A deep learning-based automatic evaluation method for the clearance environment of a meteorological observation station

Abstract

The present application relates to the technical field of image recognition, in particular to a meteorological observation station clearance environment automatic evaluation method based on deep learning, comprising the following steps: collecting panoramic image to calculate distortion parameters, establishing pre-processing feature set, using coding-decoding network to fuse attention mechanism to extract features and adjust sampling interval to generate semantic feature matrix, screening pixel category to extract boundary coordinates to construct skyline sequence, calculating coordinate offset and field of view angle mapping to obtain shielding elevation angle to generate final environment evaluation data. In the present application, the edge distortion is overcome by jointly correcting the field of view deformation of multiple source data, the global correlation is captured relying on the network attention mechanism to avoid the loss of position features, the sampling rate is dynamically scaled based on the feature distribution to retain high-frequency details and accurately locate the skyline boundary, the real skyline sequence is extracted combined with the effective field of view parameter to map and convert the real elevation angle, and the large calculation deviation caused by complex terrain is effectively eliminated to greatly improve the clearance judgment precision.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an automated assessment method for the airspace environment of meteorological observation stations based on deep learning. Background Technology

[0002] Image recognition technology refers to the technical system that uses computers to collect, analyze, and understand image information. Its core aspects include image acquisition, image preprocessing, feature extraction, target detection, pattern recognition, and result output. It acquires scene image data through camera equipment and then analyzes it based on information such as pixel grayscale changes, texture distribution, edge contours, and spatial structure to determine the location, shape, category, and changes of target objects. This technology is widely used in security monitoring, remote sensing mapping, autonomous driving, industrial inspection, and meteorological observation. In particular, the continuous processing and interpretation of multi-source image data under complex environmental conditions has become an important aspect of this field.

[0003] The traditional deep learning-based automated assessment method for the airspace environment of meteorological observation stations refers to the determination of the unobstructed space around the meteorological observation station. This involves deploying fixed cameras around the station to acquire environmental images from different directions, normalizing the size of the acquired images and converting them into pixel matrices, which are then input into a pre-constructed multi-layer convolutional network structure. Edge and texture information are extracted through layer-by-layer convolution operations, and multiple pooling operations are performed to reduce spatial resolution. The extracted feature vectors are then input into a fully connected structure for classification to identify obstructing objects such as buildings, vegetation, and terrain undulations. Simultaneously, by combining the set observation elevation and azimuth angle ranges, the identification results are spatially mapped and compared point by point according to a preset airspace angle threshold, thereby completing the process of determining the airspace environment of the observation station.

[0004] Relying on multiple fixed cameras to acquire multi-directional images can easily lead to fragmented fields of view and aggravated edge distortion. Forced size normalization can easily destroy the actual proportional structure of the original scene. Layer-by-layer convolution and continuous pooling operations inevitably cause serious loss of subtle spatial location feature information. Fully connected structures forcibly map and classify, which severely sever the spatial correlation of pixels, resulting in extremely blurry edge positioning. Performing mechanical mapping at a fixed angle lacks an adaptive adjustment mechanism for the deformation features of the real environment, and produces a serious elevation angle offset in complex terrain, which leads to significant errors in the final evaluation results. Summary of the Invention

[0005] To address the technical problems of relying on multiple fixed cameras to acquire multi-directional images, which easily leads to fragmented fields of view and aggravated edge distortion; forced size normalization easily destroys the actual proportional structure of the original scene; layer-by-layer convolution and continuous pooling operations inevitably cause serious loss of subtle spatial location feature information; forced mapping and classification of fully connected structures severely sever the spatial correlation of pixels, resulting in extremely blurry edge localization; and mechanical mapping performed at a fixed angle lacks an adaptive adjustment mechanism for the deformation features of the real environment, producing serious elevation angle offsets in complex terrain, thus leading to significant errors in the final evaluation results, this invention provides an automated evaluation method for the airspace environment of meteorological observation stations based on deep learning.

[0006] To achieve the above objectives, this invention employs an automated assessment method for the airspace environment of meteorological observation stations based on deep learning, comprising the following steps: S1: Collect panoramic images and station illumination data of meteorological observation stations, calculate distortion correction parameters of panoramic images of meteorological observation stations based on station illumination data, adjust pixel positions of panoramic images of meteorological observation stations and perform brightness equalization processing, and construct a preprocessed image feature set. S2: Input the preprocessed image feature set into a scene semantic segmentation model with an encoder-decoder structure, use the channel-space joint attention mechanism to perform weighted operations to extract spatial context features, calculate the dilated convolution sampling interval parameter and adaptively adjust it based on the feature distribution statistical results to generate an environmental semantic feature matrix. S3: Based on the environmental semantic feature matrix, perform classification judgment to filter environmental occlusion pixels and sky category pixels, extract the spatial boundary coordinate values ​​of environmental occlusion pixels and sky category pixels, arrange the spatial boundary coordinate values ​​in the horizontal direction, and construct the skyline coordinate sequence. S4: Extract the vertical coordinate offset values ​​of the elements in the skyline coordinate sequence, perform angle mapping transformation based on the preset field of view values ​​to calculate the occlusion elevation angle values ​​of the horizontal azimuth elements, and generate site environment assessment data.

[0007] As a further aspect of the present invention, the preprocessed image feature set includes corrected panoramic images and equalized brightness data; the environmental semantic feature matrix includes semantic category labels, spatial context weights, and category confidence for each pixel; the skyline coordinate sequence includes horizontal and vertical coordinate values ​​corresponding to each horizontal direction; and the site environment assessment data includes occlusion elevation angles and clearance angles in each direction.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect panoramic images and station illumination data of meteorological observation station, extract various light intensity values ​​in the station illumination data, compare the deviation from the preset calibration threshold to calculate the light source deviation ratio, perform derivative calculation to solve the variation coefficients, extract the variation coefficients of each order and fill them according to the index position to obtain the distortion correction parameters. S102: Based on the distortion correction parameters, read the two-dimensional coordinate vectors of each pixel in the panoramic image of the meteorological observation station, calculate the product between the distortion correction parameters and the two-dimensional coordinate vectors to extract the horizontal and vertical offsets, and perform translation mapping operation on the two-dimensional coordinate vectors according to the horizontal and vertical offsets to generate a pixel position reconstruction matrix. S103: Traverse the pixel positions to reconstruct the grayscale components within the matrix, accumulate the grayscale components and divide by the total number of pixels to obtain the average grayscale value, divide the average grayscale value by the color reference value to calculate the scaling ratio, multiply the scaling ratio with the grayscale components, extract texture features and stitch them together to establish a preprocessed image feature set.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the preprocessed image feature set into the encoder of the scene semantic segmentation model to perform convolution to extract spatial activation components, multiply the channel weight values ​​with the planar spatial weights to calculate the joint attention bias term, and perform dot product between the joint attention bias term and the spatial activation components to establish a spatial context feature tensor. S202: Calculate the squared difference of the activation distribution values ​​within the spatial context feature tensor from the mean, accumulate the squared differences to obtain the feature distribution variance, and perform proportional scaling and rounding down on the preset reference spacing according to the ratio between the feature distribution variance and the preset discrete threshold to obtain the adaptive sampling spacing parameter. S203: Adjust the dilation rate of the convolution kernel within the decoder according to the adaptive sampling interval parameter, input the spatial context feature tensor into the decoder to perform step sampling according to the dilation rate, extract semantic representation vectors at each resolution scale and perform concatenation operation to generate an environmental semantic feature matrix.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Read the corresponding feature vector in the environmental semantic feature matrix, read the category reference template, calculate the similarity between the feature vector and the occlusion and sky vectors in the category reference template, filter the pixels with similarity greater than the preset judgment threshold and perform category assignment, and establish a scene pixel category set. S302: Traverse the neighborhood window of the grid nodes in the pixel category set of the scene, read the category labels of the center node and the adjacent nodes, calculate the absolute value of the difference between the two labels, retrieve the center node corresponding to the absolute value greater than zero as the edge mutation point, read the plane grid coordinates of the mutation point, and obtain the boundary coordinate dataset. S303: Extract the horizontal coordinate values ​​corresponding to the grid coordinates in the boundary coordinate dataset, perform sorting operation according to the incrementing rule, read the corresponding vertical coordinate values ​​according to the sorting order, associate the horizontal coordinate values ​​with the vertical coordinate values, and obtain the skyline coordinate sequence.

[0011] As a further aspect of the present invention, the preset judgment threshold is the upper limit of the interval constructed by the median and standard deviation of the similarity distribution interval calculated based on the feature vector and the occlusion vector and the sky vector in the category reference template, respectively. The pixel points with similarity greater than the preset judgment threshold are selected and the category assignment is performed by calculating the similarity between the feature vector of each pixel point and the occlusion vector and the sky vector in the category reference template, respectively, and the category label corresponding to the maximum similarity is selected as the category label of the pixel point.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Read the spatial pixel grid points in the skyline coordinate sequence, extract the vertical coordinates of the spatial pixels, calculate the absolute difference between the longitudinal coordinates of the image reference baseline, construct the grid single-point offset based on the absolute difference and merge them in horizontal spatial order to obtain the coordinate offset value set. S402: Call the coordinate offset value set, divide the camera's vertical field of view by the total number of pixel rows in the vertical direction of the image to calculate the single pixel angular resolution, multiply the single-point offset of each grid by the single pixel angular resolution to obtain the actual elevation angle transformation, integrate the actual elevation angle transformation, and generate the occlusion elevation angle distribution matrix. S403: For the aforementioned occlusion elevation angle distribution matrix, calculate the horizontal azimuth angle based on the horizontal index of the grid nodes and the horizontal angular resolution. Add the horizontal azimuth angle to the initial azimuth parameter of the geographic north direction to calibrate the absolute azimuth angle. Perform key-value mapping between the actual elevation angle transformation and the absolute azimuth angle to establish site environment assessment data.

[0013] As a further aspect of the present invention, the method further includes: S5: Call the site environment assessment data, transmit the site environment assessment data to the central server node through the site communication terminal, compare and calculate the site environment assessment data with the preset airspace threshold, and integrate and construct a dynamic digital assessment file based on the calculation results; The dynamic digital assessment archive includes the site's airspace compliance status, airspace deviation values ​​in each direction, and historical assessment records.

[0014] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Extract the original payload sequence from the site environment assessment data, modulate the original payload sequence into a continuous analog waveform according to the baseband modulation mapping table and transmit it to the central server node, and extract the digital payload unit by analyzing the analog-to-digital conversion of the continuous analog waveform according to the demodulation parameters, and generate the restored assessment data matrix. S502: Call the restored evaluation data matrix to extract the actual elevation angle measurement value, subtract the actual elevation angle measurement value from the preset clearance threshold parameter to obtain the relative space difference, extract the grid node coordinates for reference values ​​greater than zero, and assemble the grid node coordinates and the relative space difference to obtain the set of out-of-limit deviation indicators. S503: Read the absolute azimuth angle according to the set of out-of-limit deviation indicators, write it into the local node structure tree template identifier area, fill the relative spatial difference into the second-level node of the local node structure tree template, perform encoding merging on the nodes in the local node structure tree template, and establish a dynamic digital evaluation file.

[0015] As a further aspect of the present invention, the preset airspace threshold parameter is a fixed upper limit value of a range determined according to the statistical distribution of historical elevation angle measurements in the site environment assessment data, and the relative spatial difference is the retained value after non-negative filtering in the difference sequence calculated point-by-point between the actual elevation angle measurement and the preset airspace threshold parameter.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the inherent drawbacks of field of view fragmentation and deformation distortion are overcome by jointly acquiring ambient lighting and panoramic images to extract deformation parameters and reconstruct pixel coordinate distribution. Relying on the joint encoding and decoding architecture, the deep fusion channel and spatial attention mechanism capture global contextual dependencies to completely avoid the loss of positional information. The convolution sampling interval is dynamically scaled according to the discreteness of feature distribution to retain high-frequency detail representation in order to accurately delineate the boundary of scene semantic categories. The coordinate sequence of real skyline undulations is extracted and combined with effective visual field parameters to perform multi-dimensional viewpoint mapping transformation to dynamically restore the three-dimensional occlusion elevation angle. The traditional mechanical angle comparison mode is completely abandoned, effectively eliminating the calculation error caused by complex undulating terrain conditions, thereby greatly improving the accuracy of judgment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides an automated assessment method for the airspace environment of meteorological observation stations based on deep learning, comprising the following steps: S1: Collect panoramic images and station illumination data of meteorological observation stations, calculate distortion correction parameters of panoramic images of meteorological observation stations based on station illumination data, adjust pixel positions of panoramic images of meteorological observation stations and perform brightness equalization processing, and construct a preprocessed image feature set. S2: Input the preprocessed image feature set into the scene semantic segmentation model with encoder-decoder structure, use the channel-space joint attention mechanism to perform weighted operation to extract spatial context features, calculate the dilated convolution sampling interval parameter and adaptively adjust it based on the feature distribution statistical results to generate the environmental semantic feature matrix; S3: Based on the environmental semantic feature matrix, perform classification and judgment to filter environmental occlusion pixels and sky category pixels, extract the spatial boundary coordinate values ​​of environmental occlusion pixels and sky category pixels, arrange the spatial boundary coordinate values ​​in the horizontal direction, and construct the skyline coordinate sequence. S4: Extract the vertical coordinate offset values ​​of elements in the skyline coordinate sequence, perform angle mapping transformation based on the preset field of view values ​​to calculate the occlusion elevation angle values ​​of horizontal azimuth elements, and generate site environment assessment data. S5: Call the site environment assessment data, transmit the site environment assessment data to the central server node through the site communication terminal, compare and calculate the site environment assessment data with the preset airspace threshold, and integrate and construct a dynamic digital assessment archive based on the calculation results; The preprocessed image feature set includes corrected panoramic images and equalized brightness data; the environmental semantic feature matrix includes semantic category labels, spatial context weights, and category confidence for each pixel; the skyline coordinate sequence includes horizontal and vertical coordinate values ​​corresponding to each horizontal direction; the site environmental assessment data includes occlusion elevation angles and clearance angles in each direction; and the dynamic digital assessment archive includes the site clearance compliance status, clearance deviation values ​​in each direction, and historical assessment records.

[0022] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect panoramic images and station illumination data of meteorological observation station, extract various light intensity values ​​in the station illumination data, compare the deviation from the preset calibration threshold to calculate the light source deviation ratio, perform derivative calculation to solve the variation coefficients, extract the variation coefficients of each order and fill them according to the index position to obtain the distortion correction parameters. A panoramic industrial camera and ambient light sensor, deployed 5.5 meters above the ground at the top of the meteorological observation station, acquire panoramic images of the station at a rate of 2 frames per second, with a horizontal resolution of 4096 pixels and a vertical resolution of 2048 pixels, while simultaneously recording the station's illumination data at corresponding times. Ambient light intensity values ​​distributed at eight different standard azimuths are extracted from the station's illumination data. The resistance change rate of the internal photoresistor is read and converted into actual illuminance data. A pre-set standard reference illuminance, calculated as 6000 lux (based on the average light intensity measured over 10 consecutive days at noon on a cloudless, clear day), is read. The extracted light intensity values ​​are then compared with the standard reference illuminance to obtain the absolute deviation. This absolute deviation is then divided by the standard reference illuminance to obtain the preliminary light source deviation for each azimuth. The eight preliminary light source deviations are summed, and the sum is divided by 8 to calculate the overall light source deviation ratio. For example, if the actual illuminance in the due south direction is 5600 lux, the absolute value of the difference between this and 6000 lux is 400. Dividing 400 by 6000 gives 0.066. Simultaneously, if the illuminance in the due north direction is 3200 lux, its absolute deviation is 2800. Dividing this by 6000 gives a preliminary light source deviation of 0.466. Adding the eight values, including 0.066 and 0.466, and dividing by 8 gives a light source deviation ratio of 0.254. Next, for the illuminance values ​​in adjacent directions, using the azimuth angle change in degrees as the independent variable and the illuminance value as the dependent variable, a difference operation is performed instead of a derivative calculation. The specific operation involves subtracting the light intensity value of the previous direction from the light intensity value of the later direction, and then dividing the result by the angular difference between the two directions. For example, if the angular difference is 45 degrees, the difference between 5600 lux in the south direction and 4800 lux in the southeast direction is 800, which, when divided by 45, yields a first-order variational coefficient of 17.77. Further, a second round of adjacent subtraction and division operations is performed on the obtained first-order variational coefficients to obtain the second-order variational coefficients. A 3x8 empty matrix is ​​created, using the azimuth index corresponding to each extracted variational coefficient as the column coordinate and the order index as the row coordinate. The corresponding variational coefficient values ​​are then filled into the corresponding coordinate positions in the empty matrix. For example, the first-order variational coefficient of 17.77 for the south direction is filled into the first row and first column, ultimately generating a distortion correction parameter matrix containing all spatial gradient information.

[0023] Table 1: Site Illumination and Variation Coefficients

[0024] Table 1 shows the actual measurement data for some directions and the corresponding coefficients obtained after variational calculation.

[0025] S102: Based on the distortion correction parameters, read the two-dimensional coordinate vectors of each pixel in the panoramic image of the meteorological observation station, calculate the product between the distortion correction parameters and the two-dimensional coordinate vectors to extract the horizontal and vertical offsets, and perform translation mapping operations on the two-dimensional coordinate vectors according to the horizontal and vertical offsets to generate a pixel position reconstruction matrix. Based on the distortion correction parameter matrix generated in the above steps, the positions of all pixels in the panoramic image of the meteorological observation station are read row by row and column by column. The horizontal and vertical coordinates of the currently processed pixel in its image plane are extracted, and the horizontal and vertical coordinates are combined to construct a two-dimensional coordinate vector for each pixel. The horizontal and vertical coordinate values ​​in the two-dimensional coordinate vector are read, and the horizontal coordinate values ​​are multiplied by the azimuth coefficient corresponding to the first row of the distortion correction parameter matrix. At the same time, the vertical coordinate values ​​are multiplied by the azimuth coefficient corresponding to the second row of the distortion correction parameter matrix. The results of the two multiplication operations are used as the horizontal offset in the horizontal direction and the vertical offset in the vertical direction of the current pixel, respectively. For example, if a pixel's horizontal coordinate is 1500 and its vertical coordinate is 800, the horizontal correction coefficient in the corresponding distortion correction parameter matrix is ​​extracted to be 1.02, and the vertical correction coefficient is 0.95. Multiplying 1500 by 1.02 yields the corrected expected horizontal coordinate of 1530. Subtracting the original horizontal coordinate 1500 from 1530 gives a horizontal offset of 30. Similarly, multiplying the vertical coordinate 800 by 0.95 yields 760, and subtracting the original vertical coordinate 800 from 760 gives a vertical offset of -40. Following these calculated horizontal and vertical offsets, addition operations are performed on the corresponding horizontal and vertical coordinate values ​​in the original two-dimensional coordinate vector to complete the translation mapping operation. For example, adding the offset of 30 to the original horizontal coordinate 1500 gives a new horizontal coordinate of 1530, and adding -40 to the original vertical coordinate 800 gives a new vertical coordinate of 760. Create a new empty matrix with the same resolution as the original panoramic image (4096 x 2048). Use the new horizontal and vertical coordinates after the translation and mapping operation as index positions, and fill the color values ​​of the corresponding pixels into the new empty matrix. After traversing and completing the translation and filling of all pixels, generate a pixel position reconstruction matrix.

[0026] S103: Traverse pixel positions to reconstruct grayscale components within the matrix, accumulate grayscale components and divide by the total number of pixels to obtain the average grayscale value, divide the average grayscale value by the color reference value to calculate the scaling ratio, multiply the scaling ratio with the grayscale components, extract texture features and stitch them together to establish a preprocessed image feature set. Iterate through each independent pixel unit in the pixel location reconstruction matrix established in the above steps, extracting the grayscale component value representing brightness information from the color channel data of each pixel unit. This grayscale component value is obtained by multiplying the values ​​of the red, green, and blue color channels by their respective standard brightness weights and summing them. Set an accumulator variable with an initial value of 0. As the traversal process progresses, add the grayscale component value of each extracted pixel to this accumulator variable one by one until all pixels in the matrix have been traversed. Then obtain the total number of pixels in the image, which is the product of 4096 and 2048, 8388608. Divide the total value in the accumulator variable by the total number of pixels to obtain the average grayscale value of the entire image. Read the pre-calibrated color reference value. The color reference value is set with reference to the median grayscale value of historical images under standard meteorological visibility greater than 10 kilometers and sufficient lighting conditions. By sorting and extracting the median from 30 days of historical normal observation images, the color reference value is finally set to 128. The calculated average grayscale value is used as the dividend, and the set color reference value of 128 is used as the divisor to perform a division operation, calculating the scaling ratio used to adjust the brightness of the current image. For example, when the accumulated average grayscale value of the image is 96, 96 is divided by the color reference value of 128, resulting in a scaling ratio of 0.75. Then, the pixel position reconstruction matrix is ​​re-traversed, and the calculated scaling ratio of 0.75 is multiplied by the original grayscale component of each pixel. If the original grayscale component of a pixel is 100, it is multiplied by 0.75 to obtain the adjusted new grayscale component of 75. Based on this brightness uniformity, the grayscale difference distribution between the center pixel and its eight neighboring pixels is calculated using local windows consisting of 3 by 3 pixels, extracting texture feature values ​​reflecting local roughness. The extracted texture feature values ​​are vectorized into one dimension according to the spatial arrangement of the original pixels, and the texture feature vectors of all local windows are concatenated end-to-end to establish a preprocessed image feature set.

[0027] Please see Figure 3 The specific steps of S2 are as follows: S201: Call the preprocessed image feature set to input the encoder in the scene semantic segmentation model to perform convolution to extract the spatial activation component, multiply the channel weight value with the planar spatial weight to calculate the joint attention bias term, and perform dot product with the spatial activation component to establish the spatial context feature tensor. The preprocessed image feature set established by the aforementioned operations is converted into a floating-point data tensor and then input into the encoder within the scene semantic segmentation model. This encoder consists of three consecutive cascaded convolutional layers. The input layer receives texture feature data with a resolution of 4096×2048 and one channel. The first hidden layer uses a 5×5 convolutional kernel with a stride of 2 to extract shallow edge information and output 64 channels. The second and third hidden layers both use 3×3 convolutional kernels with a stride of 1 for sliding feature extraction. After each convolutional operation, a linear rectified activation function is forcibly applied for nonlinear mapping, ultimately outputting a spatial activation component with 256 channels and a resolution one-quarter of the original. After obtaining the spatial activation component, the global average value of this component in the channel dimension is extracted as the channel weight value, and the local maximum response value of this component in the two-dimensional spatial dimension is extracted as the planar spatial weight. The extracted channel weight values ​​and the planar spatial weights are directly multiplied to calculate the joint attention bias term. For example, the channel weight of the 10th channel is extracted to be 0.65, and its corresponding planar spatial weight is 0.80. Multiplying 0.65 by 0.80 yields a joint attention bias term of 0.52 for this channel. Next, the original feature values ​​corresponding to the spatial activation components are read, and the calculated joint attention bias term 0.52 is multiplied element-wise with the original feature values. Assuming the original feature value at a certain spatial location is 1.5, multiplying it by 0.52 yields a feature value of 0.78 after the attention mechanism is fused. This multiplication operation is repeated for all channels and all spatial locations within the spatial activation components. The updated values ​​are then rearranged and combined according to the original tensor dimensions to establish a spatial context feature tensor containing both global and local attention features.

[0028] S202: Calculate the squared difference of the activation distribution values ​​within the spatial context feature tensor from the mean, accumulate the squared differences to obtain the feature distribution variance, and perform proportional scaling and rounding down on the preset baseline spacing based on the ratio between the feature distribution variance and the preset discrete threshold to obtain the adaptive sampling spacing parameter. For the spatial context feature tensor established in the above steps, the feature activation values ​​of all channels within the tensor are traversed, and the arithmetic mean of all feature activation values ​​is calculated as the mean of the activation distribution. Then, all feature activation values ​​are traversed again, and the calculated mean of the activation distribution is subtracted from each specific feature activation value to obtain the deviation of each value. This deviation is then multiplied by itself to calculate the deviation-mean squared difference. A cumulative variable is established, and all calculated deviation-mean squared differences are added together. The sum is then divided by the total number of feature activation values ​​to obtain the feature distribution variance reflecting the overall feature dispersion. A pre-set discrete threshold is read. This preset discrete threshold is set based on the feature dispersion state under unobstructed airspace conditions of multiple standard weather stations. By calculating the feature distribution variance of 500 baseline unobstructed images and taking their 95th percentile values, the preset discrete threshold is finally set to 2.5. The calculated feature distribution variance is used as the numerator, and the preset discrete threshold of 2.5 is used as the denominator. A division operation is performed between the two to obtain their ratio. The preset baseline spacing value is read. This baseline spacing refers to the default interval parameter between convolution kernel elements in a conventional dilated convolution operation, and is set to a baseline value of 2. The previously calculated ratio is then multiplied by the preset baseline spacing for scaling. For example, when the calculated feature distribution variance is 3.8, 3.8 is divided by the preset discrete threshold of 2.5 to obtain a ratio of 1.52. This ratio of 1.52 is then multiplied by the preset baseline spacing of 2 to obtain a scaled floating-point spacing value of 3.04. This floating-point spacing value is then rounded down, removing the decimal part of 3.04 to obtain the final integer parameter of 3, thus generating an adaptive sampling spacing parameter suitable for the current image feature complexity.

[0029] S203: Adjust the dilation rate of the convolutional kernel within the decoder according to the adaptive sampling interval parameter, input the spatial context feature tensor into the decoder to perform step sampling according to the dilation rate, extract semantic representation vectors at each resolution scale and perform concatenation operation to generate the environmental semantic feature matrix. Based on the adaptive sampling interval parameter generated in the above steps, the dilation rate of the convolutional kernels in the decoder layer of the scene semantic segmentation model is dynamically adjusted. This decoder consists of four symmetrical inverse convolutional layers, responsible for progressively restoring low-resolution features to the original image size. Each layer uses a skip connection operation to fuse the detailed information from the corresponding layer in the encoder. The extracted adaptive sampling interval parameter is directly assigned to the dilation rate attribute of the dilated convolutional layer in the decoder, changing the spacing between regular adjacent samples to the interval span set by this adaptive parameter. The previously established spatial context feature tensor is fed into the decoder as input data, and each layer of the network strictly follows the adjusted dilation rate to perform step sampling operations. For example, when the adaptive sampling interval parameter is set to 3, when the convolutional kernel slides on the spatial context feature tensor, the physical spacing between its internal computational elements is forcibly stretched to 3 pixels, thereby significantly expanding the network's feature extraction field of view without increasing the total number of computational parameters. As the tensor data is passed and amplified layer by layer in the decoder, intermediate feature results at the corresponding resolution scale are output at each specific depth layer. Semantic representation vectors were extracted at 1 / 4, 1 / 2, and full resolution scales. For the extracted semantic representation vectors at different resolution scales, a bilinear interpolation algorithm was used to force the spatial dimensions of all vectors to be stretched and aligned to the same 4096×2048 pixel size as the full resolution. After complete spatial alignment, the semantic representation vectors from different levels were concatenated one by one along the data channel dimension, so that each pixel contained multi-scale semantic expression information, ultimately generating a deeply fused environmental semantic feature matrix.

[0030] Please see Figure 4 The specific steps of S3 are as follows: S301: Read the corresponding feature vector in the environmental semantic feature matrix, read the category reference template, calculate the similarity between the feature vector and the occlusion and sky vectors in the category reference template, filter the pixels with similarity greater than the preset judgment threshold and perform category assignment, and establish a set of scene pixel categories. From the established environmental semantic feature matrix, feature vectors containing multi-scale information are read point-by-point according to spatial grid coordinates. A pre-calibrated category reference template is read from the local storage module. This template contains a large number of high-dimensional feature vectors extracted from standard samples, including building vectors, terrain vectors, vegetation vectors representing environmental occlusions, and a sky vector representing a clean sky background. For each currently read pixel feature vector, an inner product operation is performed with the building vector, terrain vector, and vegetation vector in the category reference template. The result is divided by the product of the magnitudes of the two vectors to obtain the cosine similarity, which is used as the corresponding building similarity, terrain similarity, and vegetation similarity. Similarly, the same operation is performed with the sky vector to obtain the sky similarity. To construct adaptive judgment conditions, the system pre-calculates the similarity distribution interval of all pixels in the current panoramic image. The similarity datasets of all pixels with respect to each environmental occlusion vector are statistically analyzed, and the median and standard deviation of this dataset are calculated. The median and standard deviation are read, and then added together to construct an upper limit value reflecting the upper limit of data fluctuation. This upper limit value is then directly set as the preset judgment threshold. For example, if the median of the sky similarity of the current image is 0.75 and the standard deviation is 0.10, adding them together gives 0.85, which is the preset judgment threshold. The filtering process then begins, checking the similarity of each pixel with the building vector, terrain vector, vegetation vector, and sky vector. For example, if a pixel has a sky similarity of 0.92, a building similarity of 0.15, a terrain similarity of 0.10, and a vegetation similarity of 0.20, its sky similarity of 0.92 is determined to be greater than the preset judgment threshold of 0.85. Furthermore, since 0.92 is the largest among all similarities, the sky category label corresponding to the largest similarity of 0.92 is selected, and a sky category assignment operation is performed on the pixel. Repeat the above calculation and comparison filtering operations for all pixels, and integrate the mapping relationship containing all pixel grid coordinates and corresponding category labels to establish a scene pixel category set.

[0031] Table 2: Scene Pixel Similarity and Category Assignment Table

[0032] Table 2 shows the category labels assigned to adjacent pixels after similarity comparison and threshold judgment.

[0033] S302: Traverse the neighborhood window of the grid nodes in the pixel category set of the scene, read the category labels of the center node and the adjacent nodes, calculate the absolute value of the difference between the two labels, retrieve the center node corresponding to the absolute value greater than zero as the edge mutation point, read the plane grid coordinates of the mutation point, and obtain the boundary coordinate dataset. Traverse the established set of scene pixel categories. In a 2D grid coordinate system, create a 3x3 grid node neighborhood window centered on each pixel node. During the sliding traversal, extract the center node within the current neighborhood window and read its associated category label value. Simultaneously, read the category label values ​​of the eight adjacent nodes surrounding the center node. Since the category labels are encoded using integers, for example, assign the number 0 to the sky category, the number 1 to the building category, the number 2 to the terrain category, and the number 3 to the vegetation category. Calculate the difference between the category label of the center node and the category label of each adjacent node, and take the absolute value of this difference. Establish a retrieval rule: if at least one of the eight absolute value results formed by the center node and its eight adjacent nodes is greater than zero, then the center node's location is considered to have undergone a category state switch, and it is retrieved and marked as an edge abrupt change point. For example, if a central node has a category label of 0 (sky category) and its adjacent node to its right has a category label of 3 (vegetation category), subtracting 3 from 0 yields a negative 3, and the absolute value is 3. Because this absolute value is greater than zero, the central node is immediately identified as an edge abrupt change point. For nodes identified as edge abrupt change points, their inherent 2D planar grid coordinates within the panoramic image are read; for example, their horizontal coordinate is 2400 and their vertical coordinate is 1350. The planar grid coordinates of all nodes identified as edge abrupt change points are collected and merged, and node coordinates with absolute values ​​all equal to 0 (i.e., located within homogeneous regions) are removed. All the collected edge abrupt change point coordinates are stored in a separate data sequence, ultimately obtaining a boundary coordinate dataset containing only boundary location information.

[0034] S303: Extract the horizontal coordinate values ​​corresponding to the grid coordinates in the boundary coordinate dataset, perform sorting operations according to the increasing rule, read the corresponding vertical coordinate values ​​according to the sorting order, associate the horizontal coordinate values ​​with the vertical coordinate values, and obtain the skyline coordinate sequence. The boundary coordinate dataset obtained through the aforementioned operations is used to read each grid coordinate information stored in the dataset one by one, and the corresponding horizontal coordinate value is extracted individually. All extracted horizontal coordinate values ​​are stored in a one-dimensional array, and the quicksort algorithm is called to sort all horizontal coordinate values ​​in the array according to an ascending order. During this ascending sorting process, the initial position index of each horizontal coordinate value in the original boundary coordinate dataset is strictly recorded to ensure that the coordinate correspondence is not disrupted. After the sorting operation is completed, based on the sorted order, the recorded initial position index is used to return to the original boundary coordinate dataset to read the corresponding vertical coordinate value. For example, if there are two points in the original dataset, point A has a horizontal coordinate of 150 and a vertical coordinate of 80, and point B has a horizontal coordinate of 120 and a vertical coordinate of 95, after extracting the horizontal coordinates and sorting them according to the ascending order, the horizontal coordinate of point B (120) will precede the horizontal coordinate of point A (150). Then, according to the sorting order, the vertical coordinate value of point B (95) is read first, followed by the vertical coordinate value of point A (80). The reordered horizontal coordinate values ​​are associated one-to-one with their corresponding vertical coordinate values, assembling them into ordered coordinate pairs. If multiple points with identical horizontal coordinate values ​​are found during the sorting process, their vertical coordinate values ​​are extracted and compared. Only the point with the smallest vertical coordinate value, i.e., the highest point on the skyline in physical space, is retained, and the remaining overlapping points are removed. After the above sorting, association, and deduplication operations, all processed ordered coordinate pairs are concatenated to obtain a skyline coordinate sequence extending continuously from the leftmost to the rightmost edge of the image.

[0035] Please see Figure 5 The specific steps of S4 are as follows: S401: Read the spatial pixel grid points in the skyline coordinate sequence, extract the vertical coordinates of the spatial pixels, calculate the absolute difference between the vertical coordinates of the reference baseline of the image, construct the grid single-point offset based on the absolute difference and merge them in horizontal spatial order to obtain the coordinate offset value set. Read the previously generated continuous skyline coordinate sequence and traverse all spatial pixel grid points contained within the sequence. For each spatial pixel grid point, extract its vertical coordinate in the two-dimensional coordinate system, i.e., the vertical pixel row number of that grid point. Read the vertical coordinate value of the image reference baseline from the system's environmental calibration parameters. This reference baseline represents the horizontal horizon line captured by the meteorological observation camera when it is in an absolutely horizontal state, and its vertical coordinate value is fixed at 1024 after the camera completes level calibration. Perform a subtraction operation between the extracted spatial pixel vertical coordinate and the preset image reference baseline vertical coordinate 1024, and calculate the absolute value of the final result of the subtraction operation to calculate the absolute difference between the two. For example, the horizontal coordinate of the 500th spatial pixel grid point in the skyline coordinate sequence is 1200, and its vertical coordinate is 824. Subtract the vertical coordinate of this point from the reference baseline vertical coordinate 1024 to calculate the absolute difference of 200. The absolute difference of 200 represents the physical pixel span by which the skyline grid point deviates from the ideal horizontal plane in vertical space. Based on each calculated absolute difference, a one-to-one mapping is established between it and the horizontal coordinate of the currently processed grid point, thus constructing a single-point grid offset reflecting the vertical obstruction in a single horizontal direction. Next, strictly following the horizontal spatial order of the skyline coordinate sequence (from smallest to largest), all constructed single-point grid offsets are sequentially stored in a continuous one-dimensional linear list and merged. After traversing and processing all spatial pixel grid points, a complete set of coordinate offset values ​​is finally obtained, recording the vertical undulation of each point within the entire horizontal viewport.

[0036] S402: Call the coordinate offset value set, divide the camera's vertical field of view by the total number of pixel rows in the vertical direction of the image to calculate the single pixel angular resolution, multiply the single-point offset of each grid by the single pixel angular resolution to obtain the actual elevation angle transformation, integrate the actual elevation angle transformation, and generate the occlusion elevation angle distribution matrix. The system retrieves the pre-defined camera hardware parameters from the constructed coordinate offset value set and obtains the vertical field of view of the camera lens. This vertical field of view is a fixed value determined through precise measurement using optical instruments, for example, set to 90 degrees. Simultaneously, it reads the total number of pixel rows in the vertical direction of the acquired image; this value represents the vertical resolution parameter of the panoramic image, fixed at 2048 rows. Using the extracted camera vertical field of view value of 90 as the dividend and the total number of pixel rows in the vertical direction of the image of 2048 as the divisor, a division calculation is performed to determine the actual physical angular span represented by a single pixel in vertical space, i.e., the single-pixel angular resolution. For example, dividing 90 by 2048 yields a single-pixel angular resolution of 0.0439 degrees. The system then sequentially extracts the pre-calculated grid single-point offset values ​​from the coordinate offset value set. Each extracted grid single-point offset value is multiplied directly by the previously calculated single-pixel angular resolution of 0.0439 to obtain the actual elevation angle change caused by obstructions in each direction. For example, the grid offset of the 500th horizontal position, calculated in the previous step, is 200 pixels. Multiplying the value 200 by the single-pixel angular resolution of 0.0439, we calculate the actual elevation angle change caused by obstruction at this position as 8.78 degrees, meaning the obstruction elevation angle at this location is 8.78 degrees. For all offset data in the coordinate offset value set, repeat the above extraction and multiplication operations to obtain all actual elevation angle changes covering the entire horizontal viewpoint. Establish a one-dimensional empty data array, and fill all the obtained actual elevation angle changes into this array one by one according to the horizontal position arrangement for integration. After integrating all angle data, a blocking elevation angle distribution matrix is ​​finally generated that comprehensively reflects the actual blocking elevation angle distribution status in various directions around the observation station.

[0037] S403: For the occlusion elevation angle distribution matrix, calculate the horizontal azimuth angle based on the horizontal index of the grid nodes and the horizontal angular resolution. Add the horizontal azimuth angle to the initial azimuth parameter of the geographic north direction to calibrate the absolute azimuth angle. Perform key-value mapping between the actual elevation angle transformation and the absolute azimuth angle to establish site environmental assessment data. For the occlusion elevation angle distribution matrix generated in the previous step, each data item contained in the matrix is ​​read sequentially. During the reading process, the horizontal index value of the grid node corresponding to the currently processed data item is extracted simultaneously. The horizontal field of view of the camera is set to 360 degrees, and the horizontal resolution of the image is 4096 pixels. Dividing 360 by 4096 yields a horizontal angular resolution of 0.0878 degrees. The extracted horizontal index value of the grid node is multiplied by this horizontal angular resolution of 0.0878 to calculate the horizontal azimuth angle of the current pixel relative to the leftmost starting point of the panoramic image. For example, when the horizontal index value of the 1000th pixel is 1000, multiplying it by 0.0878 yields a horizontal azimuth angle of 87.8 degrees at that location. The pre-calibrated initial azimuth parameter of geographic true north is read from the built-in electronic compass of the meteorological station. This initial azimuth parameter records the angular offset of the leftmost starting point of the panoramic image relative to the true geographic true north; for example, this initial azimuth parameter is set to 12.5 degrees. The calculated horizontal azimuth angle is added to the initial azimuth parameter of geographic north to determine the absolute azimuth of the data item in the real world. For example, adding 12.5 degrees to 87.8 degrees yields an absolute azimuth of 100.3 degrees for the grid node. If the sum exceeds 360 degrees, it is subtracted to keep the value within the standard azimuth range. The actual elevation change corresponding to the horizontal index is extracted from the shading elevation angle distribution matrix; for example, a value of 8.78 degrees is extracted. A key-value pair storage structure is established, using the calibrated absolute azimuth of 100.3 degrees as the unique index key and the extracted actual elevation change of 8.78 degrees as the corresponding associated data value, and a key-value mapping operation is performed. After traversing the entire matrix to complete the absolute azimuth calibration and key-value mapping of all data, the structured data containing the precise absolute azimuth and actual shading elevation angle is packaged to finally establish complete site environment assessment data.

[0038] Please see Figure 6 The specific steps of S5 are as follows: S501: Extract the original payload sequence from the site environment assessment data, modulate the original payload sequence into a continuous analog waveform according to the baseband modulation mapping table and transmit it to the central server node, and extract the digital payload unit by analyzing the analog number of the continuous analog waveform according to the demodulation parameters to generate the restored assessment data matrix. Based on the complete site environment assessment data established in the aforementioned steps, the system sequentially reads the absolute azimuth and corresponding actual elevation angle transformations within the data, converts the values ​​into a continuous data bit stream using binary encoding, and extracts the original payload sequence. It then calls the baseband modulation mapping table pre-stored on the local communication interface. This mapping table specifies the conversion rules between bit stream data and analog signal waveform characteristics, and orthogonal amplitude modulation (AEM) technology is used to process the data. According to the baseband modulation mapping table, each group of binary bit segments in the original payload sequence is mapped to analog signal characteristic parameters with specific amplitude and phase. Subsequently, a digital-to-analog converter (DAC) uses these characteristic parameters to generate a continuous analog waveform in the radio frequency band. This continuous analog waveform is transmitted to the remote central server node via a communication antenna in the form of radio electromagnetic waves. The receiving device in the central server node captures the attenuated electromagnetic wave signal and initially amplifies it into a continuous analog waveform at the receiving end. The server node reads the built-in demodulation parameters, which include the carrier frequency synchronized with the transmitter and the phase detection rule settings. Based on these demodulation parameters, analog-to-digital parsing is performed on the received continuous analog waveform. The specific operation involves discretizing the analog waveform using an analog-to-digital converter at a specific sampling rate, and then using demodulation parameters to inversely reconstruct the binary values ​​behind the discrete samples, extracting digital payload units composed of 0s and 1s. Error correction and format reconstruction operations are performed on the extracted digital payload units. Following the same key-value pair structure as the original site environment assessment data, the bitstream is converted back into decimal floating-point numbers in the form of absolute azimuth and actual elevation angle transformations. The reconstructed structured data with clearly defined azimuth coordinates and corresponding elevation angle information is stored in a two-dimensional spatial array, ultimately generating a restored assessment data matrix that is completely consistent with the transmitter data.

[0039] S502: Call the restored evaluation data matrix to extract the actual elevation angle measurement value, subtract the actual elevation angle measurement value from the preset clearance threshold parameter to obtain the relative space difference, extract the grid node coordinates for reference values ​​greater than zero, and assemble the grid node coordinates and the relative space difference to obtain the set of out-of-limit deviation indicators. The system retrieves the restoration assessment data matrix generated by the central server and extracts the actual elevation angle measurements for each azimuth according to the data's sorting sequence. Based on the meteorological observation station's clearance standards, a preset clearance threshold parameter is read. This preset clearance threshold parameter is determined by statistically distributing historical elevation angle measurements from the target station over the past three years, calculating the maximum elevation angle for all historically unobstructed observation days, and adding a 0.5-degree safety margin upwards to establish a fixed upper limit for the range. For example, the preset clearance threshold parameter determined by historical data is 5.0 degrees. Each extracted actual elevation angle measurement is used as the minuend, and the preset clearance threshold parameter of 5.0 degrees is used as the subtrahend, resulting in a relative spatial difference. For example, if the extracted actual elevation angle measurement for a certain azimuth is 6.2 degrees, subtracting the preset clearance threshold parameter of 5.0 degrees yields a relative spatial difference of 1.2 degrees. The differences calculated for all corresponding azimuths are aggregated into a difference sequence, and a non-negative filtering operation is performed on this sequence, removing all values ​​less than or equal to zero and retaining only those greater than zero as the retained values ​​after non-negative filtering. For reference values ​​greater than zero, it indicates that the obstruction at that azimuth has encroached upon the standard airspace protection zone for meteorological observation. From the reconstructed assessment data matrix, the coordinates of the grid node associated with that location are extracted using the azimuth index corresponding to the value. For example, for a difference of 1.2 degrees, the extracted grid node's horizontal coordinate is 1500 and its vertical coordinate is 820. The extracted grid node coordinates are used as spatial identifiers and assembled with the relative spatial difference obtained through non-negative filtering to form a data structure entity containing the coordinates of the exceeding limit location and the severity of the exceeding limit. After traversal, all assembled data structure entities are combined to obtain a set of exceeding deviation indicators focusing on the specific azimuth of the illegal obstruction.

[0040] Table 3: Actual Elevation Angle and Exceedance Deviation Index

[0041] Table 3 shows in detail the screening results of out-of-limit indicators generated after subtracting the threshold from the actual elevation angle measurement for some locations.

[0042] S503: Read the absolute azimuth angle based on the set of out-of-limit deviation indicators, write it into the identifier area of ​​the local node structure tree template, fill the relative spatial difference into the second-level node of the local node structure tree template, perform encoding and merging on the nodes in the local node structure tree template, and establish a dynamic digital evaluation file. Based on the established set of out-of-limit deviation indicators, each out-of-limit data record is traversed, and the absolute azimuth angle associated with each record is read one by one. In the system backend of the central server, a hierarchical local node structure tree template is initialized. The read absolute azimuth angle is used as the basic classification label for environmental risk, and is directly fixed to the first layer of the local node structure tree template, i.e., the labeling area, using a write operation, making this azimuth angle the root index for subsequent attachment of detailed risk data. For example, if the read absolute azimuth angle of the out-of-limit direction is 122.5 degrees, a basic node named 122.5 degrees is created in the labeling area. Next, the out-of-limit deviation indicator set is accessed again to extract the relative spatial difference corresponding to this absolute azimuth angle. The extracted relative spatial difference, such as the previously calculated 1.2 degrees, along with detailed data containing coordinate information, is written into the sub-layer node belonging to the absolute azimuth angle node in the local node structure tree template using a data filling command. Following the above logic, the absolute azimuth angle is continuously read and the relative spatial difference is filled in until all data in the entire set of out-of-limit deviation indicators are allocated and mounted to the corresponding hierarchical nodes in the structure tree. After the tree-like filling of all data items is completed, an encoding and merging operation is performed on all hierarchical nodes within the entire local node structure tree template. Specifically, the hierarchical relationship of the tree and the numerical information stored in the nodes are serialized into text encoding using a standard data interaction format to eliminate blank redundancy between levels and package them into a unified structured archive file. This file contains comprehensive information such as the specific azimuth of the out-of-limit occurrence, the severity of the out-of-limit, and spatial grid positioning. Finally, the encoded and merged archive file is persistently stored on the server hard drive to establish a dynamic digital assessment archive.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A deep learning-based automated assessment method for the airspace environment of meteorological observation stations, characterized in that, Includes the following steps: S1: Collect panoramic images and station illumination data of meteorological observation stations, calculate distortion correction parameters of panoramic images of meteorological observation stations based on station illumination data, adjust pixel positions of panoramic images of meteorological observation stations and perform brightness equalization processing, and construct a preprocessed image feature set. S2: Input the preprocessed image feature set into a scene semantic segmentation model with an encoder-decoder structure, use the channel-space joint attention mechanism to perform weighted operations to extract spatial context features, calculate the dilated convolution sampling interval parameter and adaptively adjust it based on the feature distribution statistical results to generate an environmental semantic feature matrix. S3: Based on the environmental semantic feature matrix, perform classification judgment to filter environmental occlusion pixels and sky category pixels, extract the spatial boundary coordinate values ​​of environmental occlusion pixels and sky category pixels, arrange the spatial boundary coordinate values ​​in the horizontal direction, and construct the skyline coordinate sequence. S4: Extract the vertical coordinate offset values ​​of the elements in the skyline coordinate sequence, perform angle mapping transformation based on the preset field of view values ​​to calculate the occlusion elevation angle values ​​of the horizontal orientation elements, and generate site environment assessment data.

2. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 1, characterized in that, The preprocessed image feature set includes corrected panoramic images and equalized brightness data; the environmental semantic feature matrix includes semantic category labels, spatial context weights, and category confidence for each pixel; the skyline coordinate sequence includes horizontal and vertical coordinate values ​​corresponding to each horizontal direction; and the site environment assessment data includes occlusion elevation angles and clearance angles in each direction.

3. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect panoramic images and station illumination data of meteorological observation station, extract various light intensity values ​​in the station illumination data, compare the deviation from the preset calibration threshold to calculate the light source deviation ratio, perform derivative calculation to solve the variation coefficients, extract the variation coefficients of each order and fill them according to the index position to obtain the distortion correction parameters. S102: Based on the distortion correction parameters, read the two-dimensional coordinate vectors of each pixel in the panoramic image of the meteorological observation station, calculate the product between the distortion correction parameters and the two-dimensional coordinate vectors to extract the horizontal and vertical offsets, and perform translation mapping operation on the two-dimensional coordinate vectors according to the horizontal and vertical offsets to generate a pixel position reconstruction matrix. S103: Traverse the pixel positions to reconstruct the grayscale components within the matrix, accumulate the grayscale components and divide by the total number of pixels to obtain the average grayscale value, divide the average grayscale value by the color reference value to calculate the scaling ratio, multiply the scaling ratio with the grayscale components, extract texture features and stitch them together to establish a preprocessed image feature set.

4. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the preprocessed image feature set into the encoder of the scene semantic segmentation model to perform convolution to extract spatial activation components, multiply the channel weight values ​​with the planar spatial weights to calculate the joint attention bias term, and perform dot product between the joint attention bias term and the spatial activation components to establish a spatial context feature tensor. S202: Calculate the squared difference of the activation distribution values ​​within the spatial context feature tensor from the mean, accumulate the squared differences to obtain the feature distribution variance, and perform proportional scaling and rounding down on the preset reference spacing according to the ratio between the feature distribution variance and the preset discrete threshold to obtain the adaptive sampling spacing parameter. S203: Adjust the dilation rate of the convolution kernel within the decoder according to the adaptive sampling interval parameter, input the spatial context feature tensor into the decoder to perform step sampling according to the dilation rate, extract semantic representation vectors at each resolution scale and perform concatenation operation to generate an environmental semantic feature matrix.

5. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Read the corresponding feature vector in the environmental semantic feature matrix, read the category reference template, calculate the similarity between the feature vector and the occlusion and sky vectors in the category reference template, filter the pixels with similarity greater than the preset judgment threshold and perform category assignment, and establish a scene pixel category set. S302: Traverse the neighborhood window of the grid nodes in the pixel category set of the scene, read the category labels of the center node and the adjacent nodes, calculate the absolute value of the difference between the two labels, retrieve the center node corresponding to the absolute value greater than zero as the edge mutation point, read the plane grid coordinates of the mutation point, and obtain the boundary coordinate dataset. S303: Extract the horizontal coordinate values ​​corresponding to the grid coordinates in the boundary coordinate dataset, perform sorting operation according to the incrementing rule, read the corresponding vertical coordinate values ​​according to the sorting order, associate the horizontal coordinate values ​​with the vertical coordinate values, and obtain the skyline coordinate sequence.

6. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 5, characterized in that, The preset judgment threshold is the upper limit of the interval constructed by the median and standard deviation of the similarity distribution interval calculated based on the feature vector and the occlusion vector and sky vector in the category reference template, respectively. The filter is the pixel point with a similarity greater than the preset judgment threshold and the category assignment is performed by calculating the similarity between the feature vector of each pixel point and the occlusion vector and sky vector in the category reference template, respectively, and selecting the category label corresponding to the maximum similarity as the pixel point category label.

7. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Read the spatial pixel grid points in the skyline coordinate sequence, extract the vertical coordinates of the spatial pixels, calculate the absolute difference between the longitudinal coordinates of the image reference baseline, construct the grid single-point offset based on the absolute difference and merge them in horizontal spatial order to obtain the coordinate offset value set. S402: Call the coordinate offset value set, divide the camera's vertical field of view by the total number of pixel rows in the vertical direction of the image to calculate the single pixel angular resolution, multiply the single-point offset of each grid by the single pixel angular resolution to obtain the actual elevation angle transformation, integrate the actual elevation angle transformation, and generate the occlusion elevation angle distribution matrix. S403: For the aforementioned occlusion elevation angle distribution matrix, calculate the horizontal azimuth angle based on the horizontal index of the grid nodes and the horizontal angular resolution. Add the horizontal azimuth angle to the initial azimuth parameter of the geographic north direction to calibrate the absolute azimuth angle. Perform key-value mapping between the actual elevation angle transformation and the absolute azimuth angle to establish site environment assessment data.

8. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 1, characterized in that, The method further includes: S5: Call the site environment assessment data, transmit the site environment assessment data to the central server node through the site communication terminal, compare and calculate the site environment assessment data with the preset airspace threshold, and integrate and construct a dynamic digital assessment file based on the calculation results; The dynamic digital assessment archive includes the site's airspace compliance status, airspace deviation values ​​in each direction, and historical assessment records.

9. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 8, characterized in that, The specific steps of S5 are as follows: S501: Extract the original payload sequence from the site environment assessment data, modulate the original payload sequence into a continuous analog waveform according to the baseband modulation mapping table and transmit it to the central server node, and extract the digital payload unit by analyzing the analog-to-digital conversion of the continuous analog waveform according to the demodulation parameters, and generate the restored assessment data matrix. S502: Call the restored evaluation data matrix to extract the actual elevation angle measurement value, subtract the actual elevation angle measurement value from the preset clearance threshold parameter to obtain the relative space difference, extract the grid node coordinates for reference values ​​greater than zero, and assemble the grid node coordinates and the relative space difference to obtain the set of out-of-limit deviation indicators. S503: Read the absolute azimuth angle according to the set of out-of-limit deviation indicators, write it into the local node structure tree template identifier area, fill the relative spatial difference into the second-level node of the local node structure tree template, perform encoding merging on the nodes in the local node structure tree template, and establish a dynamic digital evaluation file.

10. The automated assessment method for the airspace environment of meteorological observation stations based on deep learning according to claim 9, characterized in that, The preset airspace threshold parameter is a fixed upper limit value of an interval determined according to the statistical distribution of historical elevation angle measurements in the site environmental assessment data. The relative spatial difference is the retained value after non-negative filtering in the difference sequence calculated point by point between the actual elevation angle measurement and the preset airspace threshold parameter.

Images