historical scenes are called to assist in identification.

[0115] Based on the above embodiments, Figure 9 This is a structural block diagram of a fire point pixel multi-model fusion recognition device provided in an embodiment of the present invention. (Reference) Figure 9 The fire point pixel multi-model fusion recognition device provided in this embodiment specifically includes: a data preprocessing module 21, a model input module 22, and a fire point recognition module 23.

[0116] The data preprocessing module 21 is configured to acquire remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data; the model input module 22 is configured to input the initial remote sensing data into a trained basic model to obtain corresponding basic feature vectors, wherein the basic model includes a spatial model, a temporal model, and a distinguishing model; the fire point identification module 23 is configured to determine the basic vector weights corresponding to the basic feature vectors, calculate a fused feature vector based on the basic feature vectors and their corresponding basic vector weights, and output the fire point pixel identification result based on the fused feature vectors.

[0117] Based on the above embodiments, the data preprocessing module 21 includes: a pixel count unit, configured to acquire the total number of pixels in the remote sensing data to be identified and detect the number of cloud pixels in the remote sensing data to be identified; a microwave data unit, configured to calculate cloud coverage based on the total number of pixels and the number of cloud pixels, and acquire microwave remote sensing data corresponding to the remote sensing data to be identified if the cloud coverage is greater than a preset coverage threshold; and an initial data unit, configured to obtain initial remote sensing data based on the remote sensing data to be identified and the microwave remote sensing data.

[0118] Based on the above embodiments, the data preprocessing module 21 further includes: an elevation data unit, configured to locate the region to be identified corresponding to the remote sensing data to be identified, and obtain regional elevation information corresponding to the region to be identified; a slope data unit, configured to determine the regional slope and regional aspect corresponding to the region to be identified based on the regional elevation information; and a data adjustment unit, configured to adjust the spectral features in the remote sensing data to be identified according to the regional slope and the regional aspect to obtain initial remote sensing data.

[0119] Based on the above embodiments, the fire point identification module 23 includes: a confidence data and standard data unit, configured to acquire the identification confidence level corresponding to the basic feature vector and the standard remote sensing data corresponding to the basic model; a data similarity unit, configured to calculate the data similarity between the remote sensing data to be identified and the standard remote sensing data; and a vector weight unit, configured to calculate the basic vector weight corresponding to the basic feature vector based on the identification confidence level and the data similarity.

[0120] Based on the above embodiments, the fire point pixel multi-model fusion recognition device further includes a model training module, which includes: a small number of category units configured to acquire remote sensing data samples corresponding to the basic model and extract fire point regions and fire point edge regions from the remote sensing data samples; a data augmentation unit configured to generate fire point augmentation samples based on the fire point regions and generate edge augmentation samples based on the fire point edge regions; and a data merging unit configured to merge the remote sensing data samples, the fire point augmentation samples, and the edge augmentation samples to obtain model training samples, and train the basic model using the model training samples.

[0121] Based on the above embodiments, the model training module further includes: a recognition accuracy unit, configured to input remote sensing data samples into the basic model to obtain sample recognition results, and determine the recognition accuracy of the basic model based on the sample recognition results; and a reinforcement training unit, configured to perform reinforcement training on the basic model using reinforcement learning when the recognition accuracy is less than a preset accuracy threshold.

[0122] Based on the above embodiments, the fire point identification module 23 further includes: a fire point probability unit, configured to calculate the fire point identification probability corresponding to each pixel in the remote sensing data to be identified through a Softmax function, wherein the Softmax function is set with a temperature coefficient; and a result output unit, configured to output the fire point pixel identification result corresponding to each pixel in the remote sensing data to be identified according to the fire point identification probability.

[0123] The activity deployment budget management device provided in the above embodiments of the present invention realizes an end-to-end processing link from remote sensing data acquisition and feature extraction to fire point pixel recognition by constructing a hierarchical processing architecture consisting of a data preprocessing module 21, a model input module 22, and a fire point recognition module 23. This device can efficiently acquire remote sensing data to be identified, extract and fuse features from multiple models, and output fire point pixel recognition results, thus automating the fire point recognition task. Specifically, the data preprocessing module 21 has remote sensing data acquisition and preprocessing capabilities, used to acquire the remote sensing data to be identified and preprocess it to obtain initial remote sensing data, providing basic information for subsequent model feature extraction; the model input module 22 is used to input the initial remote sensing data into the trained basic model to obtain the corresponding basic feature vector, where the basic model includes a spatial model, a temporal model, and a discriminative model, providing a multi-dimensional description for fire point feature extraction; the fire point recognition module 23 is used to determine the basic vector weights corresponding to the basic feature vectors, calculate the fused feature vector based on the basic feature vectors and their corresponding weights, and finally output the fire point pixel recognition results based on the fused feature vector, achieving accurate identification of fire point targets. Through the collaborative processing of the above modules, the device constructs a complete closed-loop process from data acquisition and feature extraction to fire point recognition result output, which significantly improves the efficiency and accuracy of fire point pixel recognition.

[0124] The fire point pixel multi-model fusion recognition device provided in this embodiment of the invention can be used to execute the fire point pixel multi-model fusion recognition method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0125] Figure 10 This is a schematic diagram of the structure of a fire point pixel multi-model fusion recognition device provided in an embodiment of the present invention, with reference to... Figure 10 The fire point pixel multi-model fusion recognition device includes: a processor 31, a memory 32, a communication device 33, an input device 34, and an output device 35. The number of processors 31 and the number of memories 32 in the fire point pixel multi-model fusion recognition device can be one or more. The processor 31, memory 32, communication device 33, input device 34, and output device 35 of the fire point pixel multi-model fusion recognition device can be connected via a bus or other means.

[0126] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the fire point pixel multi-model fusion recognition method in any embodiment of the present invention (e.g., the data preprocessing module 21, model input module 22, and fire point recognition module 23 in the fire point pixel multi-model fusion recognition device). The memory 32 may mainly include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created according to the use of the device, etc. Furthermore, the memory 32 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0127] The communication device 33 is used for data transmission.

[0128] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory 32, thereby realizing the above-mentioned fire point pixel multi-model fusion recognition method.

[0129] Input device 34 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 35 may include display devices such as a display screen.

[0130] The fire point pixel multi-model fusion recognition device provided above can be used to execute the fire point pixel multi-model fusion recognition method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0131] This invention also provides a storage medium containing computer-executable instructions. When executed by a computer processor, the computer-executable instructions are used to perform a fire pixel multi-model fusion recognition method. This fire pixel multi-model fusion recognition method includes: acquiring remote sensing data to be identified; performing compensation preprocessing on the remote sensing data to obtain compensated remote sensing data; performing terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data; inputting the initial remote sensing data into a trained base model to obtain corresponding base feature vectors, the base model including a spatial model, a temporal model, and a distinguishing model; determining the base vector weights corresponding to the base feature vectors; calculating a fusion feature vector based on the base feature vectors and their corresponding base vector weights; and outputting a fire pixel recognition result based on the fusion feature vector.

[0132] Storage medium—any type of memory device or storage device. The term "storage medium" is intended to include: mounting media, such as CD-ROM, floppy disk, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, etc.; non-volatile memory, such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements, etc. Storage medium may also include other types of memory or combinations thereof. Furthermore, storage medium may reside in a first computer system in which a program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term "storage medium" may include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0133] Of course, the computer-executable instructions provided in the embodiments of the present invention are not limited to the fire point pixel multi-model fusion recognition method described above, but can also execute related operations in the fire point pixel multi-model fusion recognition method provided in any embodiment of the present invention.

[0134] The fire point pixel multi-model fusion recognition device, storage medium, and fire point pixel multi-model fusion recognition equipment provided in the above embodiments can execute the fire point pixel multi-model fusion recognition method provided in any embodiment of the present invention. For technical details not described in detail in the above embodiments, please refer to the fire point pixel multi-model fusion recognition method provided in any embodiment of the present invention.

[0135] The above description is merely a preferred embodiment of the present invention and the technical principles employed. The present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the claims.

Claims

1. A method for multi-model fusion recognition of fire point pixels, characterized in that, include: Acquire remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data; The initial remote sensing data is input into the trained base model to obtain the corresponding base feature vector. The base model includes a spatial model, a temporal model, and a discrimination model. Determine the base vector weights corresponding to the base feature vectors, calculate the fused feature vectors based on the base feature vectors and their corresponding base vector weights, and output the fire pixel recognition results based on the fused feature vectors.

2. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, The step of performing compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data includes: Obtain the total number of pixels in the remote sensing data to be identified, and detect the number of cloud pixels in the remote sensing data to be identified; The cloud coverage rate is calculated based on the total number of pixels and the number of cloud pixels. If the cloud coverage rate is greater than the preset coverage threshold, microwave remote sensing data corresponding to the remote sensing data to be identified is obtained. Compensated remote sensing data is obtained based on the remote sensing data to be identified and the microwave remote sensing data.

3. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, The step of performing terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data includes: Locate the region to be identified corresponding to the compensated remote sensing data, and obtain the regional elevation information corresponding to the region to be identified; Based on the regional elevation information, determine the regional slope and regional aspect of the region to be identified; The spectral characteristics in the compensated remote sensing data are adjusted according to the slope and aspect of the region to obtain the initial remote sensing data.

4. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, Determining the basic vector weights corresponding to the basic feature vectors includes: Obtain the recognition confidence level corresponding to the basic feature vector and the standard remote sensing data corresponding to the basic model; Calculate the data similarity between the remote sensing data to be identified and the standard remote sensing data; The basic vector weights corresponding to the basic feature vectors are calculated based on the recognition confidence and the data similarity.

5. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, The training method for the base model includes: Obtain remote sensing data samples corresponding to the basic model, and extract the fire point region and fire point edge region from the remote sensing data samples; Fire point augmentation samples are generated based on the fire point region, and edge augmentation samples are generated based on the fire point edge region. The remote sensing data samples, the fire point augmentation samples, and the edge augmentation samples are merged to obtain model training samples, and the base model is trained using the model training samples.

6. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, The training method for the base model includes: Remote sensing data samples are input into a base model to obtain sample recognition results, and the recognition accuracy of the base model is determined based on the sample recognition results; If the recognition accuracy is less than a preset accuracy threshold, reinforcement learning is used to reinforce the basic model.

7. The fire point pixel multi-model fusion recognition method according to claim 1, characterized in that, The step of outputting the fire pixel recognition result based on the fused feature vector includes: The probability of fire point identification for each pixel in the remote sensing data to be identified is calculated by the Softmax function, which is set with a temperature coefficient. Based on the fire point identification probability, the fire point pixel identification result corresponding to each pixel in the remote sensing data to be identified is output.

8. A fire point pixel multi-model fusion recognition device, characterized in that, include: The data preprocessing module is used to acquire remote sensing data to be identified, perform compensation preprocessing on the remote sensing data to be identified to obtain compensated remote sensing data, and perform terrain preprocessing on the compensated remote sensing data to obtain initial remote sensing data. The model input module is used to input the initial remote sensing data into the trained basic model to obtain the corresponding basic feature vector. The basic model includes a spatial model, a temporal model, and a discrimination model. The fire point recognition module is used to determine the basic vector weights corresponding to the basic feature vectors, calculate the fused feature vectors based on the basic feature vectors and their corresponding basic vector weights, and output the fire point pixel recognition results based on the fused feature vectors.

9. A fire point pixel multi-model fusion recognition device, characterized in that, include: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the fire point pixel multi-model fusion recognition method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the fire point pixel multi-model fusion recognition method as described in any one of claims 1-7.

Images