Rapid BP imaging accumulated snow detection method and system, medium and product

Abstract

The invention discloses a rapid BP imaging accumulated snow detection method and system, a medium and a product, and relates to the technical field of image processing. The method comprises the following steps: firstly, acquiring original echo data acquired by a radar system and carrying out range pulse compression processing; performing azimuth coherent accumulation on the data to obtain a source resolution image, and extracting position information of suspected accumulated snow from the source resolution image; determining a target area according to the position information and a preset boundary, and performing backward projection imaging on the area to obtain a high-resolution image; inputting the image into a pre-trained semantic segmentation network for pixel-level classification, and obtaining an accurate segmentation result of the snow area; and finally, analyzing spatial distribution and area information of accumulated snow based on a segmentation result, and generating a detection report. By implementing the technical scheme provided by the invention, the accuracy of accumulated snow detection can be improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a fast BP imaging snow detection method, system, medium, and product. Background Technology

[0002] With increasingly stringent safety requirements for airport operations, all-weather, high-precision snow monitoring of airport pavements has become a crucial element in ensuring the safe takeoff and landing of aircraft. Ground-based circular arc trajectory synthetic aperture radar (GB-ArcSAR) demonstrates significant advantages in airport pavement monitoring due to its unique scanning trajectory and flexible installation and deployment capabilities.

[0003] Currently, ground-based circular trajectory synthetic aperture radar systems primarily employ the back projection (BP) algorithm for imaging processing in airport pavement monitoring. This algorithm, by precisely focusing on each pixel, effectively overcomes the limitations of traditional range-Doppler algorithms in handling nonlinear trajectories, achieving high-quality two-dimensional imaging.

[0004] However, in practical applications, the existing GB-ArcSAR system uses a single-pass processing mode for snow cover detection on airport pavements, that is, it directly performs BP imaging and identification on the entire observation area. This global single-pass processing method is not only computationally intensive, but also prone to large errors in the snow cover area identification process due to the lack of targeted fine processing, thus reducing the accuracy of snow cover detection. Summary of the Invention

[0005] This application provides a rapid BP imaging method, system, medium, and product for snow cover detection, which can improve the accuracy of snow cover detection.

[0006] The first aspect of this application provides a rapid BP imaging method for snow detection, comprising: The raw echo data collected by the ground-based circular arc trajectory synthetic aperture radar system on the airport pavement is acquired, and the raw echo data is processed by range pulse compression to obtain the target echo data. The target echo data is coherently accumulated in the azimuth direction to obtain the source resolution azimuth image; The location information of suspected snow cover is extracted from the source resolution azimuth image, and the location information includes a range coordinate range and an azimuth coordinate range. Based on the location information and the preset boundary margin, the target area suspected of being covered by snow is determined on the imaging plane; Back-projection imaging is performed on the pixels within the target area to obtain a radar image with target resolution; The target resolution radar image is input into a pre-trained semantic segmentation network for pixel-level classification to obtain the segmentation result of the snow-covered area; Based on the segmentation results, the spatial distribution and area information of the snow-covered area are determined, and an airport runway snow cover detection report is generated.

[0007] By adopting the above technical solution, the original echo data is first processed by range pulse compression to obtain target echo data, and then coherently accumulated in the azimuth direction to obtain a source resolution azimuth image, which can initially identify the location information of suspected snow accumulation. Then, based on this location information and a preset boundary margin, the target area of ​​suspected snow accumulation is determined, and back projection imaging is performed only on the pixels within the target area, avoiding global BP imaging processing of the entire observation area and significantly reducing computational complexity. Next, the obtained target resolution radar image is input into a pre-trained semantic segmentation network for pixel-level classification. By utilizing the feature extraction and classification capabilities of the deep learning model, the boundaries of the snow accumulation area can be identified more accurately. Finally, based on the segmentation results, the spatial distribution information and area information of the snow accumulation area are determined, thereby generating an accurate airport runway snow accumulation detection report and improving the accuracy of snow accumulation detection.

[0008] Optionally, the target echo data is arranged according to the radar position to generate a data matrix; for each range cell in the data matrix, the echo complex value corresponding to the radar position is extracted to generate the azimuth echo sequence of the range cell; the phase difference between each echo complex value in the azimuth echo sequence is calculated, and phase alignment is performed according to the phase difference; the phase-aligned azimuth echo sequence is accumulated to obtain the azimuth cumulative response value of the range cell; each range cell is traversed, and the azimuth cumulative response value of each range cell is used as the pixel value of the corresponding range position to generate the source resolution azimuth image.

[0009] Optionally, the image pixel values ​​in the source resolution azimuth image are mapped to a preset numerical range to obtain a normalized image; the normalized image is input into a pre-trained convolutional neural network, and the pre-trained convolutional neural network performs forward propagation calculation on the normalized image to output a binary mask image of the suspected snow-covered area. The training samples of the convolutional neural network are the source resolution azimuth image and the corresponding snow-covered area annotation information; morphological opening operation is performed on the binary mask image to remove isolated noise pixels to obtain a denoised binary image; morphological closing operation is performed on the binary image to fill the empty areas in the suspected snow-covered area to obtain a mask of the suspected snow-covered area; the circumscribed rectangle of the suspected snow-covered area mask is extracted, and the range of the distance coordinates and the range of the azimuth coordinates of the circumscribed rectangle are determined as the position information.

[0010] Optionally, the minimum and maximum values ​​of the range coordinate range and the azimuth coordinate range are obtained from the location information; the minimum value of the range coordinate range is subtracted from the preset boundary margin to determine the range starting coordinate; the maximum value of the range coordinate range is added to the preset boundary margin to determine the range ending coordinate; the minimum value of the azimuth coordinate range is subtracted from the preset boundary margin to determine the azimuth starting coordinate; the maximum value of the azimuth coordinate range is added to the preset boundary margin to determine the azimuth ending coordinate; and the target area is determined on the imaging plane based on the range starting coordinate, the range ending coordinate, the azimuth starting coordinate, and the azimuth ending coordinate.

[0011] Optionally, based on the range start coordinates, range end coordinates, azimuth start coordinates, and azimuth end coordinates of the target area, a set of pixels requiring back-projection imaging is determined. For each pixel in the set, the slant range between the pixel and each radar position is calculated based on the pixel's spatial coordinates and the spatial coordinates of each radar position on the circular trajectory. Based on the slant range, the complex echo value corresponding to the radar position is extracted from the target echo data. A phase compensation factor is determined based on the slant range, and the phase compensation factor is multiplied by the corresponding complex echo value to obtain a compensated echo value. The compensated echo values ​​for each radar position are coherently accumulated to obtain the complex image value of the pixel. All pixels in the set are traversed, and the complex image value of each pixel is used as the pixel value at the corresponding position to generate the target resolution radar image.

[0012] Optionally, the amplitude of the complex image values ​​in the target resolution radar image is calculated to obtain an amplitude image, and the amplitude image is logarithmically transformed to obtain a logarithmic amplitude image; the logarithmic amplitude image is subjected to histogram equalization to obtain an enhanced image; the enhanced image is input into a pre-trained semantic segmentation network, the training samples of which are the target resolution radar image and the corresponding pixel-level snow cover annotation information; each pixel in the enhanced image is classified by the pre-trained semantic segmentation network to determine the probability value of each pixel as a snow cover category, pixels with a probability value greater than or equal to a preset probability threshold are marked as snow cover category pixels, and pixels with a probability value less than the preset probability threshold are marked as non-snow cover category pixels, to obtain pixel-level classification results; connected component analysis is performed on the pixel-level classification results to remove connected regions with an area smaller than a preset area threshold, to obtain the segmentation result of the snow cover region.

[0013] Optionally, the segmentation results are labeled with connected components to obtain multiple independent snow-covered regions; for each snow-covered region, the number of pixels contained in the snow-covered region is counted, and the number of pixels is converted into the actual area according to the resolution of the target resolution radar image; the centroid coordinates and boundary contour coordinates of each snow-covered region are extracted; the centroid coordinates and boundary contour coordinates are converted from the imaging plane coordinate system to the airport pavement coordinate system to obtain the target centroid coordinates and target boundary contour coordinates in the airport pavement coordinate system; an airport pavement snow cover detection report containing the actual area of ​​each snow-covered region, the target centroid coordinates, and the target boundary contour coordinates is generated.

[0014] In a second aspect, embodiments of this application provide a fast BP imaging snow detection system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the fast BP imaging snow detection system to perform the method described in the first aspect and any possible implementation thereof.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a fast BP imaging snow detection system, cause the fast BP imaging snow detection system to perform the method described in the first aspect and any possible implementation thereof.

[0016] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a fast BP imaging snow detection system, cause the fast BP imaging snow detection system to perform the method described in the first aspect and any possible implementation thereof.

[0017] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By adopting the above technical solution, the original echo data is first processed by range pulse compression to obtain target echo data, and then coherently accumulated in the azimuth direction to obtain a source resolution azimuth image, which can initially identify the location information of suspected snow accumulation. Then, based on this location information and a preset boundary margin, the target area of ​​suspected snow accumulation is determined, and back projection imaging is performed only on the pixels within the target area, avoiding global BP imaging processing of the entire observation area and significantly reducing computational complexity. Next, the obtained target resolution radar image is input into a pre-trained semantic segmentation network for pixel-level classification. By utilizing the feature extraction and classification capabilities of the deep learning model, the boundaries of the snow accumulation area can be identified more accurately. Finally, based on the segmentation results, the spatial distribution information and area information of the snow accumulation area are determined, thereby generating an accurate airport runway snow accumulation detection report and improving the accuracy of snow accumulation detection. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a rapid BP imaging snow detection method disclosed in an embodiment of this application; Figure 2 This is another schematic flowchart of a rapid BP imaging snow detection method disclosed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a system provided in an embodiment of this application.

[0019] Explanation of reference numerals in the attached figures: 301, Central Processing Unit; 302, Read-Only Memory; 303, Random Access Memory; 304, Bus; 305, Input / Output Interface; 306, Input Section; 307, Output Section; 308, Storage Section; 309, Communication Section; 310, Driver; 311, Removable Media. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0021] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0022] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0023] This application provides a rapid BP imaging method for snow cover detection, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a rapid BP imaging snow cover detection method provided in an embodiment of this application. The method is applied to a system, which refers to a hardware and software integrated platform capable of executing a rapid BP imaging snow cover detection program. The system can execute a rapid BP imaging snow cover detection program, and the method includes steps 101 to 105, as follows: Step 101: Acquire the raw echo data collected by the ground-based circular arc trajectory synthetic aperture radar system on the airport pavement, and perform range pulse compression processing on the raw echo data to obtain the target echo data.

[0024] A ground-based circular trajectory synthetic aperture radar (SAR) system is a radar system installed at a fixed location on the ground. Its antenna scans along a circular trajectory to achieve high-resolution imaging. Raw echo data consists of the reflected signals received after the radar emits electromagnetic waves that illuminate the target; these signals contain information about the target's spatial location and scattering characteristics. Airport pavement refers to the smooth surface of airport runways, taxiways, etc., used for aircraft takeoff, landing, and taxiing. Target echo data is radar signal data with a higher signal-to-noise ratio and range resolution obtained after pulse compression processing.

[0025] Specifically, firstly, after the ground-based circular trajectory synthetic aperture radar system is installed in a fixed location, its operating parameters are set, including the transmitted signal frequency, bandwidth, pulse repetition frequency, and sampling rate. The radar antenna scans along a preset circular trajectory, continuously transmitting a linear frequency modulated signal and receiving the target reflection signal during the scan. The received reflection signal undergoes analog-to-digital conversion and digital sampling to form the raw echo data matrix. Subsequently, the raw echo data undergoes range-direction pulse compression processing, specifically including: multiplying the raw echo data with the complex conjugate of the transmitted signal in the frequency domain to achieve matched filtering; and performing an inverse Fourier transform on the matched filtering result to obtain the range-direction compressed target echo data. Through this processing, the range-direction ambiguous target response in the raw echo data can be compressed into a narrow pulse, improving range resolution and enhancing the signal-to-noise ratio. For example, for a radar system with a bandwidth of 300MHz, the range resolution can reach the order of 0.5 meters after pulse compression processing. This processed target echo data will be used for subsequent imaging and target detection.

[0026] Step 102: Perform azimuth coherent accumulation on the target echo data to obtain the source resolution azimuth image; extract the location information of suspected snow cover from the source resolution azimuth image, including the range coordinate range and the azimuth coordinate range.

[0027] The source resolution azimuth image refers to a radar image with initial resolution obtained after coherent accumulation processing, where pixel values ​​reflect the target's scattering intensity. Position information refers to the spatial description of the target's location in the radar imaging coordinate system, including the range (distance direction from the radar to the target) and azimuth (scanning direction of the radar antenna) coordinate ranges.

[0028] Specifically, the first step is coherent accumulation in the azimuth direction: The target echo data is arranged according to the sampling positions of the radar on the circular trajectory, forming a two-dimensional data matrix; for the echo data within each range gate, complex numerical sequences corresponding to different radar positions are extracted; the phase difference between the echo signals of adjacent radar positions is calculated, and phase correction is performed based on the theoretical phase difference; the corrected echo signals are summed using complex numbers to obtain the accumulation result for that range gate; the above process is repeated for all range gates to generate a source-resolution azimuth image. For example, for a certain range gate, assuming there are echo data from N radar positions {S1, S2, ..., SN}, and the phase correction factors are {exp(jφ1), exp(jφ2), ..., exp(jφN)}, then the accumulation result for that range gate is Σ(Sᵢ·exp(jφᵢ)), where i ranges from 1 to N. Then, amplitude calculation and logarithmic transformation are performed on the source resolution azimuth image to enhance image contrast; an adaptive threshold is set using the CFAR detection algorithm to extract suspected target regions above the threshold; connected component analysis is performed on the detection results to remove noise regions with excessively small areas; the minimum bounding rectangle of the remaining target regions is calculated to obtain their coordinate range in the range and azimuth directions. For example, for a detected suspected snow-covered area, the coordinate range of its bounding rectangle can be represented as: range [R_min, R_max], azimuth [A_min, A_max], where R and A represent the range and azimuth coordinate values, respectively. This positional information will be used to determine the target region range for subsequent fine processing.

[0029] In one possible implementation, the target echo data is coherently accumulated in the azimuth direction to obtain a source resolution azimuth image, specifically including steps 1021-1023, as follows: Step 1021: Arrange the target echo data according to the radar position to generate a data matrix; for each range cell in the data matrix, extract the echo complex value of the corresponding radar position to generate the azimuth echo sequence of the range cell.

[0030] Radar position refers to the spatial coordinates of the sampling point on the circular trajectory of the radar antenna. The data matrix is ​​a two-dimensional complex array formed by arranging echo data according to range cells and radar positions. A range cell is the smallest resolvable unit in the range direction of the radar detection space, corresponding to echo data within a specific range. The echo complex value is the representation of the amplitude and phase information of the radar received signal in complex form. The azimuth echo sequence is a sequence of echo complex values ​​obtained from the same range cell at different radar positions.

[0031] Specifically, first, the data matrix structure is established, with rows corresponding to range cells and columns corresponding to radar positions. Let M be the number of range sampling points in the radar system and N be the number of sampling positions on the circular trajectory, then an M×N complex matrix is ​​generated. The target echo data is then rearranged, with the echo data collected from each radar position filled into the corresponding matrix columns according to range gate order. For example, for the echo data {S} collected from the k-th radar position (k=1, 2, ..., N), ... 1k S 2k , ..., SMk}, are filled into the k-th column of the data matrix. Then, for each range cell m (m=1, 2, ..., M), all elements {Sm1, Sm2, ..., SmN} of that row are extracted from the data matrix to form the azimuth echo sequence for that range cell. Each complex value contains amplitude A and phase φ information, represented as Smn=A·exp(jφ). This data organization facilitates subsequent phase alignment and coherent accumulation processing. In specific implementations, a two-dimensional array can be used to store the data matrix, and a one-dimensional array can be used to store the azimuth echo sequence for each range cell. For example, for a system with a range resolution of 0.5m and a total range of 300m, M is approximately 600; for a system with an angle sampling interval of 0.1° and a scanning range of 90°, N is approximately 900, ultimately forming a 600×900 complex data matrix.

[0032] Step 1022: Calculate the phase difference between the complex values ​​of each echo in the azimuth echo sequence, and perform phase alignment based on the phase difference.

[0033] An azimuth echo sequence is a sequence of complex echo values ​​obtained from the same range cell at different radar locations. The complex echo values ​​are complex signals containing both amplitude and phase information. Phase difference refers to the amount of phase change between echo signals obtained from adjacent radar locations. Phase alignment is a signal processing technique that compensates for phase differences to achieve coherent superposition of echo signals from different locations.

[0034] Specifically, firstly, for the azimuth echo sequence {S1, S2, ..., SN} of each range cell, the phase difference between adjacent complex values ​​of echoes is calculated. Let the complex value of the echo at the nth position be Sn = An·exp(jφn), then the phase difference Δφn between the nth and (n+1)th positions is Δφn = φn+1 - φn = arg(Sn+1·Sn*), where Sn* represents the conjugate of Sn, and arg() represents the argument operation. After calculating the phase difference sequence {Δφ1, Δφ2, ..., ΔφN-1}, it is compared with the theoretical phase difference. The theoretical phase difference δφn can be calculated based on the radar wavelength λ and the change in distance from the target to the radar ΔR: δφn = 4π·ΔR / λ. The phase value to be compensated at each position is calculated: θn = Σ(δφi - Δφi), where i ranges from 1 to n-1. Finally, phase compensation is performed on the echo sequence: S'n = Sn·exp(jθn). For example, for a radar system operating at 10 GHz with a wavelength λ = 0.03 m, if a range cell R = 100 m and the distance between adjacent radar positions is 0.1 m, then the theoretical phase difference is approximately 12°. This phase alignment process ensures that echo signals from different directions can be coherently superimposed, improving the signal-to-noise ratio. In practice, a complex number arithmetic library can be used to perform phase difference calculations and phase compensation operations. For an echo sequence with N sampling points, the computational complexity is O(N).

[0035] Step 1023: Accumulate the phase-aligned azimuth echo sequence to obtain the azimuth cumulative response value of the range cell; traverse each range cell and use the azimuth cumulative response value of each range cell as the pixel value of the corresponding range position to generate the source resolution azimuth image.

[0036] The phase-aligned azimuth echo sequence is a complex numerical sequence of echoes from the same range cell after phase correction. The azimuth cumulative response value is a complex value obtained by coherently accumulating the phase-aligned echo sequence, reflecting the scattering intensity of the target in that range cell. A range cell is the smallest resolvable unit in the range direction in the radar detection space. The source resolution azimuth image is a two-dimensional image formed by arranging the azimuth cumulative response values ​​of each range cell in range order. The pixel value is the numerical value of each point in the image, representing the scattering intensity of the target at the corresponding location.

[0037] Specifically, the first step is azimuth-oriented accumulation: For each range cell m, the phase-aligned echo sequence {S'1, S'2, ..., S'N} is subjected to complex addition to obtain the cumulative response value Am = ΣS'n (n = 1 to N). Since phase alignment is complete, the echo signals of coherent targets will constructively superimpose, while noise will be suppressed due to phase randomness. For example, for a range cell containing a target, if the amplitude of the echo signal at each position is a, and the phase is aligned, the amplitude of the accumulated response value after N positions is theoretically N·a. Then, an image matrix of size M×1 is constructed, where M is the number of range cells. According to the range cell number m (m = 1, 2, ..., M), the cumulative response value Am of each range cell is filled into the corresponding position. The amplitude |Am| of the response value is calculated and converted to a decibel value 20·log10(|Am|) as the pixel value. Finally, the pixel value is normalized and mapped to a grayscale range of 0-255. For example, for a system with M=600 sampling points, a 600×1 azimuth image is generated. In practice, array manipulation functions can be used to perform accumulation and image generation, with a computational complexity of O(M·N). Although the generated source-resolution azimuth image has a low azimuth resolution, the computational cost is low, making it suitable for fast target detection.

[0038] In one possible implementation, the location information of suspected snow cover is extracted from the source resolution azimuth image, specifically including steps 1024-1027, as follows: Step 1024: Map the image pixel values ​​in the source resolution orientation image to a preset numerical range to obtain a normalized image.

[0039] The source resolution azimuth image is the initial radar image obtained through azimuth coherent accumulation processing. The image pixel value is the numerical value of each point in the image, representing the scattering intensity of the target at the corresponding location. The preset numerical range is a pre-determined standardized numerical range, typically chosen as [0, 1] or [0, 255]. The normalized image is the image obtained by mapping the pixel values ​​of the original image to the standardized numerical range, facilitating subsequent processing and display.

[0040] Specifically, first, all pixel values ​​in the source resolution azimuth image are obtained, and the maximum and minimum pixel values ​​Vmax and Vmin are determined. The target numerical range is set to [a, b]. For each pixel value v in the source image, a linear mapping formula is used for normalization: v' = a + (v - Vmin) × (ba) / (Vmax - Vmin), where v' is the normalized pixel value. This mapping operation is performed on all pixels in the image to generate a normalized image. For example, if the source image pixel value range is [-50dB, -10dB] and the target range is [0, 255], then for the original pixel value -30dB, the normalized result is 127.5. In practice, vectorized operations can be used to process all pixels simultaneously, improving computational efficiency. To handle the influence of outliers, extreme values ​​can be removed when determining Vmax and Vmin, such as using the 1% and 99th percentiles instead of the maximum and minimum values. Normalization gives the image a uniform numerical range, enhances image contrast, and facilitates subsequent object detection and feature extraction. In practical applications, [0, 255] is usually chosen as the target interval for easy image display, or [0, 1] is chosen to meet the input requirements of deep learning models.

[0041] Step 1025: Input the normalized image into the pre-trained convolutional neural network. Perform forward propagation calculation on the normalized image through the pre-trained convolutional neural network and output a binary mask image of the suspected snow area. The training samples of the convolutional neural network are the source resolution azimuth image and the corresponding snow area annotation information.

[0042] A pre-trained convolutional neural network (CNN) is a deep learning model trained for image recognition, containing multiple convolutional layers, pooling layers, and fully connected layers. Forward propagation is the process by which the neural network sequentially passes the input data through each network layer for feature extraction and transformation. A binary mask image is an image containing only 0 and 1 pixel values, used to mark the location of target regions. Training samples are data pairs used for model training, including the input radar image and corresponding manually labeled snow-covered area information.

[0043] Specifically, the first step is to construct a convolutional neural network model structure, comprising two main parts: feature extraction and semantic segmentation. The feature extraction part employs a multi-layer convolutional structure, such as convolutional layers with 3×3 kernels, 2×2 max-pooling layers, and the ReLU activation function, extracting image features layer by layer. Example network structure: Input layer (1×M×1) → Convolutional layer 1 (16 channels) → Pooling layer 1 → Convolutional layer 2 (32 channels) → Pooling layer 2 → Convolutional layer 3 (64 channels) → Deconvolutional layer → Output layer (1×M×1). For an input image with M=600 distance units, after downsampling by the pooling layer and upsampling by the deconvolutional layer, a binary mask image of the same size is output.

[0044] The network was trained using supervised learning. Training data included source-resolution azimuth images and corresponding snow-covered region annotations. The annotations were binary images, with snow-covered regions marked as 1 and non-snow-covered regions marked as 0. The training process used a cross-entropy loss function, and the network parameters were optimized through backpropagation. The network weights were saved after training.

[0045] In actual detection, the normalized image is input into the pre-trained network, and forward propagation is performed: the image data passes through each network layer sequentially to extract multi-scale features; spatial resolution is restored through deconvolution layers; and finally, pixel-level prediction probabilities are obtained through a sigmoid activation function. The predicted probabilities are compared with a set threshold (e.g., 0.5). Pixels greater than the threshold are marked as 1 (snow-covered areas), and pixels less than the threshold are marked as 0 (non-snow-covered areas), generating a binary mask image. The entire detection process is end-to-end, requiring no manual design of a feature extractor; the model automatically learns the feature representation of snow-covered areas through training.

[0046] Step 1026: Perform morphological opening operation on the binary mask image to remove isolated noise pixels and obtain a denoised binary image; perform morphological closing operation on the binary image to fill the empty areas in the suspected snow area and obtain a mask for the suspected snow area.

[0047] A binary mask image is an image output by a convolutional neural network containing only 0 and 1 pixel values. Morphological opening is a fundamental operation in image processing, consisting of erosion followed by dilation, used to remove small targets and protrusions. Morphological closing consists of dilation followed by erosion, used to fill small holes and depressions within the target area. Isolated noise pixels are randomly occurring false detection points in non-target areas. A hole region refers to a set of non-target pixels within the target area. A suspected snow cover mask is a complete binary image of the target area obtained after morphological processing.

[0048] Specifically, the first step is the opening operation: A structuring element of appropriate size is selected (e.g., a 3×3 rectangular kernel). An erosion operation is first performed on the binary mask image, replacing each pixel value with the minimum value in its neighborhood. This erosion operation eliminates targets smaller than the structuring element. Then, a dilation operation is performed on the erosion result, replacing each pixel value with the maximum value in its neighborhood. This dilation operation restores the size of the remaining target. Through this process, isolated noise points with areas smaller than the structuring element are completely removed, while larger target regions retain their shape largely unchanged.

[0049] Next, a closing operation is performed: the result of the opening operation is first subjected to a dilation operation, using the same structuring element as the opening operation. The dilation operation expands the target boundary and fills gaps smaller than the structuring element; then, an erosion operation is performed to restore the outer boundary of the target. This processing fills small holes within the target area, making the region more complete. For example, for a binary image of size M×1, a sliding window size of 3 is set, and a local minimum (erosion) or maximum (dilation) is calculated at each location. If the 3×1 neighboring pixel values ​​at a certain location are [1, 0, 1], the erosion result is 0, and the dilation result is 1. This operation is performed on the entire image to complete the morphological processing. The processed mask of the suspected snow-covered area has more regular boundaries and a complete internal region, facilitating subsequent extraction of location information.

[0050] Step 1027: Extract the bounding rectangle of the suspected snow cover area mask, and determine the distance coordinate range and azimuth coordinate range of the bounding rectangle as location information.

[0051] The circumscribed rectangle is the smallest rectangle that completely encloses the target area, determined by the maximum and minimum coordinate values ​​of the target area. The range coordinate range is the extent of the circumscribed rectangle along the distance from the radar to the target. The azimuth coordinate range is the extent of the circumscribed rectangle along the radar antenna scanning direction. Position information is a set of numerical parameters describing the spatial location of the target area.

[0052] Specifically, the suspected snow-covered area mask is first scanned row by row and column, recording the coordinates of all pixels with a value of 1. The mask image size is M×1, where M is the number of distance units. During the scan, the distance unit indices {i1, i2, ..., iK} of all areas marked as snow-covered (pixel value 1) are recorded. The minimum value imin and the maximum value imax are extracted from these indices and converted into actual distance coordinates: Rmin = R0 + imin·ΔR, Rmax = R0 + imax·ΔR, where R0 is the starting position of the distance direction and ΔR is the distance resolution. For example, for a system with a distance resolution ΔR = 0.5m and a starting distance R0 = 50m, if imin = 100 and imax = 200, then the distance coordinate range is [100m, 150m].

[0053] Since the source resolution azimuth image has not undergone fine processing in the azimuth direction, a certain margin needs to be considered for the azimuth coordinate range. Based on the radar system's beamwidth θB and range position, the azimuth coordinate range is calculated as: Amin = -θB / 2, Amax = θB / 2. For example, for a system with a beamwidth θB = 5°, the azimuth coordinate range is [-2.5°, 2.5°]. Finally, the range coordinate range [Rmin, Rmax] and the azimuth coordinate range [Amin, Amax] are combined to form a position information vector for subsequent target area determination. This position information extraction method ensures that the complete suspected snow-covered area is included, while also considering the influence of system parameters, providing accurate spatial range information for subsequent fine processing.

[0054] Step 103: Based on the location information and preset boundary margin, determine the target area suspected to be covered by snow on the imaging plane; perform back projection imaging on the pixels within the target area to obtain a radar image with target resolution.

[0055] Specifically, the first step is to determine the target region: Obtain the range coordinates [Rmin, Rmax] and azimuth coordinates [Amin, Amax] from the location information, and set a preset boundary margin ΔB (e.g., 5m for range and 1° for azimuth). Calculate the expanded target region range: range [Rmin-ΔB, Rmax+ΔB], azimuth range [Amin-ΔB, Amax+ΔB]. Based on the system's range resolution ΔR and azimuth sampling interval ΔA, calculate the number of pixels in the target region: range pixels NR = (Rmax-Rmin+2ΔB) / ΔR, azimuth pixels NA = (Amax-Amin+2ΔB) / ΔA. Create an NR×NA complex matrix as the storage space for the target resolution image.

[0056] Then, back projection imaging is performed: For each pixel (r, a) within the target area, its position (x, y) in the ground coordinate system is calculated. Based on the sampling positions {(xn, yn, zn), n=1, 2, ..., N} of the radar on the circular trajectory, the slant range Rn from each radar position to the pixel is calculated. The complex echo values ​​{Sn} corresponding to each slant range are extracted from the target echo data, and the phase compensation factor exp(-j4πRn / λ) is calculated, where λ is the radar operating wavelength. The compensated echo signals are coherently accumulated: P(r, a)=ΣSn·exp(-j4πRn / λ). Back projection processing is performed on all pixels within the target area to obtain the target resolution radar image. For example, for a 10GHz radar system, for a target area with a pixel size of 0.1m×0.1°, if the area range is 100m×10°, then 1000×100 pixels need to be processed, and each pixel needs to have its echo data from all radar positions processed. This precise imaging method achieves high-quality target imaging by performing independent coherent processing on each pixel.

[0057] Step 104: Input the target resolution radar image into the pre-trained semantic segmentation network for pixel-level classification to obtain the segmentation results of the snow-covered area.

[0058] Target resolution radar images are high-resolution complex images obtained through back-projection imaging. The pre-trained semantic segmentation network is a deep learning model trained on a large number of labeled samples to perform pixel-level classification tasks. Pixel-level classification is the process of classifying each pixel in an image. The segmentation result is a binary image output by the network, marking the spatial distribution of snow-covered areas.

[0059] Specifically, the target resolution radar image is first preprocessed. The complex image is converted into an amplitude image, and the amplitude of the complex value of each pixel is calculated. The amplitude image is then logarithmically transformed and normalized to map the pixel values ​​to the [0, 1] interval, facilitating network processing. The preprocessed image is used as input to the semantic segmentation network.

[0060] The semantic segmentation network adopts a U-Net architecture, consisting of an encoder and a decoder. The encoder comprises multiple convolutional blocks, each containing two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. Downsampling is performed between convolutional blocks using 2×2 max pooling. The specific structure is: Input layer → Convolutional block 1 (64 channels) → Pooling → Convolutional block 2 (128 channels) → Pooling → Convolutional block 3 (256 channels) → Pooling → Convolutional block 4 (512 channels). The decoder performs upsampling through transposed convolutions and concatenates the feature maps with the corresponding layers of the encoder to achieve multi-scale feature fusion. Finally, a pixel-level predicted probability map is output through a 1×1 convolutional layer and a sigmoid activation function.

[0061] The network training uses a cross-entropy loss function, and the training samples include target-resolution radar images and manually labeled snow-covered region masks. After training, in actual detection, the preprocessed image is input into the network to obtain the probability value of each pixel belonging to the snow category. A decision threshold (e.g., 0.5) is set, and pixels with probability values ​​greater than the threshold are marked as snow-covered regions (value 1), while pixels with probability values ​​less than the threshold are marked as non-snow-covered regions (value 0), generating a binary segmentation result. Post-processing of the segmentation result includes removing small noise areas and filling internal holes. For example, for a target image of size 1000×100 pixels, the network completes processing within 0.1 seconds, outputting a segmentation mask of the same size, accurately marking the spatial distribution of snow-covered regions.

[0062] In one possible implementation, the target resolution radar image is input into a pre-trained semantic segmentation network for pixel-level classification to obtain the segmentation result of the snow-covered area, specifically including steps 1041-1044, as follows: Step 1041: Calculate the amplitude of the complex image values ​​in the target resolution radar image to obtain an amplitude image, and perform a logarithmic transformation on the amplitude image to obtain a logarithmic amplitude image; perform histogram equalization on the logarithmic amplitude image to obtain an enhanced image.

[0063] Complex image values ​​are radar echo data in complex form, consisting of real and imaginary parts. An amplitude image is a real-valued image containing only echo signal strength information. Logarithmic transformation is a mathematical operation that converts a linear scale to a logarithmic scale. A logarithmic amplitude image is the image after logarithmic transformation. Histogram equalization is an image enhancement technique that improves image contrast by redistributing pixel grayscale values. An enhanced image is the image after histogram equalization processing.

[0064] Specifically, for each pixel in the target resolution radar image, the complex value z = a + bi is calculated, and its amplitude value |z| = A is calculated, where A is the modulus of the complex number. This operation is performed on all pixels to generate amplitude images of the same size. For example, for the complex value 2 + 3i, its amplitude value is A. A logarithmic operation is performed on each pixel value A in the amplitude image, calculated using the formula V = 20·log10(A), where V is the logarithmically transformed pixel value. 20 is chosen as the coefficient to convert the amplitude value to decibels. This transformation is performed on the entire image to generate a logarithmic amplitude image. For example, a pixel with an amplitude value of 100 will have a value of 40 dB after the logarithmic transformation. Next, histogram equalization is performed. First, the distribution of pixel values ​​in the logarithmic amplitude image is statistically analyzed, and the cumulative distribution function (CDF) is calculated. Let the image size be M×N, and the pixel value range be [Vmin, Vmax]. For any pixel value v, its cumulative distribution function is c(v) = n(v) / (M×N), where n(v) is the number of pixels with values ​​less than or equal to v. Pixel values ​​are mapped according to the cumulative distribution function: v' = Vmin + (Vmax - Vmin)·c(v), where v' is the equalized pixel value. This mapping operation is performed on all pixels in the image to generate an enhanced image. For example, for a logarithmic amplitude image with a dynamic range of [-60dB, 0dB], if a pixel value is -40dB and its cumulative distribution function is 0.6, then the equalized pixel value is -24dB. This processing improves image contrast and makes target features clearer.

[0065] Step 1042: Input the enhanced image into the pre-trained semantic segmentation network. The training samples of the semantic segmentation network are the target resolution radar image and the corresponding pixel-level snow cover annotation information.

[0066] The enhanced image is a radar image processed by histogram equalization. The pre-trained semantic segmentation network is an image segmentation model built and trained based on a deep learning framework. Training samples are data pairs used for network training, including input images and annotation information. The target resolution radar image is a high-resolution complex image processed by backpropagation. The pixel-level snow cover annotation information is a manually annotated binary image that accurately marks the spatial distribution of snow cover areas.

[0067] Specifically, a semantic segmentation network is constructed using the U-Net architecture. The network input is a single-channel image, and the output is a binary segmentation map. The encoder contains four downsampling stages. Each stage uses two 3×3 convolutional layers to extract features, with the number of channels being 64, 128, 256, and 512 respectively. After the convolutional layers, a batch normalization layer and a ReLU activation function are applied, and 2×2 max pooling is used for feature map downsampling. The decoder restores spatial resolution through four upsampling stages. Each stage first uses a 2×2 transposed convolution for upsampling, then concatenates it with the feature map of the corresponding layer of the encoder, performing two 3×3 convolution operations, with the number of channels being 256, 128, 64, and 32 respectively. Finally, a 1×1 convolutional layer is used to convert the feature map into a single channel, and the pixel-level predicted probability is output through the Sigmoid function. The training process uses standardized target-resolution radar images and corresponding snow cover masking as training samples. The cross-entropy loss function is used to measure the difference between the predicted results and the labeled information. The Adam optimizer is used to update network parameters, with a learning rate of 0.001, a batch size of 4, and 100 training epochs. Data augmentation is performed on the training data through random cropping and flipping to improve the model's generalization ability. For an input image of size 1000×100, the network contains approximately 8 million trainable parameters, enabling real-time pixel-level classification after training.

[0068] Step 1043: Classify each pixel in the enhanced image using a pre-trained semantic segmentation network, determine the probability value of each pixel as the snow category, mark pixels with probability values ​​greater than or equal to a preset probability threshold as snow category pixels, and mark pixels with probability values ​​less than the preset probability threshold as non-snow category pixels, and obtain pixel-level classification results.

[0069] Snow-category pixels are those identified as snow-covered areas. Non-snow-category pixels are those identified as non-snow-covered areas. The pixel-level classification result is a binary image labeled with the category of each pixel.

[0070] Specifically, the enhanced image is input into the pre-trained semantic segmentation network, and the prediction results are obtained through the forward propagation calculation of the network. During the forward propagation of the network, the input image sequentially passes through the convolutional layer and pooling layer of the encoder to generate multi-scale feature maps; then, through the deconvolution layer and feature fusion of the decoder, the spatial resolution is restored; finally, through the 1×1 convolutional layer and the Sigmoid activation function, a probability map of the same size as the input image is output. Each pixel value in the probability map represents the probability that the position belongs to the snow cover category, and the value is between 0 and 1. A preset probability threshold T (such as T = 0.5) is set to perform binary processing on the probability map: traverse each pixel in the probability map, record its position coordinates (i, j) and probability value P(i, j). When P(i, j) ≥ T, mark it as 1 at the corresponding position in the output image, indicating the snow cover category; when P(i, j) < T, mark it as 0, indicating the non-snow cover category. For example, for an image of 1000×100 pixels, 100,000 pixel points are processed in sequence. If the predicted probability of a certain pixel is 0.8, which is greater than the threshold of 0.5, then this pixel is classified as the snow cover category. After all pixels are judged, a binary classification result image is obtained, where the area with a value of 1 represents the detected snow cover area, and the area with a value of 0 represents the background area. This classification method based on the probability threshold realizes the accurate division of the snow cover area, and the classification result directly reflects the spatial distribution characteristics of the snow cover.

[0071] Step 1044: Perform connected component analysis on the pixel-level classification results, remove the connected regions with an area smaller than the preset area threshold, and obtain the segmentation results of the snow cover area.

[0072] A connected region is a set of regions composed of pixels with the same pixel value and adjacent positions. The preset area threshold is a judgment criterion for filtering small-area noise regions. The segmentation result of the snow cover area is the final binary mask image after post-processing. The area is the number of pixels included in the connected region.

[0073] Specifically, connected component labeling is performed on pixel-level classification results using the 8-neighborhood connectivity criterion, which considers the eight neighboring pixels around each pixel. Starting from the top left corner of the image, the system scans row by row. When an unlabeled pixel with a value of 1 is encountered, it is used as a seed point and labeled as a new connected region. The 8 neighboring pixels of the seed point are checked, and pixels with a value of 1 are added to the current connected region and assigned the same label. This process is repeated recursively until no more neighboring pixels with a value of 1 can be found. After labeling a connected region, the scanning of unprocessed pixels continues, repeating the above process until all pixels with a value of 1 have been processed. The area of ​​each connected region in the labeled image is counted, which is the number of pixels with the same label number. A preset area threshold S is set (e.g., S = 100 pixels). All connected regions are traversed, and all pixel values ​​in regions with an area less than S are set to 0, while regions with an area greater than or equal to S are retained. For example, for a 1000×100 pixel classification result image, assuming 10 connected regions are detected with areas of 50, 200, 80, 300, 60, 400, 90, 250, 70, and 150 pixels respectively, and an area threshold S=100 pixels is set, then the five regions with areas of 200, 300, 400, 250, and 150 pixels are retained, while the other five smaller regions are deleted. This post-processing method based on connected component analysis effectively removes discrete noise points and small false detection regions from the classification results, improving the accuracy of snow area detection.

[0074] Step 105: Based on the segmentation results, determine the spatial distribution and area information of the snow-covered area, and generate an airport runway snow cover detection report.

[0075] Spatial distribution information is a set of geometric parameters describing the location and shape of a snow-covered area, including coordinate range, centroid location, and area boundaries. Area information is a numerical parameter representing the area occupied by the snow-covered region. An airport runway snow cover inspection report is a standardized document describing the snow distribution. A snow-covered area is a segmented and marked area of ​​snow cover.

[0076] Specifically, the spatial information of the segmentation results is extracted first. A boundary tracking algorithm is used to extract the contour point set for each snow-covered area, and the contour is represented using an 8-direction chain code. Statistical analysis is performed on the contour point set to calculate the spatial range of the area: traversing the contour point coordinates (xi, yi), the coordinates of the four vertices of the smallest bounding rectangle are extracted, and the coordinate range of the area in the range and azimuth directions is recorded. The centroid coordinates of the area are calculated: Xc = ΣxiA / N, Yc = ΣyiA / N, where N is the total number of pixels in the area, and A is the pixel area. The area of ​​each snow-covered area is calculated: the number of pixels in the area is multiplied by the actual area of ​​a single pixel (determined by the range resolution and azimuth resolution). For example, for a system with a range resolution of 0.1m and an azimuth resolution of 0.1°, a snow-covered area containing 1000 pixels has an actual area of ​​10 square meters. The standard format for generating the detection report includes: detection time, radar system parameters (operating frequency, resolution), a summary of the detection results (number of snow-covered areas, total area), and detailed information for each snow-covered area (number, location range, area, centroid coordinates). All information is organized according to a predetermined format to generate a test report document that is easy for humans to view and for the system to archive. This standardized output method allows airport managers to quickly understand the snow distribution on the runway, providing a basis for snow removal operations.

[0077] In one possible implementation, based on the segmentation results, the spatial distribution and area information of the snow-covered area are determined, and an airport runway snow cover detection report is generated. Specifically, this includes steps 1051-1053, as follows: Step 1051: Label the connected components of the segmentation results to obtain multiple independent snow-covered regions; for each snow-covered region, count the number of pixels contained in the snow-covered region, and convert the number of pixels into the actual area according to the resolution of the target resolution radar image.

[0078] Specifically, connected component labeling is performed first, using a two-pass scanning algorithm. The first pass scans the image from left to right and top to bottom. When a pixel with a value of 1 is encountered, the labels of processed pixels in its 4-neighborhood or 8-neighborhood are checked. If no labeled pixel exists in the neighborhood, a new label is assigned; if a labeled pixel exists, the minimum label value is assigned, and the label equivalence relationship is recorded. The second pass normalizes the labels, unifying equivalent labels to the minimum value. For example, for a 100×100 image, if three connected regions are detected, they are labeled as 1, 2, and 3 respectively. Then, the area of ​​each labeled region is calculated: a label counter array count[] is created, the labeled image is scanned, and for each label value i, the counter count[i] is incremented. For the range resolution ΔR and azimuth resolution ΔA of the target resolution radar image, the actual area of ​​a single pixel is calculated: S_pixel = ΔR × R × ΔA, where R is the range coordinate value of the pixel. The actual area is obtained by multiplying the number of pixels by the area of ​​a single pixel: S_real = count[i] × S_pixel. For example, if a connected region contains 1000 pixels, has a distance resolution of 0.1 meters, an azimuth resolution of 0.1 degrees, and a target distance of 100 meters, then the area of ​​a single pixel is approximately 0.017 square meters, and the actual area of ​​the region is 17 square meters. This calculation process is repeated for all marked regions to obtain complete area information. This pixel-based area calculation method takes into account the geometric characteristics of radar imaging and provides accurate snow cover area data.

[0079] Step 1052: Extract the centroid coordinates and boundary contour coordinates of each snow-covered area; convert the centroid coordinates and boundary contour coordinates from the imaging plane coordinate system to the airport pavement coordinate system to obtain the target centroid coordinates and target boundary contour coordinates in the airport pavement coordinate system.

[0080] Specifically, the snow-covered area features are first extracted in the imaging plane coordinate system. For each marked area, the centroid coordinates are calculated: traversing the position coordinates (ri, ai) of all pixels within the area, where ri is the distance coordinate and ai is the azimuth coordinate, the weighted average values ​​rc = Σri / N and ac = Σai / N are calculated, where N is the total number of pixels in the area, yielding the centroid position (rc, ac). A boundary tracking algorithm is then used to extract the contour coordinates: selecting the top-left pixel of the area as the starting point, and sequentially checking the 8 neighboring pixels in a clockwise direction, recording the boundary pixel positions until returning to the starting point. For each boundary pixel, its distance coordinate rb and azimuth coordinate ab are recorded, resulting in a contour point sequence {(rb1, ab1), (rb2, ab2), ..., (rbM, abM)}, where M is the number of contour points. Then, a coordinate system transformation is performed: Let the radar installation location in the airport pavement coordinate system be (x0, y0, z0), and the angle between the radar scanning reference direction and true north be θ0. For any point (r, a) in the imaging plane, its three-dimensional coordinates in the radar coordinate system are first calculated: x'=r·cos(a), y'=r·sin(a), z'=0. Then, a coordinate rotation and translation transformation is performed: x=x'·cos(θ0)-y'·sin(θ0)+x0, y=x'·sin(θ0)+y'·cos(θ0)+y0, z=z'+z0. This transformation is then applied sequentially to the centroid coordinates and the coordinates of all contour points to obtain the target position information in the airport pavement coordinate system. For example, if the centroid of a snow-covered area has coordinates (100m, 30°) in the imaging plane, the radar position is (0, 0, 5m), and the reference direction angle is 45°, then the transformed centroid coordinates are approximately (61.2m, 86.6m, 5m). This coordinate transformation method establishes a correspondence between the image space and the actual space, facilitating the determination of the specific location of the snow-covered area on the airport pavement.

[0081] Step 1053: Generate an airport runway snow cover detection report containing the actual area of ​​each snow-covered area, the target centroid coordinates, and the target boundary contour coordinates.

[0082] Specifically, the inspection report document is constructed according to a standardized format, including two parts: a report header and inspection results. The report header records basic inspection parameters: inspection time, radar operating parameters (center frequency, transmit power, beamwidth), inspection range (distance range, azimuth range), and coordinate reference information (radar position coordinates, reference azimuth angle). The inspection results section first provides a statistical overview: the total number of snow-covered areas N detected, and the cumulative coverage area Stotal = ΣSi, where Si is the area of ​​the i-th area. Then, detailed information for each snow-covered area is recorded sequentially according to its area number: area number, area size (unit: square meters), centroid location (east coordinate XC, north coordinate YC, elevation coordinate ZC), number of contour points M, and contour point coordinate sequence {(X1, Y1, Z1), (X2, Y2, Z2), ..., (XM, YM, ZM)}. For example, a typical inspection report might include the following: Inspection time: 2023-12-22 10:00:00; Radar frequency: 10GHz; 3 snow-covered areas detected, total area: 52.5 square meters; Area 1: 15.2 square meters, centroid coordinates: (125.3m, 45.6m, 0m), number of contour points: 50; Area 2: 20.8 square meters, centroid coordinates: (180.2m, 62.4m, 0m), number of contour points: 65; Area 3: 16.5 square meters, centroid coordinates: (220.5m, 38.7m, 0m), number of contour points: 55. Saving the inspection report in a standard text format facilitates manual viewing and system parsing. This standardized report format comprehensively records the spatial characteristics of snow distribution, providing precise location and extent information for airport pavement snow removal operations.

[0083] In the above embodiments, a basic target area determination framework was achieved through location information extraction and area range calculation. To further improve the spatial positioning accuracy of snow-covered areas and reduce the impact of coordinate range truncation on area integrity assessment, this application also provides a boundary-expanding imaging method. This method determines the expanded target area range by analyzing the boundary features of location information and performs refined back projection processing, enabling the system to more accurately acquire the snow-covered area distribution status in complex pavement environments. The following is a combination of... Figure 2 Another fast BP imaging snow detection method is described in the embodiments of this application: Please see Figure 2 This is a flowchart illustrating a rapid BP imaging snow detection method in an embodiment of this application.

[0084] Step 201: Obtain the minimum and maximum values ​​of the distance coordinate range and the azimuth coordinate range from the location information.

[0085] Specifically, traverse and compare the position information data, and extract the boundary values in the range direction and azimuth direction respectively. First, initialize four variables: the minimum range value Rmin is set to the maximum detection range of the system (such as 1000 meters), the maximum range value Rmax is set to 0, the minimum azimuth value Amin is set to 360 degrees, and the maximum azimuth value Amax is set to 0. Read each sampling point coordinate (ri, ai) in the position information in sequence, where ri is the range value and ai is the azimuth angle. Perform update operations on each range value ri: if ri < Rmin, then update Rmin = ri; if ri > Rmax, then update Rmax = ri. Perform update operations on each azimuth angle ai: if ai < Amin, then update Amin = ai; if ai > Amax, then update Amax = ai. For example, the position information of a certain radar system contains 1000 sampling points, and the coordinate values are distributed within the range of 100 meters to 500 meters in range and 15 degrees to 75 degrees in azimuth. After traversing and comparing, Rmin = 100 meters, Rmax = 500 meters, Amin = 15 degrees, and Amax = 75 degrees are obtained. After completing the traversal, the precise spatial range of the observation area is obtained: the range direction range [Rmin, Rmax], and the azimuth direction range [Amin, Amax]. This boundary extraction method determines the data space range to be processed and provides the necessary parameter information for subsequent imaging processing.

[0086] Step 202: Subtract the preset boundary margin from the minimum value of the range direction coordinate range to determine the starting coordinate in the range direction; add the preset boundary margin to the maximum value of the range direction coordinate range to determine the ending coordinate in the range direction; subtract the preset boundary margin from the minimum value of the azimuth direction coordinate range to determine the starting coordinate in the azimuth direction; add the preset boundary margin to the maximum value of the azimuth direction coordinate range to determine the ending coordinate in the azimuth direction.

[0087] Specifically, preset boundary margin parameters ΔR and ΔA are set, where ΔR is the range margin (e.g., 10 meters) and ΔA is the azimuth margin (e.g., 2 degrees). The range coordinate range [Rmin, Rmax] is expanded: the starting range coordinate Rs = Rmin - ΔR is calculated, and the ending range coordinate Re = Rmax + ΔR is calculated. The azimuth coordinate range [Amin, Amax] is expanded: the starting azimuth coordinate As = Amin - ΔA is calculated, and the ending azimuth coordinate Ae = Amax + ΔA is calculated. When determining the expansion range, the boundary validity must be checked: if Rs is less than the radar's minimum detection range (e.g., 50 meters), Rs is set to the minimum detection range; if Re is greater than the radar's maximum detection range (e.g., 1000 meters), Re is set to the maximum detection range; if As is less than 0 degrees, As is set to 0 degrees; if Ae is greater than 360 degrees, Ae is set to 360 degrees. For example, if the original range of an observation area is 100 meters to 500 meters in distance and 15 degrees to 75 degrees in azimuth, with a preset boundary margin ΔR = 10 meters and ΔA = 2 degrees, then the expanded processing range is: [90 meters, 510 meters] in distance and [13 degrees, 77 degrees] in azimuth. By adding boundary margins to expand the processing range, the edge data truncation effect is avoided, improving the quality of the imaging results. This range expansion method limits the excessive increase in computation while ensuring processing accuracy.

[0088] Step 203: Determine the target area on the imaging plane based on the range starting coordinates, range ending coordinates, azimuth starting coordinates, and azimuth ending coordinates.

[0089] Specifically, the pixel size of the target region is first calculated based on the range resolution ΔR and the azimuth resolution ΔA. The number of pixels in the range direction is Nr = (Re - Rs) / ΔR, and the number of pixels in the azimuth direction is Na = (Ae - As) / ΔA, where Rs and Re are the start and end coordinates in the range direction, and As and Ae are the start and end coordinates in the azimuth direction, respectively. The calculation results are rounded up to obtain the pixel matrix size of the target region, Nr × Na. The coordinate mapping relationship is established: for a pixel at position (i, j) in the matrix, its corresponding range coordinate r = Rs + i · ΔR, and azimuth coordinate a = As + j · ΔA, where the value of i ranges from [0, Nr - 1], and the value of j ranges from [0, Na - 1]. For example, if the target area has a distance range of [90 meters, 510 meters] and an azimuth range of [13 degrees, 77 degrees], and the distance resolution ΔR = 0.5 meters and the azimuth resolution ΔA = 0.1 degrees, then the number of pixels in the distance direction Nr = 840, the number of pixels in the azimuth direction Na = 640, and the target area size is 840 × 640 pixels. For the pixel position (100, 200), its corresponding actual coordinates are: distance r = 90 + 100 × 0.5 = 140 meters, azimuth angle a = 13 + 200 × 0.1 = 33 degrees. This region division method based on uniform sampling establishes a precise correspondence between the image space and the actual space, providing a standardized data structure for subsequent imaging processing. By reasonably selecting resolution parameters, both imaging accuracy and data processing volume are ensured.

[0090] In one possible implementation, after determining the target area on the imaging plane based on the temperature field distribution data and carbon dioxide concentration field distribution data of each monitoring section, steps 2031-2034 are further included, as follows: Step 2031: Determine the set of pixels that need to be back-projected based on the distance start coordinates, distance end coordinates, azimuth start coordinates, and azimuth end coordinates of the target area.

[0091] Specifically, firstly, the sampling interval is set according to the system parameters. The range sampling interval ΔR is selected as half of the range resolution (e.g., 0.05 meters), and the azimuth sampling interval ΔA is selected as half of the azimuth resolution (e.g., 0.05 degrees). This oversampling setting improves imaging accuracy. The number of sampling points in the target area is calculated based on the sampling interval: the number of range sampling points M = (Re - Rs) / ΔR, and the number of azimuth sampling points N = (Ae - As) / ΔA, where Rs and Re are the start and end coordinates of the range, and As and Ae are the start and end coordinates of the azimuth. A two-dimensional sampling grid is established: for the i-th range sampling point and the j-th azimuth sampling point, their coordinate values ​​ri = Rs + i · ΔR, aj = As + j · ΔA, where i ranges from [0, M-1] and j ranges from [0, N-1]. The coordinate information of all sampling points is stored in an array in the form {(r0, a0), (r0, a1), ..., (rM-1, aN-1)}. For example, if the target area ranges from 90 meters to 510 meters in distance and 13 degrees to 77 degrees in azimuth, with a sampling interval ΔR = 0.05 meters and ΔA = 0.05 degrees, then the number of distance sampling points M = 8400, the number of azimuth sampling points N = 1280, and the total number of sampling points is 10.752 million. The sampling points are organized as follows: an M×N two-dimensional array is established to store the sampling point indices, where array element (i, j) corresponds to the sampling point at coordinates (ri, aj). This regularized sampling scheme ensures complete coverage of the imaging area and provides a standardized processing object for the back projection algorithm. By adjusting the sampling interval, a balance can be achieved between computational efficiency and imaging quality.

[0092] Step 2032: For each pixel in the pixel set, calculate the slant range between the pixel and each radar position based on the spatial coordinates of the pixel and the spatial coordinates of each radar position on the arc trajectory; based on the slant range, extract the echo complex value corresponding to the radar position from the target echo data.

[0093] Specifically, let the radar installation height be h and the rotation radius be R. For the radar position at azimuth angle a, its spatial coordinates are (R·cos(a), R·sin(a), h). For the pixel P with coordinates (rp, ap) within the target area, its spatial coordinates are (rp·cos(ap), rp·sin(ap), 0). Calculate the slant range from radar position A to pixel P: First, calculate the x-direction distance difference dx = xA - xP, the y-direction distance difference dy = yA - yP, and the z-direction distance difference dz = zA - zP. Substitute the distance differences in the three directions into the three-dimensional distance calculation formula to obtain the slant range d. Convert the slant range d into a range sampling index: n = round((d - R0) / ΔR), where R0 is the starting sampling position in the range direction, ΔR is the range sampling interval, and round represents the rounding operation. Extract the corresponding complex value S(m, n) from the target echo data matrix S based on the range sampling index n and the azimuth sampling index m. For example, a radar installed at a height of 5 meters and a rotation radius of 2 meters has pixel coordinates of (100 meters, 30 degrees). The radar coordinates at an azimuth angle of 45 degrees are (1.414 meters, 1.414 meters, 5 meters). The slant range is calculated to be 100.12 meters using three-dimensional distance calculation. If the starting position for distance sampling is 50 meters and the sampling interval is 0.1 meters, the corresponding distance sampling number is 501. The complex value at position (m, 501) is extracted from the echo data matrix, where m is the azimuth sampling number corresponding to 45 degrees. This echo data extraction method based on geometric relationships achieves accurate mapping from spatial points to signal data, providing a data foundation for back projection accumulation.

[0094] Step 2033: Determine the phase compensation factor based on the slant range, and multiply the phase compensation factor by the corresponding complex echo value to obtain the compensated echo value; coherently accumulate the compensated echo values ​​for each radar position to obtain the complex image value of the pixel.

[0095] Specifically, phase compensation and accumulation are performed on the echo signals from each radar location. First, the phase compensation factor is calculated: for a radar system with an operating wavelength of λ, the slant range from radar location A to pixel P is d, and the phase compensation factor is exp(j·4π·d / λ), where j is the imaginary unit. The corresponding complex echo value S(m, n) is obtained from the echo data matrix and multiplied by the phase compensation factor to obtain the compensated echo value. This operation is repeated for all radar locations, performing complex addition to obtain the complex image value of the pixel. Taking a Ka-band radar system as an example, with an operating frequency of 35 GHz and a wavelength of approximately 8.6 mm, the slant ranges of a pixel corresponding to five radar locations are 100.12 m, 100.15 m, 100.18 m, 100.21 m, and 100.24 m, respectively. Five phase compensation factors are calculated. For these five radar locations, complex values ​​were extracted from the echo data, namely 1.2+0.5j, 1.1+0.6j, 1.0+0.7j, 0.9+0.8j, and 0.8+0.9j. Each complex value was multiplied by its corresponding phase compensation factor to obtain five compensated echo values. Complex addition was then performed on these five compensated echo values ​​to obtain the complex image value for that pixel. This phase-compensated coherent accumulation method achieves target energy focusing and improves the imaging signal-to-noise ratio. This process was repeated for all pixels within the target area to complete the back-projection imaging of the entire region.

[0096] Step 2034: Traverse all pixels in the pixel set, use the complex image value of each pixel as the pixel value at the corresponding position, and generate a target resolution radar image.

[0097] Specifically, based on the number of range samples M and azimuth samples N for the target area, an M×N complex matrix is ​​constructed as the target resolution radar image. The data type of the image matrix is ​​set to complex, with each element containing a real part and an imaginary part. The pixel set is traversed in ascending order of range and then azimuth: for the pixel in the i-th row and j-th column, the value of i ranges from [0, M-1], and the value of j ranges from [0, N-1]. The complex image value of this pixel is read from the back projection processing result, denoted as z = a + bi, where a is the real part and b is the imaginary part. The complex image value z is directly assigned to the element at position (i, j) of the image matrix. For example, if the image size is 1000×800 pixels, and the pixel at position (500, 400) has a complex image value of 1.5 + 2.3j, then this complex value is stored in the 500th row and 400th column of the image matrix. After assigning values ​​to all pixels, the complete target resolution radar image is obtained. This direct mapping method preserves all the information of the complex image, providing a complete data foundation for subsequent image analysis and processing. For storage format, a binary file is used to store the complex matrix, with each complex value occupying 8 bytes of space, of which the real and imaginary parts each occupy 4 bytes, for a total file size of 8 × M × N bytes.

[0098] The following describes a fast BP imaging snow detection system according to an embodiment of the present invention from the perspective of hardware processing. Please refer to [link to relevant documentation]. Figure 3 This is a schematic diagram of a rapid BP imaging snow detection system in an embodiment of this application.

[0099] It should be noted that, Figure 3 The structure of the rapid BP imaging snow detection system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0100] like Figure 3 As shown, a rapid BP imaging snow detection system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 302 or a program loaded from storage section 308 into random access memory (RAM) 303, such as executing the methods in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0101] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0102] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0103] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0105] Specifically, a rapid BP imaging snow detection system according to this embodiment includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements a rapid BP imaging snow detection method provided in the above embodiment.

[0106] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the fast BP imaging snow detection system described in the above embodiments; or it may exist independently and not assembled into the fast BP imaging snow detection system. The storage medium carries one or more computer programs, which, when executed by a processor of the fast BP imaging snow detection system, cause the fast BP imaging snow detection system to implement the fast BP imaging snow detection method based on IoT data encryption transmission provided in the above embodiments.

Claims

1. A rapid BP imaging method for snow cover detection, characterized in that, The method includes: The raw echo data collected by the ground-based circular arc trajectory synthetic aperture radar system on the airport pavement is acquired, and the raw echo data is processed by range pulse compression to obtain the target echo data. The target echo data is coherently accumulated in the azimuth direction to obtain the source resolution azimuth image; The location information of suspected snow cover is extracted from the source resolution azimuth image, and the location information includes a range coordinate range and an azimuth coordinate range. Based on the location information and the preset boundary margin, the target area suspected of being covered by snow is determined on the imaging plane; Back-projection imaging is performed on the pixels within the target area to obtain a radar image with target resolution; The target resolution radar image is input into a pre-trained semantic segmentation network for pixel-level classification to obtain the segmentation result of the snow-covered area; Based on the segmentation results, the spatial distribution and area information of the snow-covered area are determined, and an airport runway snow cover detection report is generated.

2. The method according to claim 1, characterized in that, The step of performing azimuth coherent accumulation on the target echo data to obtain a source resolution azimuth image includes: The target echo data are arranged according to the radar position to generate a data matrix; For each range cell in the data matrix, the echo complex value of the corresponding radar position is extracted to generate the azimuth echo sequence of the range cell; Calculate the phase difference between the complex values ​​of each echo in the azimuth echo sequence, and perform phase alignment based on the phase difference; The phase-aligned azimuth echo sequence is accumulated to obtain the azimuth cumulative response value of the range cell; By traversing each of the distance cells, the cumulative azimuth response value of each distance cell is used as the pixel value of the corresponding distance position to generate the source resolution azimuth image.

3. The method according to claim 1, characterized in that, The step of extracting the suspected snow location information from the source resolution azimuth image includes: The image pixel values ​​in the source resolution orientation image are mapped to a preset numerical range to obtain a normalized image; The normalized image is input into a pre-trained convolutional neural network. The pre-trained convolutional neural network performs forward propagation calculation on the normalized image and outputs a binary mask image of the suspected snow area. The training samples of the convolutional neural network are the source resolution azimuth image and the corresponding snow area annotation information. Perform morphological opening operation on the binary mask image to remove isolated noise pixels and obtain a denoised binary image; Perform morphological closing operation on the binary image to fill the voids in the suspected snow area and obtain the mask of the suspected snow area; Extract the bounding rectangle of the suspected snow-covered area mask, and determine the range of distance coordinates and the range of azimuth coordinates of the bounding rectangle as the location information.

4. The method according to claim 1, characterized in that, The step of determining the suspected snow-covered target area on the imaging plane based on the location information and the preset boundary margin includes: Obtain the minimum and maximum values ​​of the distance coordinate range and the minimum and maximum values ​​of the azimuth coordinate range from the location information; The starting coordinates of the distance direction are determined by subtracting the preset boundary margin from the minimum value of the distance coordinate range. The maximum value of the distance coordinate range is added to the preset boundary margin to determine the distance termination coordinate; The starting coordinates of the azimuth direction are determined by subtracting the preset boundary margin from the minimum value of the azimuth coordinate range. Add the maximum value of the azimuth coordinate range to the preset boundary margin to determine the azimuth termination coordinate; The target region is determined on the imaging plane based on the range start coordinate, the range end coordinate, the azimuth start coordinate, and the azimuth end coordinate.

5. The method according to claim 4, characterized in that, The step of performing back-projection imaging on pixels within the target area to obtain a radar image of target resolution includes: Based on the distance start coordinates, distance end coordinates, azimuth start coordinates, and azimuth end coordinates of the target area, determine the set of pixels that need to be back-projected into image. For each pixel in the pixel set, the slant distance between the pixel and each radar position is calculated based on the spatial coordinates of the pixel and the spatial coordinates of each radar position on the arc trajectory. Based on the slant range, extract the complex echo value corresponding to the radar position from the target echo data; The phase compensation factor is determined based on the slant distance, and the phase compensation factor is multiplied by the corresponding echo complex value to obtain the compensated echo value; The compensated echo values ​​at each radar location are coherently accumulated to obtain the complex image value of the pixel. Traverse all pixels in the pixel set, and use the complex image value of each pixel as the pixel value at the corresponding position to generate the target resolution radar image.

6. The method according to claim 1, characterized in that, The step of inputting the target resolution radar image into a pre-trained semantic segmentation network for pixel-level classification to obtain the segmentation result of the snow-covered area includes: The amplitude of the complex image values ​​in the target resolution radar image is calculated to obtain an amplitude image, and the amplitude image is logarithmically transformed to obtain a logarithmic amplitude image. The logarithmic magnitude image is subjected to histogram equalization to obtain an enhanced image; The enhanced image is input into a pre-trained semantic segmentation network, the training samples of which are target resolution radar images and corresponding pixel-level snow cover annotation information; The pre-trained semantic segmentation network is used to classify each pixel in the enhanced image, determine the probability value of each pixel as a snow category, mark pixels with probability values ​​greater than or equal to a preset probability threshold as snow category pixels, and mark pixels with probability values ​​less than the preset probability threshold as non-snow category pixels, thereby obtaining pixel-level classification results. Connectivity analysis is performed on the pixel-level classification results to remove connected regions with an area smaller than a preset area threshold, thereby obtaining the segmentation results of the snow-covered region.

7. The method according to claim 1, characterized in that, Based on the segmentation results, the spatial distribution and area information of the snow-covered area are determined, and an airport runway snow cover detection report is generated, including: The segmentation results are labeled with connected components to obtain multiple independent snow-covered regions; For each snow-covered area, the number of pixels contained in the snow-covered area is counted, and the number of pixels is converted into the actual area according to the resolution of the target resolution radar image. Extract the centroid coordinates and boundary contour coordinates of each snow-covered region; The centroid coordinates and the boundary contour coordinates are converted from the imaging plane coordinate system to the airport pavement coordinate system to obtain the target centroid coordinates and target boundary contour coordinates in the airport pavement coordinate system. Generate an airport runway snow cover detection report containing the actual area of ​​each snow-covered area, the target centroid coordinates, and the target boundary contour coordinates.

8. A rapid BP imaging snow cover detection system, characterized in that, The rapid BP imaging snow detection system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the rapid BP imaging snow detection system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the fast BP imaging snow detection system, the fast BP imaging snow detection system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the fast BP imaging snow detection system, the fast BP imaging snow detection system performs the method as described in any one of claims 1-7.

Images