A fire zone scale adaptive segmentation method and system based on multi-scale spatial feature fusion based on time series

Abstract

The present invention relates to the field of fire zone image segmentation, and specifically to a method and system for fire zone scale adaptive segmentation based on time series multi-scale spatial feature fusion. The method specifically includes a fire zone data acquisition step, a multi-scale feature fusion step, a time series extraction step and a fire zone segmentation step. Through multi-dimensional technological innovation, the method systematically solves the core problems of edge fuzzy, scale sensitivity and insufficient time series correlation in fire zone segmentation, providing high-precision and high-reliability technical support for fire monitoring and emergency response.

Description

Technical Field

[0001] The present invention relates to the field of fire zone image segmentation, and in particular to a fire zone scale adaptive segmentation method and system based on time series multi-scale spatial feature fusion. Background Art

[0002] Fire zone segmentation is a technology that monitors and analyzes target areas, accurately segmenting and predicting fire points. By monitoring fire activity in real time, this technology helps relevant departments promptly identify fire hazards and implement emergency measures. Currently, fire zone segmentation has been widely applied in a variety of fields, including forest fire prevention and control, industrial fire supervision, urban residential fire monitoring, and field fire monitoring, providing timely fire zone location and fire analysis. Therefore, fire zone segmentation technology has important research significance and can provide strong support for fire early warning, post-disaster rescue, and the optimization of fire prevention and control strategies.

[0003] The task of fire area segmentation can usually be solved by analyzing radar reflectivity maps, because the heat and gases generated by the fire will significantly change the temperature and density of the surrounding air, forming strong convection, which in turn forms a specific signal pattern on the radar reflectivity map. Currently, researchers usually use optical flow and deep learning methods to process radar reflectivity maps. Traditional optical flow methods can estimate dynamic changes in radar images, but often face the problems of low resolution and difficulty in accurately processing small targets. In contrast, deep learning methods, with their powerful feature extraction and learning capabilities, can effectively process multi-scale radar echo targets, significantly improving the accuracy of recognition and segmentation, and have gradually become the focus of this research field.

[0004] However, existing fire zone segmentation methods based on radar reflectivity data still have some problems. First, the radar reflectivity map of the fire point area usually has significant edge complexity. Fire points are often concentrated in a specific area, resulting in the reflection intensity of this area being much higher than that of the surrounding area. The edge area of ​​the fire point has a weaker reflection intensity, showing an irregular edge complexity. This leads to poor segmentation performance of the target edge area. Second, the fire point area shows significant scale differences in the radar reflectivity map. From small-scale initial fire points to large-scale burning areas, the fire area scale varies greatly in different situations, resulting in insufficient adaptability to the target scale. Finally, the radar reflectivity map of the fire point usually shows obvious temporal changes. The spread speed, intensity changes, and airflow disturbances of the fire all have significant temporal correlation, but existing methods do not fully consider the temporal correlation between data.

[0005] Therefore, in order to solve the above problems, this application proposes a fire zone scale adaptive segmentation method and system based on time series multi-scale spatial feature fusion. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a fire zone scale adaptive segmentation method and system based on time series multi-scale spatial feature fusion.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion, including:

[0009] The fire area data acquisition step is to obtain the radar reflectivity map of the target area and perform feature extraction to obtain the image to be processed;

[0010] a multi-scale feature fusion step, dividing the image to be processed into a high-resolution feature map and a low-resolution feature map according to a resolution parameter, respectively introducing the high-resolution feature map and the low-resolution feature map into a high-resolution pathway and a low-resolution pathway for feature enhancement to obtain high-resolution output features and low-resolution output features; and combining the high-resolution output features and the low-resolution output features with the detail information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing;

[0011] In a time series extraction step, the output feature map is subjected to time feature fusion by a time feature extraction module to obtain a final output feature with deep feature fusion in the time dimension;

[0012] The fire area segmentation step performs fire area segmentation through a multi-level attention module according to the final output features.

[0013] As a further improvement of the present invention, the feature enhancement of the high-resolution path includes subjecting the high-resolution feature map to initial convolution to extract local features and simplify the feature dimension, extracting key features from the extracted feature map through a Bottleneck module consisting of two convolutional layers, and then further optimizing the features through a Bottleneck module consisting of two convolutional layers to obtain detailed features and contextual information of the high-resolution feature map.

[0014] As a further improvement of the present invention, the feature enhancement of the high-resolution path also includes inputting the feature map after the initial convolution of the high-resolution feature map into the high-resolution independent convolution branch for secondary convolution, and fusing the feature map after the secondary convolution with the feature map after two layers of Bottleneck processing, and adjusting the spatial resolution through a downsampling operation to obtain a high-resolution output feature.

[0015] As a further improvement of the present invention, the feature enhancement of the low-resolution path includes subjecting the low-resolution feature map to initial convolution to extract local features and simplify the feature dimension, extracting key features from the extracted feature map through a Bottleneck module composed of two convolutional layers, and then further optimizing the features through a Bottleneck module composed of two convolutional layers to obtain detailed features and contextual information of the low-resolution feature map.

[0016] As a further improvement of the present invention, the feature enhancement of the low-resolution path also includes inputting the feature map after the initial convolution of the low-resolution feature map into a low-resolution independent convolution branch for secondary convolution, and fusing the feature map after the secondary convolution with the feature map after two layers of Bottleneck processing to obtain a low-resolution output feature.

[0017] As a further improvement of the present invention, the feature splicing includes splicing the high-resolution features and the low-resolution features along the channel dimension to form a 4C-dimensional feature map, compressing the channel to 2C through a 1×1 convolutional layer and retaining the core features of the feature map, and performing a nonlinear transformation on the compressed feature map to obtain an output feature map.

[0018] As a further improvement of the present invention, the time feature extraction module includes: t Extract the branch, expand the time step of the output feature map to 3T to form 3T×C×H×W features, perform three 1×1 convolutions in sequence, perform ReLUctant activation function and batch normalization after each convolution, and then perform element-by-element addition to obtain the intermediate feature F t1 .

[0019] As a further improvement of the present invention, the time feature extraction module further includes: c Extract the branch, concatenate the three time-step features of the output feature map along the channel dimension to form T×3C×H×W features and compress them to 2C along the channel dimension, and divide them into two groups of sub-features F along the channel dimension c1 and F c2 Get two groups of sub-feature maps with the same shape as the output feature map, and convert the intermediate feature F t1 With the first group of sub-features F c1 Perform element-level addition to complete the initial feature fusion, and then process the fused features through the convolution layer and ReLU activation function to obtain the feature F including deep temporal information. t2 , and at the same time the second group of sub-features F c2 After another convolutional layer for feature enhancement, the t2 Element-level addition is performed to achieve further feature fusion, and the final fused time features are processed through a convolution layer and ReLU activation function to obtain the final output features.

[0020] A fire zone scale adaptive segmentation system based on time series multi-scale spatial feature fusion, including:

[0021] The fire area data acquisition module obtains the radar reflectivity map of the target area and performs feature extraction to obtain the image to be processed;

[0022] A multi-scale feature fusion module divides the image to be processed into a high-resolution feature map and a low-resolution feature map according to a resolution parameter, and imports the high-resolution feature map and the low-resolution feature map into a high-resolution pathway and a low-resolution pathway respectively for feature enhancement to obtain high-resolution output features and low-resolution output features; the high-resolution output features and the low-resolution output features are combined with the detailed information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing;

[0023] A time series extraction module performs time feature fusion on the output feature map through a time feature extraction module to obtain a final output feature with deep feature fusion in the time dimension;

[0024] The fire area segmentation module performs fire area segmentation through a multi-level attention module according to the final output features.

[0025] The beneficial effects of the present invention are: through the multi-scale feature fusion module, the detailed information of high-resolution features and the global semantics of low-resolution features are effectively integrated, which significantly improves the adaptability of the model to fire zone targets of different scales; the temporal feature extraction module innovatively integrates the dynamic change information in the time dimension, captures the continuity characteristics of fire spread, and enhances the model's ability to utilize temporal correlation data; the multi-level attention mechanism accurately focuses on the edge area of ​​the target, optimizes the interactive learning of local features and overall distribution, and significantly improves the segmentation accuracy of traditional methods in complex edge scenes. In addition, the design of independent branches and residual connections reduces the computational complexity while ensuring the feature expression ability, thereby improving the real-time and robustness of the model. Through multi-dimensional technological innovation, the present invention systematically solves the core problems of edge fuzziness, scale sensitivity and insufficient temporal correlation in fire zone segmentation, providing high-precision and high-reliability technical support for fire monitoring and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion in this application;

[0027] Figure 2 This is a block diagram of the spatial feature fusion module in the system of this application;

[0028] Figure 3 This is the block diagram of the time feature extraction module of this application. DETAILED DESCRIPTION

[0029] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom," "top," "inner," and "outer" refer to directions toward or away from the geometric center of a particular component, respectively.

[0030] Existing fire zone segmentation methods based on radar reflectivity data suffer from poor performance in segmenting target edges, insufficient adaptability to target scales, and a failure to consider temporal correlations between data. To address these issues, a fire zone scale-adaptive segmentation method based on time series multi-scale spatial feature fusion is proposed. This method gradually addresses the limitations of existing techniques through the steps of fire zone data acquisition, multi-scale feature fusion, time series extraction, and fire zone segmentation.

[0031] The problems with fire zone segmentation primarily involve three aspects: poor segmentation performance at the target's edge, insufficient adaptability to the target scale, and failure to consider temporal correlations between data. First, the poor segmentation performance at the target's edge is due to the significant edge complexity of the radar reflectivity map of the fire's area. The reflection intensity at the edge of the fire is weak, resulting in an irregular, complex edge. Second, the fire's area exhibits significant scale differences in the radar reflectivity map. From small, initial fires to large burning areas, the scale of the fire varies significantly across different scenarios. Finally, the radar reflectivity map of the fire typically exhibits significant temporal variations, with significant temporal correlations between the fire's spread rate, intensity variations, and airflow disturbances.

[0032] Radar reflectivity maps are obtained by transmitting electromagnetic waves and receiving reflected signals. The reflectivity factor reflects the intensity of the radar echo and is primarily used to characterize the meteorological characteristics of the fire area. The reflectivity factor of the fire area reflects the impact of the thermal air and smoke generated by the fire on the radar signal. The radar scan covers 360 azimuth points and 800 range points, with a detection range of 200 kilometers and a tangential resolution of 0.25 kilometers. The reflectivity map is converted from polar coordinates to Cartesian coordinates and converted to a high-resolution 1600×1600 pixel data using bilinear interpolation.

[0033] The fire area data acquisition step includes obtaining a radar reflectivity map of the target area and performing feature extraction to obtain an image to be processed. In the multi-scale feature fusion step, the image to be processed is divided into a high-resolution feature map and a low-resolution feature map according to the resolution parameter, and is respectively introduced into the high-resolution pathway and the low-resolution pathway for feature enhancement to obtain high-resolution output features and low-resolution output features. The high-resolution output features and the low-resolution output features are combined with the detailed information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing. In the temporal extraction step, the output feature map is subjected to temporal feature fusion through a temporal feature extraction module to obtain a final output feature with deep feature fusion in the temporal dimension. In the fire area segmentation step, the fire area is segmented according to the final output features through a multi-level attention module.

[0034] Compared with the existing technology, this method has advantages in many aspects. Through the multi-scale feature fusion module, it can effectively capture the different spatial hierarchical features of the fire area and improve the adaptability of the model to targets of different scales. Through the time feature extraction module, it can capture more temporal correlation features and effectively utilize the temporal correlation information between data. Through the multi-level attention module, key local area features and overall distribution features are extracted at the same time to improve the segmentation performance of the target edge area. Through the above steps, this method can effectively solve the problems of poor target edge area segmentation performance, insufficient target scale adaptability and failure to consider the temporal correlation between data in the existing technology, and significantly improve the accuracy and robustness of fire area segmentation.

[0035] like Figures 1 to 3 As shown, including:

[0036] The fire area data acquisition step obtains the radar reflectivity map of the target area and performs feature extraction to obtain the image to be processed.

[0037] In the multi-scale feature fusion step, the image to be processed is divided into a high-resolution feature map and a low-resolution feature map based on the resolution parameter. The high-resolution feature map and the low-resolution feature map are respectively fed into the high-resolution and low-resolution pathways for feature enhancement to obtain high-resolution and low-resolution output features. The high-resolution and low-resolution output features are combined to form the output feature map by combining the detailed information of the high-resolution feature map and the global semantics of the low-resolution feature map.

[0038] In the time series extraction step, the output feature map is subjected to time feature fusion through the time feature extraction module to obtain the final output feature with deep feature fusion in the time dimension.

[0039] In the fire area segmentation step, the fire area is segmented through a multi-level attention module based on the final output features.

[0040] Furthermore, if Figures 1 to 3 As shown in the figure, the high-resolution feature map is subjected to initial convolution to extract local features and simplify the feature dimension. The extracted feature map is passed through a Bottleneck module composed of two convolutional layers to extract key features, and then further optimized and extracted through a Bottleneck module composed of two convolutional layers to obtain the detailed features and context information of the high-resolution feature map.

[0041] The initial convolution is used to extract local features and simplify feature dimensions, while the bottleneck module is used to extract key features and optimize them. The initial convolution reduces feature dimensions, lowering computational complexity while simultaneously extracting important local feature information. The first bottleneck module extracts key features through two convolutional layers and further optimizes the feature representation. The second bottleneck module further optimizes the features, extracting detailed features and contextual information. These steps enable the effective extraction of key features and optimization of feature dimensions during feature enhancement of high-resolution feature maps.

[0042] The initial convolution can be implemented by using a 3×3 or 5×5 convolution kernel to extract local features and simplify the feature dimension. The Bottleneck module can be implemented by combining two layers of 1×1 convolution and a middle layer of 3×3 convolution. The 1×1 convolution is used for dimensionality reduction and dimensionality increase, and the 3×3 convolution is used for feature extraction and optimization. Figures 1 to 3 As shown in Figure 2, batch normalization and ReLU activation functions can be added to each Bottleneck module to improve the training stability and nonlinear expression ability of the model.

[0043] This application uses a combination of initial convolution and two Bottleneck modules in the feature enhancement process of high-resolution feature maps to achieve effective extraction of key features and optimization of feature dimensions. Compared with the existing technology, this application can significantly reduce the computational complexity while maintaining the detailed information of the high-resolution feature map, and improve the accuracy and efficiency of feature extraction. As a result, this application provides an efficient and accurate method for the feature enhancement process of high-resolution feature maps.

[0044] Furthermore, if Figures 1 to 3 As shown in the figure, the feature map after the initial convolution of the high-resolution feature map is input into the high-resolution independent convolution branch for secondary convolution, and the feature map after the secondary convolution is fused with the feature map after the two-layer Bottleneck processing. The spatial resolution is adjusted through the downsampling operation to obtain the high-resolution output feature.

[0045] After the initial convolution of the high-resolution feature map, the feature map undergoes a second convolution with the high-resolution independent convolution branch. The resulting feature map is then fused with the feature map processed by two Bottleneck modules. The Bottleneck module, consisting of two convolutional layers, extracts and optimizes key features. The fused feature map is then downsampled to adjust the spatial resolution, ultimately producing high-resolution output features. This process, through the combination of high-resolution independent convolution and the Bottleneck module, enables feature map optimization and fusion at different levels, thereby improving feature representation and resolution.

[0046] The secondary convolutions in the high-resolution independent convolution branches can use different kernel sizes and strides to accommodate different feature extraction requirements. The two convolutional layers in the Bottleneck module can further optimize feature extraction by introducing different activation functions and normalization layers. Downsampling can be implemented through pooling or convolutional layers to ensure that the spatial resolution of the feature map is adjusted to an appropriate scale.

[0047] By feeding the high-resolution feature map after initial convolution into a high-resolution independent convolution branch for secondary convolution, fusing the secondary convolutional feature map with the feature map after two layers of bottleneck processing, and finally adjusting the spatial resolution through downsampling, the method can effectively enhance the detailed features and contextual information of the high-resolution feature map. Compared with existing technologies, the method of this application can better preserve and optimize high-resolution features, thereby improving the accuracy and robustness of fire area segmentation.

[0048] Furthermore, if Figures 1 to 3 As shown in the figure, the low-resolution feature map is subjected to initial convolution to extract local features and simplify the feature dimension. The extracted feature map is passed through a Bottleneck module composed of two convolutional layers to extract key features, and then further optimized and extracted through a Bottleneck module composed of two convolutional layers to obtain the detailed features and context information of the low-resolution feature map.

[0049] Feature enhancement in the low-resolution pathway involves the following steps: First, the low-resolution feature map undergoes an initial convolution to extract local features and simplify the feature dimension. Specifically, the initial convolution can be implemented using a standard convolutional layer, whose kernel size and stride can be adjusted according to actual needs to ensure that the feature map dimension is appropriately simplified while retaining the necessary local feature information. Next, the extracted feature map is passed through a two-convolutional layer-based bottleneck module to extract key features. The first convolutional layer of the bottleneck module primarily extracts key features and reduces computational complexity, while the second convolutional layer further optimizes the feature representation. In this way, key features from the low-resolution feature map can be effectively extracted. Finally, feature optimization is further performed through a two-convolutional layer-based bottleneck module to obtain detailed features and contextual information from the low-resolution feature map. The bottleneck module design can be combined with residual connections to preserve input feature information, prevent gradient vanishing, and accelerate network convergence.

[0050] When implementing low-resolution path feature enhancement, there are many possible implementation methods. For example, in the selection of the initial convolution, different convolution kernel sizes and step sizes can be selected according to the specific application scenario to adapt to different feature extraction requirements. In addition, the convolution layer in the Bottleneck module can also be adjusted according to the specific situation of the feature map, such as adjusting the convolution kernel size, step size and the number of convolution layers to optimize the feature extraction effect. Further, if Figures 1 to 3 As shown in Figure 2, operations such as batch normalization and activation functions (such as ReLU) can be introduced into the Bottleneck module to improve the stability and nonlinear expression ability of feature extraction.

[0051] Through the above technical solution, the present application can effectively solve the problem of feature extraction and optimization of low-resolution feature maps. Compared with the existing technology, the low-resolution path feature enhancement method of the present application can better extract the detailed features and contextual information in the low-resolution feature map, thereby improving the overall feature fusion and segmentation effect. In particular, by using the combined design of initial convolution and multiple Bottleneck modules, it is possible to fully tap the useful information in the low-resolution feature map while ensuring computational efficiency, thereby improving the accuracy and robustness of fire zone segmentation.

[0052] Furthermore, if Figures 1 to 3 As shown in FIG, the feature enhancement of the low-resolution pathway also includes inputting the feature map after the initial convolution of the low-resolution feature map into the low-resolution independent convolution branch for secondary convolution, and fusing the feature map after the secondary convolution with the feature map after the two-layer Bottleneck processing to obtain the low-resolution output feature.

[0053] Feature enhancement in the low-resolution pipeline involves several steps. First, after an initial convolution, the low-resolution feature map is fed into a low-resolution independent convolution branch for a second convolution. The second-convolved feature map is then fused with the feature map processed by two Bottleneck layers. These steps further extract and optimize the features of the low-resolution feature map, resulting in low-resolution output features. This approach allows for better extraction of detail and contextual information from the low-resolution feature map, enhancing feature fusion and addressing the issue of insufficient low-resolution feature extraction.

[0054] After the initial convolution, the low-resolution feature map is fed into a low-resolution independent convolution branch for secondary convolution. Independent convolution can employ standard convolution operations, extracting features through a combination of multiple convolutional layers and activation functions. The feature map after secondary convolution is fused with the feature map processed by two Bottleneck layers. A Bottleneck module typically consists of two convolutional layers, the first of which reduces feature dimensionality, and the second restores it. This design allows for the extraction of more feature information while maintaining computational efficiency. The fused feature map is further processed to produce low-resolution output features.

[0055] By using this method, the present invention can further enhance the feature extraction and fusion effects in the low-resolution path. Compared with the existing technology, it can better extract the details and context information of the low-resolution feature map, solve the problem of insufficient low-resolution feature extraction, and improve the overall segmentation performance.

[0056] Furthermore, if Figures 1 to 3 As shown in the figure, the high-resolution features and low-resolution features are concatenated along the channel dimension to form a 4C-dimensional feature map. The channel is compressed to 2C through a 1×1 convolutional layer and the core features of the feature map are retained. The compressed feature map is nonlinearly transformed to obtain the output feature map.

[0057] Feature splicing involves concatenating high-resolution and low-resolution features along the channel dimension to form a 4C-dimensional feature map. A 1×1 convolutional layer compresses the channel to 2C while preserving the core features of the feature map. A nonlinear transformation is then performed on the compressed feature map to produce the final output feature map. This technique preserves the core information of the feature map by concatenating high-resolution and low-resolution features and compressing the channel using a 1×1 convolutional layer. The feature map is then further optimized through nonlinear transformations. This process effectively addresses the information integration issue in multi-scale feature fusion, improving the efficiency and accuracy of feature fusion.

[0058] Furthermore, if Figures 1 to 3As shown in the figure, the specific implementation of feature splicing includes the following possible variations. First, different nonlinear transformation methods can be used, such as ReLU activation function, Sigmoid activation function, or other suitable nonlinear activation functions, to enhance the nonlinear expression ability of the feature map. In addition, a batch normalization layer can be added after the 1×1 convolution layer to stabilize the training process and accelerate convergence. As a preferred embodiment, a spatial attention mechanism can also be applied to the spliced ​​feature map to further improve the expression ability of the feature map and the focus on key areas.

[0059] like Figure 2 As shown, the multi-scale feature fusion module of this application includes:

[0060] In the high-resolution pathway, the input feature map first undergoes a preliminary convolution operation to extract local features and reduce the dimensionality of the input features. The feature map then passes through two stacked Bottleneck modules. Each Bottleneck module consists of two convolutional layers. The first convolutional layer primarily extracts key features and reduces computational complexity, while the second convolutional layer further refines the feature representation. The Bottleneck module also incorporates residual connections, which effectively preserve input feature information, prevent gradient vanishing, and accelerate network convergence. By stacking two Bottleneck layers, the module is able to extract deeper contextual information and local details. This allows it to fully extract detailed features and contextual information about the fire spot area, thereby more comprehensively characterizing its morphological characteristics. The high-frequency pathway also includes a parallel, independent branch to process the feature map after the preliminary convolution. After further convolution, this branch is fused with the feature map processed by the Bottleneck module. This step preserves the original information and enhances feature diversity. Finally, the fused feature map undergoes downsampling to adjust the spatial resolution to match that of the low-resolution pathway, resulting in the final output of the high-resolution pathway. .

[0061] ;

[0062] The meaning of the above formula is the output feature map of the high-resolution pathway, F Hout represents the output feature map of the high-resolution pathway, F H The input feature map of the high-resolution path, Conv() represents the convolution layer, Bot() represents the Bottleneck module, and Downsample() represents downsampling. The processing flow of the low-resolution path is similar to that of the high-resolution path, but no downsampling operation is required. After the input low-resolution feature map is processed by the convolution and Bottleneck modules, the final output of the low-resolution path can be obtained. This operation can retain more semantic information and global features, and plays a significant role in capturing the spatial distribution characteristics of the fire area.

[0063] ;

[0064] The above formula represents the output feature map of the low-resolution path, F Lout represents the output feature map of the low-resolution pathway, F l Represents the input features of the low-resolution pathway. Finally, the output features of the high-resolution pathway and the low-resolution pathway are concatenated through a splicing operation, combining the detailed information of the high-resolution features with the global semantics of the low-resolution features. The concatenated feature map is then subjected to a 1×1 convolution operation to compress the feature dimension, simplify the feature expression, and highlight important information, thereby obtaining the final output of the module. .

[0065] ;

[0066] F Mout Represents the feature map after feature splicing, Concat() represents feature splicing, Conv 1×1 () represents a 1×1 convolutional channel. This application utilizes feature concatenation technology to effectively fuse high-resolution and low-resolution features during multi-scale feature fusion. Compared to existing technologies, this solution preserves the core information of feature maps by concatenating high-resolution and low-resolution features and utilizing 1×1 convolutional layers for channel compression. It further optimizes the feature maps through nonlinear transformations. This improves the efficiency and accuracy of feature fusion and ensures the effectiveness of multi-scale feature fusion.

[0067] Furthermore, if Figures 1 to 3 As shown in Figure 1, the temporal feature extraction module includes an Ft extraction branch, which expands the time step of the output feature map to 3T to form a 3T×C×H×W feature, performs three 1×1 convolutions in sequence, and performs ReLU activation function and batch normalization after each convolution, and then performs element-by-element addition to obtain the intermediate feature Ft1.

[0068] The temporal feature extraction module includes an Ft extraction branch that adds temporal information by expanding the time step of the output feature map to 3T. It then performs three 1×1 convolutions, each followed by a ReLU activation function and batch normalization, to gradually extract and fuse temporal features. Finally, the intermediate feature Ft1 is generated through element-by-element addition. This technical feature addresses technical issues with temporal feature extraction by increasing feature extraction and fusion along the temporal dimension, enhancing the model's responsiveness to temporal dynamics and predictive accuracy.

[0069] Specifically, the Ft extraction branch of the temporal feature extraction module first triples the time step of the output feature map, from T to 3T, to form a 3T×C×H×W feature map. The purpose of this step is to enrich the feature expression by adding information in the temporal dimension. Next, the feature map undergoes three 1×1 convolution operations in sequence, each followed by a ReLU activation function and batch normalization. The 1×1 convolution is used to reduce computational complexity and the number of parameters while maintaining the spatial resolution of the feature map. The ReLU activation function introduces nonlinear characteristics, increasing the expressive power of the model. Batch normalization helps stabilize the training process and accelerate convergence. Finally, the feature maps after the three convolutions are element-wise added to obtain the intermediate feature Ft1. This step further enhances the feature expression in the temporal dimension by fusing features from different time steps.

[0070] Through the aforementioned technical means, the temporal feature extraction module of this application effectively addresses the technical challenges of temporal feature extraction. Compared to existing technologies, this module significantly enhances the model's responsiveness to dynamic temporal changes and predictive accuracy by adding temporal dimension information and gradually integrating temporal features. This allows for more accurate capture of changing information within time series, improving the accuracy and robustness of fire zone segmentation.

[0071] Furthermore, if Figures 1 to 3 As shown, the temporal feature extraction module also includes an Fc extraction branch, which concatenates the three time-step features of the output feature map along the channel dimension to form a T×3C×H×W feature and compresses it to 2C along the channel dimension, and divides it into two groups of sub-features Fc1 and Fc2 along the channel dimension to obtain two groups of sub-feature maps with the same shape as the output feature map, and adds the intermediate feature Ft1 to the first group of sub-features Fc1 at the element level to complete the initial feature fusion, and further processes the fused feature through the convolution layer and the ReLU activation function to obtain the feature Ft2 including deep temporal information, and at the same time, the second group of sub-features Fc2 are enhanced by another convolution layer and then added to Ft2 at the element level to achieve further feature fusion, and the final fused time feature is processed by a convolution layer and the ReLU activation function to obtain the final output feature.

[0072] The technical solution of the temporal feature extraction module includes two main branches: Ft extraction branch and Fc extraction branch. The Ft extraction branch is responsible for expanding the time step of the output feature map to 3T to form 3T×C×H×W features, and performing three 1×1 convolutions in sequence. After each convolution, the ReLU activation function and batch normalization processing are performed, and then element-by-element addition is performed to obtain the intermediate feature Ft1. Figures 1 to 3As shown in the figure, the Fc extraction branch concatenates the three time-step features of the output feature map along the channel dimension to form a T×3C×H×W feature, compresses it to 2C along the channel dimension, and then divides it into two groups of sub-features Fc1 and Fc2 along the channel dimension. The intermediate feature Ft1 is element-wise added to the first group of sub-features Fc1 to complete the initial feature fusion. The fused features are processed through a convolution layer and a ReLU activation function to obtain the feature Ft2 that includes deep temporal information. At the same time, the second group of sub-features Fc2 is enhanced through another convolution layer and then element-wise added to Ft2 to achieve further feature fusion. Finally, the fused temporal features are processed through a convolution layer and a ReLU activation function to obtain the final output features.

[0073] There are many variations for the implementation of the Fc extraction branch. For example, the number and parameters of the convolutional layers can be adjusted according to the specific application scenario to meet different feature extraction requirements. Figures 1 to 3 As shown in the figure, the ReLU activation function can be replaced with other suitable activation functions, such as Leaky ReLU or ELU, to improve the nonlinear expression ability during feature extraction. In addition, batch normalization can also be replaced with other normalization methods, such as layer normalization or instance normalization, as needed to improve the training stability and convergence speed of the model.

[0074] like Figure 3 As shown, the time feature extraction module of this application includes:

[0075] This application uses the temporal feature extraction module TFE in the decoder part. In this module, we process the features of three consecutive time steps at the same time. The input feature format is T×C×H×W (Time×Channel×Height×Width), where T represents the number of time steps, C represents the number of channels, H represents the spatial height, and W represents the spatial width. First, all the input features are processed in two ways. The first is the feature reshaping operation, which forms the feature F by stacking the features of multiple time steps in the time dimension. t , represents the time branch feature, which is in the form of (3×T)×C×H×W. This operation enhances the continuity of the time series. The second processing method is the splicing operation of the channel dimension to obtain the feature F c , represents the channel branch feature, which is in the form of T×(3×C)×H×W. This operation enhances the diversity of features in each time step.

[0076]

[0077]

[0078]

[0079] F timei ∈R T×C×H×W Represents the temporal feature F of the input timei The characteristic form is T×C×H×W, F t ∈R （3 ×T）×C×H×W The form of expressing the time branch feature is (3×T)×C×H×W, F c ∈R T×(3×C)×H×W The form of representing channel branch features is T×(3×C)×H×W.

[0080] Next, F t Through the convolution layer, ReLU activation function and batch normalization (BN) processing, the first intermediate feature F is obtained t1 .

[0081]

[0082] ReLUctant() represents the activation function, and BN() represents the normalization. This step strengthens the nonlinear representation ability of the feature and stabilizes the network learning process. At the same time, the feature map F c The channel dimension is first compressed to 2C, and the time dimension T is expanded to 3T, and then divided into two groups of sub-features along the channel dimension to obtain two groups of sub-features with the feature map F. t Sub-feature maps of the same shape, F c1 represents the first channel branch sub-feature, F c2 Represents the second channel branch sub-feature, Reshape() represents dimension transformation, and Split() represents channel grouping.

[0083]

[0084]

[0085] Afterwards, F c1 With F t1 Perform element-level addition to achieve preliminary feature fusion. This process preliminarily integrates the continuity features of the time series and the features within the time step.

[0086] The fused features are further processed through the convolution layer and ReLU activation function to obtain the second intermediate feature F containing deep temporal information. t2 .

[0087] At the same time, F c2 After another convolutional layer to further enhance the feature representation, it is combined with F t2Element-wise addition is performed to achieve further feature fusion, enhancing the depth and richness of the composite temporal feature expression. Finally, the fused temporal features are processed through a convolutional layer and a ReLU activation function to output the final output features of the module.

[0088] The output contains high-level features that have been deeply fused in the time dimension, providing refined temporal feature representation for subsequent fire point segmentation.

[0089] This application further enhances the ability to extract and fuse temporal features by adding an Fc extraction branch to the temporal feature extraction module. Compared to existing technologies, this application can more effectively capture deep-level feature information in the temporal dimension, thereby improving the accuracy and robustness of fire zone segmentation. This solves the problem of insufficient utilization of temporal correlation features in existing technologies and significantly improves the model's performance in fire zone segmentation tasks.

[0090] Furthermore, if Figures 1 to 3 As shown, a time-series-based multi-scale spatial feature fusion fire zone scale adaptive segmentation system includes: a fire zone data acquisition module, which obtains a radar reflectivity map of the target area and performs feature extraction to obtain a to-be-processed image; a multi-scale feature fusion module, which divides the to-be-processed image into a high-resolution feature map and a low-resolution feature map according to a resolution parameter, and imports the high-resolution feature map and the low-resolution feature map into a high-resolution path and a low-resolution path respectively for feature enhancement to obtain high-resolution output features and low-resolution output features; the high-resolution output features and the low-resolution output features are combined with the detail information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing; a time series extraction module, which performs time feature fusion on the output feature map through a time feature extraction module to obtain a final output feature with deep feature fusion in the time dimension; a fire zone segmentation module, which performs fire zone segmentation through a multi-level attention module according to the final output features.

[0091] The system includes a fire zone data acquisition module, a multi-scale feature fusion module, a time series extraction module, and a fire zone segmentation module. The fire zone data acquisition module is used to obtain the radar reflectivity map of the target area and perform feature extraction, solving the problems of data acquisition and preliminary processing. The multi-scale feature fusion module solves the problems of multi-scale feature extraction and fusion by dividing the image to be processed into a high-resolution feature map and a low-resolution feature map, and performing feature enhancement in the high-resolution path and the low-resolution path. The time series extraction module solves the problem of time series feature extraction by performing time feature fusion on the output feature map through the time feature extraction module. The fire zone segmentation module uses a multi-level attention module to perform fire zone segmentation on the final output features, solving the problem of fire zone segmentation accuracy.

[0092] Through a modular design, this technical solution addresses the issues of poor performance in target edge segmentation, insufficient adaptability to target scales, and a failure to consider temporal correlations between data during fire zone segmentation. The fire zone data acquisition module ensures data accuracy and integrity, the multi-scale feature fusion module improves adaptability to targets of varying scales, the temporal extraction module enhances the capture of temporal correlation features, and the fire zone segmentation module improves segmentation accuracy and detail processing capabilities through a multi-level attention mechanism.

[0093] The technical features of the fire zone data acquisition module include the use of radar reflectivity maps, which are extracted through a specific algorithm to ensure the high accuracy and completeness of the acquired data. The key to the multi-scale feature fusion module is to divide the image into high-resolution and low-resolution feature maps and perform feature enhancement in different pathways. The high-resolution pathway mainly extracts detailed features, while the low-resolution pathway captures global semantic information. The time series extraction module fuses the feature maps in the temporal dimension through the temporal feature extraction module to ensure that the dynamic changes in the characteristics of consecutive time steps are captured. The fire zone segmentation module segments the final features through a multi-level attention mechanism, improving the accuracy of segmentation and the ability to process details.

[0094] As a preferred implementation, high-resolution feature maps can be extracted and optimized using initial convolution and the Bottleneck module. Low-resolution feature maps are similarly processed. Feature concatenation compresses the channels through a 1×1 convolutional layer and performs nonlinear transformations to ensure that the core information of the feature map is preserved. The temporal feature extraction module includes an Ft extraction branch and an Fc extraction branch, which achieve deep temporal feature fusion through multiple convolutions and activation function processing.

[0095] This application solves several issues in existing fire zone segmentation technologies through the collaborative work of multiple modules. Compared with existing technologies, the system provided by this application has significant advantages in edge region segmentation performance, target scale adaptability, and temporal correlation feature capture, thereby improving the accuracy and reliability of fire zone segmentation.

[0096] The above shows and describes the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, which are only some embodiments. Without departing from the spirit and scope of the present invention, various improvements and supplements made are considered to be within the scope of protection of the present invention.

Claims

1. A fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion, characterized by: include: The fire area data acquisition step is to obtain the radar reflectivity map of the target area and perform feature extraction to obtain the image to be processed; a multi-scale feature fusion step, dividing the image to be processed into a high-resolution feature map and a low-resolution feature map according to a resolution parameter, respectively introducing the high-resolution feature map and the low-resolution feature map into a high-resolution pathway and a low-resolution pathway for feature enhancement to obtain high-resolution output features and low-resolution output features; and combining the high-resolution output features and the low-resolution output features with the detail information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing; In a time series extraction step, the output feature map is subjected to time feature fusion by a time feature extraction module to obtain a final output feature with deep feature fusion in the time dimension; a fire area segmentation step, performing fire area segmentation through a multi-level attention module according to the final output features; The time feature extraction module includes: t Extract the branch, expand the time step of the output feature map to 3T to form 3T×C×H×W features, perform three 1×1 convolutions in sequence, perform ReLUctant activation function and batch normalization after each convolution, and then perform element-by-element addition to obtain the intermediate feature F t1 ; The time feature extraction module also includes: c Extract the branch, concatenate the three time-step features of the output feature map along the channel dimension to form T×3C×H×W features and compress them to 2C along the channel dimension, and divide them into two groups of sub-features F along the channel dimension c1 and F c2 Get two groups of sub-feature maps with the same shape as the output feature map, and convert the intermediate feature F t1 With the first group of sub-features F c1 Perform element-level addition to complete the initial feature fusion, and then process the fused features through the convolution layer and ReLU activation function to obtain the feature F including deep temporal information. t2 , and at the same time the second group of sub-features F c2 After another convolutional layer for feature enhancement, the t2 Element-level addition is performed to achieve further feature fusion, and the final fused time features are processed through a convolution layer and ReLU activation function to obtain the final output features.

2. The fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion according to claim 1 is characterized in that: The feature enhancement of the high-resolution pathway includes subjecting the high-resolution feature map to initial convolution to extract local features and simplify feature dimensions, extracting key features from the extracted feature map through a Bottleneck module consisting of two convolutional layers, and then further optimizing the features through another Bottleneck module consisting of two convolutional layers to extract detailed features and contextual information from the high-resolution feature map.

3. The fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion according to claim 2 is characterized in that: The feature enhancement of the high-resolution path also includes inputting the feature map after the initial convolution of the high-resolution feature map into the high-resolution independent convolution branch for secondary convolution, and fusing the feature map after the secondary convolution with the feature map after two layers of Bottleneck processing, and adjusting the spatial resolution through a downsampling operation to obtain a high-resolution output feature.

4. The fire zone scale adaptive segmentation method based on time series multi-scale spatial feature fusion according to claim 1 is characterized in that: The feature enhancement of the low-resolution path includes subjecting the low-resolution feature map to initial convolution to extract local features and simplify the feature dimension, extracting key features from the extracted feature map through a Bottleneck module consisting of two convolutional layers, and then further optimizing the features through a Bottleneck module consisting of two convolutional layers to obtain detailed features and contextual information of the low-resolution feature map.

5. The method for fire zone scale adaptive segmentation based on time series multi-scale spatial feature fusion according to claim 4 is characterized in that: The feature enhancement of the low-resolution path also includes inputting the feature map after the initial convolution of the low-resolution feature map into the low-resolution independent convolution branch for secondary convolution, and fusing the feature map after the secondary convolution with the feature map after the two-layer Bottleneck processing to obtain a low-resolution output feature.

6. The method for fire zone scale adaptive segmentation based on time series multi-scale spatial feature fusion according to claim 1 is characterized in that: The acquisition of the output feature map includes splicing the high-resolution features and the low-resolution features along the channel dimension to form a 4C-dimensional feature map, compressing the channel to 2C through a 1×1 convolution layer and retaining the core features of the feature map, and performing a nonlinear transformation on the compressed 4C-dimensional feature map to obtain the output feature map.

7. A fire zone scale adaptive segmentation system based on time series multi-scale spatial feature fusion, characterized by: include: The fire area data acquisition module obtains the radar reflectivity map of the target area and performs feature extraction to obtain the image to be processed; A multi-scale feature fusion module divides the image to be processed into a high-resolution feature map and a low-resolution feature map according to a resolution parameter, and imports the high-resolution feature map and the low-resolution feature map into a high-resolution pathway and a low-resolution pathway respectively for feature enhancement to obtain high-resolution output features and low-resolution output features; the high-resolution output features and the low-resolution output features are combined with the detailed information of the high-resolution feature map and the global semantics of the low-resolution feature map to obtain an output feature map by splicing; A time series extraction module performs time feature fusion on the output feature map through a time feature extraction module to obtain a final output feature with deep feature fusion in the time dimension; a fire zone segmentation module, which performs fire zone segmentation according to the final output features through a multi-level attention module; The time feature extraction module includes: t Extract the branch, expand the time step of the output feature map to 3T to form 3T×C×H×W features, perform three 1×1 convolutions in sequence, perform ReLUctant activation function and batch normalization after each convolution, and then perform element-by-element addition to obtain the intermediate feature F t1 ; The time feature extraction module also includes: c Extract the branch, concatenate the three time-step features of the output feature map along the channel dimension to form T×3C×H×W features and compress them to 2C along the channel dimension, and divide them into two groups of sub-features F along the channel dimension c1 and F c2 Get two groups of sub-feature maps with the same shape as the output feature map, and convert the intermediate feature F t1 With the first group of sub-features F c1 Perform element-level addition to complete the initial feature fusion, and then process the fused features through the convolution layer and ReLU activation function to obtain the feature F including deep temporal information. t2 , and at the same time the second group of sub-features F c2 After another convolutional layer for feature enhancement, the t2 Element-level addition is performed to achieve further feature fusion, and the final fused time features are processed through a convolution layer and ReLU activation function to obtain the final output features.

Images