Visibility analysis method based on satellite observation and continuous images

Abstract

The invention provides a visibility analysis method based on satellite observation and continuous images, and the method comprises the steps: obtaining the continuous image data of a fixed camera, the hourly visibility observation data of a ground monitoring station, and the satellite inversion visibility data, and carrying out the multi-source data space dynamic fusion based on the above data. An appropriate interpolation algorithm is used according to different data sources, the contribution weight of each data source is dynamically adjusted based on the data distribution density, it is ensured that the model has the accurate prediction capacity in space, then model building, training and fine adjustment are conducted, visibility recognition of continuous images is finally achieved, and in the model application stage, the visibility recognition efficiency is improved. The visibility recognition is carried out by using real-time images of the camera instead of depending on ground monitoring station data and satellite data. According to the method, a time weighting mechanism is introduced, camera images at continuous moments can be fully utilized, the stability and precision of visibility prediction are improved, and the situation that noise or abnormal values at a single moment have great influence on a result is avoided.

Description

Technical Field

[0001] The invention relates to the technical field of atmosphere detection, and in particular to a visibility analysis method based on satellite observation and continuous images. Background Art

[0002] Visibility refers to the maximum distance at which a person with normal vision can identify a target in a specific background environment. Among the many influencing factors, various weather phenomena such as fog, haze, sand, smoke, dust, snow, rain, etc. will have a very significant impact on it. Especially local weather phenomena, such as sand, fog and storms, are more likely to cause a sharp drop in visibility, thus causing low visibility.

[0003] Visibility plays a vital role in urban air pollution control. In cities, good visibility helps people to detect the degree of air pollution in time and take corresponding prevention and control measures. At the same time, it is also closely related to public transportation safety. Whether it is vehicles on the road or rail transit in the city, sufficient visibility is required to ensure driving safety. In ocean terminals and waterways, traffic safety is closely related to visibility. In the case of low visibility, the navigation of ships faces huge risks and is likely to cause serious collision accidents. Low visibility is very likely to cause serious safety accidents, posing a huge threat to people's lives and property safety. Once an accident occurs, it will not only cause casualties, but also bring a heavy blow to the family, and also cause significant losses to the social economy. In view of this, it is very important to accurately classify atmospheric visibility and implement refined monitoring.

[0004] At present, in my country's meteorological departments, the mainstream visibility observation and identification methods are mainly divided into manual observation and instrument measurement. Manual observation is performed by observers at several specific times of the day, relying on visual observation of the clarity of targets at different distances to determine the visibility value. The manual observation visibility mentioned here usually refers to the effective horizontal visibility, that is, the maximum horizontal distance of the target that can be seen in more than half of the surrounding field of vision. However, manual observation has many defects. First, the number of observations is small, and it is impossible to grasp the changes in visibility in real time. Secondly, the observation results are greatly affected by the subjective factors of the observer, and different observers may get different results. Finally, the accuracy is relatively low, which is difficult to meet the needs of modern meteorological observation.

[0005] The instrumental measurement method uses optical equipment to measure specific optical properties of the atmosphere, such as extinction coefficient and transmittance, and calculates the visibility value. Common instruments include forward scattering visibility meters, single-ended projection portable measurement systems and other visibility detection instruments. The results obtained by the instrumental measurement method are relatively accurate. These instruments can accurately measure the optical properties in the atmosphere and calculate more accurate visibility values. However, such instruments are expensive in design and production, which may put some meteorological departments under great economic pressure. Moreover, these instruments require special installation and calibration, which not only requires professional technicians, but also takes a lot of time and effort. In addition, the higher the cost, the less flexible it is. Once installed in a fixed location, it is difficult to move and adjust.

[0006] In addition, in recent years, some studies have attempted to apply machine learning methods to perform visibility inversion based on images. This method uses computer technology to analyze and process a large number of images to obtain visibility values. However, this method still has some problems. For example, the quality of image acquisition is unstable and may be affected by factors such as lighting and weather. The image data is not synchronized with the environmental monitoring data, which affects the accuracy of the inversion results to a certain extent. Under low visibility conditions, the recognition effect is also unsatisfactory, and it is difficult to accurately determine the visibility value. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a visibility analysis method based on satellite observation and continuous images. By constructing a visibility analysis system with strong generalization ability based on vision and satellite observation capabilities, high-precision visibility observation is achieved.

[0008] To achieve the above object, the technical solution adopted by the present invention is:

[0009] A visibility analysis method based on satellite observation and continuous images, characterized in that it comprises the following steps:

[0010] S1. Data preparation:

[0011] Build a system database to collect continuous image data from fixed cameras, hourly visibility observation data from ground monitoring stations, and satellite inversion visibility data.

[0012] S11, fixed camera continuous image data:

[0013] The continuous image data of fixed cameras is provided by fixed cameras. The data of fixed cameras comes from traffic cameras, urban security cameras or special cameras deployed by other meteorological departments. The cameras continuously collect images at fixed intervals to form a time-series image data set.

[0014] Data preprocessing for the time series image dataset:

[0015] S111, Image size unification:

[0016] All images are scaled to a uniform size, and the image size is uniformly interpolated using bilinear interpolation or cubic spline interpolation algorithms.

[0017] I resize = resize(I original , 800, 600)

[0018] Among them I resize is the processed image, I original is the original image;

[0019] S112, color channel standardization:

[0020] In order to meet the input requirements of the neural network, the three RGB color channels of the image need to be normalized separately. The normalization formula is as follows:

[0021]

[0022] Where I(x, y, c) is the value of the color channel c of the pixel (x, y), μ c and μ c are the mean and standard deviation of color channel c, respectively.

[0023] S113, denoising and enhancement:

[0024] Use median filtering or bilateral filtering to denoise the image:

[0025] I denoise =medianFilter(I norm , k)

[0026] Where k is the filter window size,

[0027] The image is enhanced by contrast stretching and brightness adjustment.

[0028] S114, Data Enhancement:

[0029] The image data is enhanced by random rotation, cropping, and flipping:

[0030] I aug =randomAugment(I denoise )

[0031] The data-augmented images will be used together with the original images for visibility regression model training.

[0032] S115, Time Alignment:

[0033] Time alignment is performed through interpolation algorithms to ensure that the timestamp of the image is consistent with the hourly visibility observation data of the ground monitoring station and the visibility data inverted by the satellite. If the camera image is missing at a certain moment, it can be filled by interpolating the images at adjacent moments:

[0034]

[0035] This ensures that there is continuous image data at all time points for subsequent analysis;

[0036] S12. Hourly visibility observation data of ground monitoring stations:

[0037] The hourly visibility observation data of ground monitoring stations are provided by the meteorological department. The meteorological department monitors visibility through ground monitoring stations in various cities. The data of ground monitoring stations include hourly visibility data and meteorological data. The meteorological data is used to assist in analyzing visibility changes.

[0038] Data preprocessing for hourly visibility observation data:

[0039] S121, Time Alignment:

[0040] Temporally interpolate the hourly visibility observation data to make it consistent with the camera image data collected by the fixed camera:

[0041]

[0042] Where V(t) is the visibility value at time t, t 0 and t 1 For two consecutive hourly moments,

[0043] S122. Missing value processing:

[0044] If the data of ground monitoring stations are missing at a certain moment, they can be supplemented by interpolation or using interpolation methods based on historical data;

[0045] S13. Satellite inversion visibility data:

[0046] The satellite-derived visibility data is obtained through polar-orbiting satellites or geostationary satellites, and its resolution depends on the corresponding satellite system.

[0047] S131. Remove outliers:

[0048] In the preprocessing stage, outliers in the satellite-retrieved visibility data are eliminated, and error data are detected through upper and lower limit constraints and statistical methods;

[0049] S132, Time Alignment:

[0050] The time gaps in the satellite-inverted visibility data are filled by linear interpolation to ensure that the camera image data, the hourly visibility observation data of the ground monitoring station and the satellite-inverted visibility data correspond to each other at the same time. S14. All processed data will be stored in the system database and classified and managed by region and time.

[0051] Data management includes: storage of raw data and storage of processed data.

[0052] Each processed camera image data, hourly visibility observation data and satellite-derived visibility data will be time-stamped and spatially labeled to ensure consistency with the input of the visibility regression model;

[0053] S2. Dynamic fusion of multi-source data space:

[0054] By introducing the mesh model, the three types of data, namely camera image data, hourly visibility observation data and satellite inversion visibility data, are spatially fused and weighted as follows:

[0055] The corresponding interpolation algorithm is used according to different data sources, and the contribution weight of each data source is dynamically adjusted based on the data distribution density to ensure that the visibility regression model has accurate prediction capabilities in space.

[0056] Based on the highest resolution grid of satellite inversion visibility data, the target area is gridded in geographic space. The grid division is based on satellite resolution. The grid where the fixed camera is located is used as the core grid, and eight surrounding grids are expanded outward to form a 3x3 nine-square grid structure.

[0057] Among these nine grids, the data of the core grid is integrated with fixed cameras, ground monitoring stations and satellite systems, while the peripheral grids are used for interpolation of satellite data and ground station data.

[0058] The spatial data fusion method is adjusted according to the distribution of ground monitoring stations within the grid, and is divided into four cases:

[0059] Case 1: No site

[0060] There are no ground monitoring stations in the core grid and the eight surrounding grids. Due to the lack of ground monitoring station data, the visibility of the core grid is completely dependent on the satellite system. The calculation process does not require an interpolation algorithm. The visibility value is determined only based on the observation data of the satellite system. The data weight distribution is as follows:

[0061] Satellite data weight within the core grid: ω sat =1

[0062] Ground station data weight: ωsite =0

[0063] Core grid visibility algorithm formula:

[0064] V core =V sat (core)

[0065] Among them, V core is the visibility of the core grid after data fusion, V sat (core) is the satellite observation visibility data within the core grid,

[0066] Case 2: Clustered Sites

[0067] There is at least one ground monitoring station in the core grid, but no station in the surrounding eight grids. In this case, it is necessary to combine the station data in the core grid with the satellite data.

[0068] If there is only one ground monitoring site in the core grid, the data of the ground monitoring site is directly used as the site visibility V of the core grid. site (core),

[0069] If the core grid has more than one ground monitoring station, interpolation processing is required between the ground monitoring stations:

[0070]

[0071] Among them, V(x i ) represents the visibility observation value of site i, d(x 0 , x i ) is the camera position x 0 The distance from site i, p is the distance attenuation factor,

[0072] Data weight distribution:

[0073] Satellite data weight within the core grid: ω sat =0.5

[0074] Data weight of ground monitoring stations within the core grid: ω site =0.5

[0075] Taking into account the satellite data and site data of the core grid, the visibility calculation formula of the core grid is:

[0076] V core =ω site ×V site (core)+ω sat ×V sat (core)

[0077] Case 3: Sparse sites

[0078] There is no ground monitoring station in the core grid, but there is at least one ground monitoring station in the surrounding 8 grids. In this case, the visibility of the core grid needs to be interpolated using the data of the surrounding ground monitoring stations.

[0079] If there is only one ground monitoring station in the surrounding grid, the data of the ground monitoring station is directly used as the ground monitoring station visibility V of the surrounding grid. site (out),

[0080] If there are multiple ground monitoring stations in the surrounding grid, use the Kriging interpolation method to interpolate the ground monitoring station data of the surrounding grid into the core grid:

[0081]

[0082] Among them, V(x i ) is the visibility observation value of the surrounding ground monitoring station i, λ i is the interpolation weight, determined based on geographic spatial relevance, satisfying

[0083] Data weight distribution:

[0084] Satellite data weight of the core grid: ω sat =0.8

[0085] Ground monitoring station data weight: ω site =0.2

[0086] The interpolated ground monitoring station data is combined with the satellite data of the core grid, and the visibility calculation formula of the core grid is:

[0087] V core =ω site ×V site (out)+ω sat ×V sat (core)

[0088] Case 4: Multisite

[0089] There are ground monitoring stations in the core grid and the 8 surrounding grids. At this time, the ground monitoring station data of the core grid and the surrounding grids need to be processed separately. The ground monitoring station data in the core grid is interpolated using the inverse distance weighted method, and the ground monitoring station data in the surrounding grid is interpolated to the core grid using the Kriging interpolation method. The interpolation algorithm for the ground monitoring stations in the core grid is as follows:

[0090]

[0091] The interpolation algorithm for the ground monitoring sites of the surrounding grid is:

[0092]

[0093] Data weight distribution:

[0094] The ground monitoring station data weight of the core grid: ω site (core) = 0.6

[0095] Satellite data weight of the core grid: ω sat =0.3

[0096] The data weight of the ground monitoring stations in the surrounding grid: ω site (out) = 0.1

[0097] Combining the interpolation data of ground monitoring stations of the core and surrounding grids, and the satellite data of the core grid, the visibility calculation formula of the core grid is:

[0098] V core =ω site (core)×V site (core)+ω sat ×V sat (core)+ω site (out)×V site (out);

[0099] S3. Visibility regression model training:

[0100] S31. Visual image feature extraction:

[0101] The backbone network is used to extract features from visual images taken by fixed cameras to obtain high-dimensional features of the images. The backbone network is used to extract features from satellite data acquired by the satellite system to obtain high-dimensional features of the satellite data. The satellite data is standardized to obtain satellite values.

[0102] S32. Establish a fully connected regression model:

[0103] A fully connected neural network regression model is established by combining the high-dimensional features of the image extracted from the visual image and the high-latitude features or satellite values ​​extracted from the satellite data with the visibility label data;

[0104] S33, model training: use multi-region training samples with long time scales to perform model training;

[0105] S34, Model fine-tuning: In model training, the model trained with samples from multiple locations and long time scales needs to be fine-tuned when used in the field. In the model application, a model fine-tuning function is set to achieve fine-tuning of the model;

[0106] S4. Visibility identification:

[0107] In the application stage of the visibility regression model, we no longer rely on ground monitoring station data and satellite inversion visibility data, but use real-time images from fixed cameras for visibility recognition.

[0108] By introducing a time weighting mechanism, we can make full use of the real-time images of fixed cameras at consecutive moments.

[0109] For the target time t, the image data of 2n+1 moments before and after are extracted from the image sequence taken by the fixed camera, and visibility recognition is performed at each moment.

[0110] Then, different weights are assigned to each moment based on the time difference between the image capture moment and the target moment;

[0111] Finally, the visibility value at time t is obtained by weighted calculation method.

[0112]

[0113] Where V(t+i) is the estimated visibility at time t+i; ω i is a weight inversely proportional to the time difference |i|. The smaller the time difference, the greater the weight.

[0114] Weight explanation: The weight ω corresponding to the current time t 0 Maximum, ensuring that the visibility regression model is dominated by the observation value at the current moment; the data before and after the moment are used as auxiliary, and the weight ω -2 ,ω -1 ,ω 1 ,ω 2 It gradually decreases with the increase of time distance.

[0115] As a preferred technical solution of the present invention: in step S11, the data collection interval of the fixed camera is every 5 minutes or 10 minutes.

[0116] As a preferred technical solution of the present invention: in step S12, the monitoring site data also includes meteorological data, and the meteorological data is used to assist in analyzing visibility changes.

[0117] As a preferred technical solution of the present invention: in step S13, the resolution range of the satellite inverted visibility data is one hundred meters to nine kilometers.

[0118] As a preferred technical solution of the present invention: in step S33, data enhancement is achieved by scaling and transposing the visual image and satellite data.

[0119] As a preferred technical solution of the present invention: in step S34, the model fine-tuning function uses manual visual method or mobile visibility monitoring equipment to measure the visibility value of the model application location, accumulates certain data to fine-tune the basic model, so as to optimize the model effect.

[0120] Compared with the prior art, the present invention has the following beneficial effects:

[0121] 1. The present invention has higher-precision visibility observation: by constructing a visibility recognition system based on satellites and visual images and combining different observation methods with artificial intelligence methods, the accuracy of visibility observation can be improved.

[0122] 2. The present invention has a wider range of visibility observation: the automated observation of visibility is no longer limited to expensive observation equipment with certain construction requirements. Satellite data has global coverage, and combined with a large number of camera resources, visibility observation can be completed more widely.

[0123] 3. The present invention has the ability to improve forecasting capabilities: For weather forecasting technology, better restoration of the current state of the atmosphere is a key factor in improving the accuracy of deducing future weather trends. High-precision, high-density visibility observations can better help restore the current state of the atmosphere, thereby improving weather forecasting capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] Figure 1 It is a schematic diagram of the structure of the present invention;

[0125] Figure 2 It is the principle diagram of neural network regression model;

[0126] Figure 3 It is a real-time meteorological collection image;

[0127] Figure 4 Satellites acquire images. DETAILED DESCRIPTION

[0128] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0129] like Figure 1-2 As shown, the present invention proposes a visibility analysis method based on satellite observation and continuous images, comprising the following steps:

[0130] S1. Data preparation:

[0131] Build a system database to collect continuous image data from fixed cameras, hourly visibility observation data from ground monitoring stations, and satellite inversion visibility data.

[0132] S11, fixed camera continuous image data:

[0133] like Figure 3 As shown in FIG, the continuous image data of the fixed camera is provided by the fixed camera. The data of the fixed camera comes from traffic cameras, urban security cameras or special cameras deployed by other meteorological departments. The camera continuously collects images at fixed intervals to form a time series image data set.

[0134] The data characteristics of these images include:

[0135] Continuous moment capture: Images at adjacent moments have a high correlation, which provides a basis for subsequent time-weighted processing;

[0136] Resolution differences: The resolutions of different cameras may vary. Common image resolutions range from 640x480 to 1920x1080 pixels.

[0137] Different fields of view and angles: The shooting angles and fields of view of different cameras may be different, so they need to be uniformly processed during preprocessing.

[0138] Before model training, the collected continuous images must be preprocessed in various ways to ensure the consistency of the image data and suitability for deep learning model training. The specific processing methods are as follows:

[0139] Data preprocessing for the time series image dataset:

[0140] S111, Image size unification:

[0141] All images are scaled to a uniform size, and bilinear interpolation or cubic spline interpolation algorithms are used to ensure that the image remains relatively clear after scaling.

[0142] I resize = resize(I original , 800, 600)

[0143] Among them I resize is the processed image, I original is the original image;

[0144] S112, color channel standardization:

[0145] In order to meet the input requirements of the neural network, the three RGB color channels of the image need to be normalized separately. The normalization formula is as follows:

[0146]

[0147] Where I(x, y, c) is the value of the color channel c of the pixel (x, y), μ c and σ care the mean and standard deviation of color channel c, respectively.

[0148] S113, denoising and enhancement:

[0149] Image denoising is to remove the influence of sensor noise and environmental interference on the image. Median filtering or bilateral filtering is usually used to denoise the image:

[0150] I denoise =medianFilter(I norm , k)

[0151] Where k is the filter window size,

[0152] At the same time, considering that some images may be affected by factors such as haze and insufficient light, image enhancement methods such as contrast stretching and brightness adjustment can be used to improve image clarity.

[0153] S114, Data Enhancement:

[0154] Data augmentation helps improve the generalization ability of the model and avoid overfitting. Data augmentation is performed on images by randomly rotating, cropping, and flipping:

[0155] I aug =randomAugment(I denoise )

[0156] The augmented images are used together with the original images to train the visibility regression model, ensuring that the model can adapt to different camera angles, lighting, and weather conditions.

[0157] S115, Time Alignment:

[0158] Since the data from fixed cameras are collected at fixed time intervals, it is necessary to ensure that the timestamp of the image is consistent with the time of the ground monitoring station and satellite data. Time alignment is performed through the interpolation algorithm to ensure that the timestamp of the image is consistent with the hourly visibility observation data of the ground monitoring station and the satellite inversion visibility data. If the camera image is missing at a certain moment, it can be filled by interpolating the images at adjacent moments:

[0159]

[0160] This ensures that there is continuous image data at all time points for subsequent analysis;

[0161] S12. Hourly visibility observation data of ground monitoring stations are shown in the following table:

[0162] Current status of 54416 (Miyun) station

[0163] Meteorological elements value Update time Instantaneous temperature 19.0 2024-11-01 12:00+0800 24 hours temperature change 1.9 2024-11-01 12:00+0800 Ground pressure 1015.3 2024-11-01 12:00+0800 Relative humidity 53 2024-11-01 12:00+0800 2-minute average wind direction (NE) 2024-11-01 12:00+0800 2-minute average wind speed (2) 2024-11-01 12:00+0800 1 hour precipitation 0.0 2024-11-01 12:00+0800 24-hour precipitation 0.0 2024-11-01 12:00+0800 10-minute average visibility 6.348 2024-11-01 12:00+0800

[0164] The hourly visibility observation data of ground monitoring stations are provided by the meteorological department. The meteorological department monitors visibility through ground monitoring stations in various cities. The data of ground monitoring stations include hourly visibility data and meteorological data. The meteorological data is used to assist in analyzing visibility changes.

[0165] Data preprocessing for hourly visibility observation data:

[0166] S121, Time Alignment:

[0167] Time synchronization with fixed camera images and satellite data is very important. Ground monitoring station data is usually hourly, but fixed camera data may be collected on a minute-by-minute basis. Therefore, the hourly visibility observation data is temporally interpolated to make it consistent with the camera image data collected by the fixed camera:

[0168]

[0169] Where V(t) is the visibility value at time t, t 0 and t 1 For two consecutive hourly moments,

[0170] S122. Missing value processing:

[0171] If the data of ground monitoring stations are missing at a certain moment, they can be supplemented by interpolation or using interpolation methods based on historical data;

[0172] S13. Satellite inversion visibility data:

[0173] Satellite-derived visibility data is obtained through polar-orbiting satellites or geostationary satellites. Figure 4 As shown in the figure, its resolution depends on the corresponding satellite system. Satellite data can provide wide-area visibility information, making up for the sparse distribution of ground stations.

[0174] S131. Remove outliers:

[0175] Since satellite observations may be affected by clouds, sensor failures, etc., there may be certain errors. Therefore, in the preprocessing stage, outliers in the satellite inversion visibility data are removed, and error data are detected through upper and lower limit constraints and statistical methods;

[0176] S132, Time Alignment:

[0177] Satellite data is usually acquired at long intervals, so linear interpolation or more complex interpolation methods are needed to fill in the time gaps to ensure that camera images, monitoring site data, and satellite data correspond to each other at the same time.

[0178] The time gaps in the satellite-inverted visibility data are filled by linear interpolation to ensure that the camera image data, the hourly visibility observation data of the ground monitoring station and the satellite-inverted visibility data correspond to each other at the same time. S14. All processed data will be stored in the system database and classified and managed by region and time.

[0179] Data management includes: storage of raw data and storage of processed data.

[0180] Each processed camera image data, hourly visibility observation data and satellite-derived visibility data will be time-stamped and spatially labeled to ensure consistency with the input of the visibility regression model;

[0181] S2. Dynamic fusion of multi-source data space:

[0182] Multi-source data spatial fusion is one of the key steps in visibility analysis system.

[0183] By introducing the mesh model, the three types of data, namely camera image data, hourly visibility observation data and satellite inversion visibility data, are spatially fused and weighted as follows:

[0184] The corresponding interpolation algorithm is used according to different data sources, and the contribution weight of each data source is dynamically adjusted based on the data distribution density to ensure that the visibility regression model has accurate prediction capabilities in space.

[0185] Based on the highest resolution grid of satellite inversion visibility data, the target area is gridded in geographic space. The grid division is based on satellite resolution. The grid where the fixed camera is located is used as the core grid, and eight surrounding grids are expanded outward to form a 3x3 nine-square grid structure.

[0186] Among these nine grids, the data of the core grid is integrated with fixed cameras, ground monitoring stations and satellite systems, while the peripheral grids are used for interpolation of satellite data and ground station data.

[0187] The spatial data fusion method is adjusted according to the distribution of ground monitoring stations within the grid, and is divided into four cases:

[0188] Case 1: No site

[0189] There are no ground monitoring stations in the core grid and the eight surrounding grids. Due to the lack of ground monitoring station data, the visibility of the core grid is completely dependent on the satellite system. The calculation process does not require an interpolation algorithm. The visibility value is determined only based on the observation data of the satellite system. The data weight distribution is as follows:

[0190] Satellite data weight within the core grid: ωsat =1

[0191] Ground station data weight: ω site =0

[0192] Core grid visibility algorithm formula:

[0193] V core =V sat (core)

[0194] Among them, V core is the visibility of the core grid after data fusion, V sat (core) is the satellite observation visibility data within the core grid,

[0195] Case 2: Clustered Sites

[0196] There is at least one ground monitoring station in the core grid, but no station in the surrounding eight grids. In this case, it is necessary to combine the station data in the core grid with the satellite data.

[0197] If there is only one ground monitoring site in the core grid, the data of the ground monitoring site is directly used as the site visibility V of the core grid. site (core),

[0198] If the core grid has more than one ground monitoring station, interpolation processing is required between the ground monitoring stations:

[0199]

[0200] Among them, V(x i ) represents the visibility observation value of site i, d(x 0 , x i ) is the camera position x 0 The distance from site i, p is the distance attenuation factor,

[0201] Data weight distribution:

[0202] Satellite data weight within the core grid: ω sat =0.5

[0203] Data weight of ground monitoring stations within the core grid: ω site =0.5

[0204] Taking into account the satellite data and site data of the core grid, the visibility calculation formula of the core grid is:

[0205] V core =ω site ×V site (core)+ω sat ×Vsat (core)

[0206] Case 3: Sparse sites

[0207] There is no ground monitoring station in the core grid, but there is at least one ground monitoring station in the surrounding 8 grids. In this case, the visibility of the core grid needs to be interpolated using the data of the surrounding ground monitoring stations.

[0208] If there is only one ground monitoring station in the surrounding grid, the data of the ground monitoring station is directly used as the ground monitoring station visibility V of the surrounding grid. site (out),

[0209] If there are multiple ground monitoring stations in the surrounding grid, use the Kriging interpolation method to interpolate the ground monitoring station data of the surrounding grid into the core grid:

[0210]

[0211] Among them, V(x i ) is the visibility observation value of the surrounding ground monitoring station i, λ i is the interpolation weight, determined based on geographic spatial relevance, satisfying

[0212] Data weight distribution:

[0213] Satellite data weight of the core grid: ω sat =0.8

[0214] Ground monitoring station data weight: ω site =0.2

[0215] The interpolated ground monitoring station data is combined with the satellite data of the core grid, and the visibility calculation formula of the core grid is:

[0216] V core =ω site ×V site (out)+ω sat ×V sat (core)

[0217] Case 4: Multisite

[0218] There are ground monitoring stations in the core grid and the 8 surrounding grids. At this time, the ground monitoring station data of the core grid and the surrounding grids need to be processed separately. The ground monitoring station data in the core grid is interpolated using the inverse distance weighted method, and the ground monitoring station data in the surrounding grid is interpolated to the core grid using the Kriging interpolation method. The interpolation algorithm for the ground monitoring stations in the core grid is as follows:

[0219]

[0220] The interpolation algorithm for the ground monitoring sites of the surrounding grid is:

[0221]

[0222] Data weight distribution:

[0223] The ground monitoring station data weight of the core grid: ω site (core) = 0.6

[0224] Satellite data weight of the core grid: ω sat =0.3

[0225] The data weight of the ground monitoring stations in the surrounding grid: ω site (out) = 0.1

[0226] Combining the interpolation data of ground monitoring stations of the core and surrounding grids, and the satellite data of the core grid, the visibility calculation formula of the core grid is:

[0227] V cor =ω site (core)×V site (core)+ω sat ×V sat (core)+ω site (out)×V site (out);

[0228] S3. Visibility regression model training:

[0229] S31. Visual image feature extraction:

[0230] The backbone network is used to extract features from visual images taken by fixed cameras to obtain high-dimensional features of the images. The backbone network is used to extract features from satellite data acquired by the satellite system to obtain high-dimensional features of the satellite data. The satellite data is standardized to obtain satellite values.

[0231] You can use CNN, Swin-Transformer, ViT, DeiT and other backbone networks to extract features of visual images. The backbone network here is pluggable. With the continuous changes in neural network technology, the neural network used to extract visual image features in the system can also be replaced at any time.

[0232] S32. Establish a fully connected regression model:

[0233] A fully connected neural network regression model is established by combining the high-dimensional features of the image extracted from the visual image and the high-latitude features or satellite values ​​extracted from the satellite data with the visibility label data;

[0234] S33, model training: use multi-region training samples with long time scales to perform model training;

[0235] S34, Model fine-tuning: In model training, the model trained with samples from multiple locations and long time scales needs to be fine-tuned when used in the field. In the model application, a model fine-tuning function is set to achieve fine-tuning of the model;

[0236] S4. Visibility identification:

[0237] In the application stage of the visibility regression model, we no longer rely on ground monitoring station data and satellite inversion visibility data, but use real-time images from fixed cameras for visibility recognition.

[0238] By introducing a time weighting mechanism, we can make full use of the real-time images of fixed cameras at consecutive moments.

[0239] Different from traditional single-image recognition, this method can make full use of camera images at consecutive moments by introducing a time-weighted mechanism, improve the stability and accuracy of visibility prediction, and avoid the significant impact of noise or outliers at a single moment on the results.

[0240] For the target time t, the image data of 2n+1 moments before and after are extracted from the image sequence taken by the fixed camera, and visibility recognition is performed at each moment.

[0241] For example, when n=2, extract the images at five times: t-2, t-1, t, t+1, and t+2, and perform visibility recognition at each time.

[0242] Then, different weights are assigned to each moment based on the time difference between the image capture moment and the target moment;

[0243] Finally, the visibility value at time t is obtained by weighted calculation method.

[0244]

[0245] Where V(t+i) is the estimated visibility at time t+i; ω i is a weight inversely proportional to the time difference |i|. The smaller the time difference, the greater the weight.

[0246] Weight explanation: The weight ω corresponding to the current time t 0 Maximum, ensuring that the visibility regression model is dominated by the observation value at the current moment; the data before and after the moment are used as auxiliary, and the weight ω -2 ,ω -1 ,ω 1 ,ω 2As the time distance increases, it gradually decreases.

[0247] For example, if the weights are calculated according to ω 0 =0.4,ω -1 =ω 1 =0.2,ω -2 =ω 2 =0.1, the calculation formula is:

[0248] V(t)=0.4×V(t)+0.2×(V(t-1)+V(t+1))+0.1×(V(t-2)+V(t+2))

[0249] This time-weighted processing can fully refer to the image data at adjacent moments based on the use of the current moment image, thereby reducing the impact of noise or instantaneous fluctuations in a single-frame image on the result and obtaining a more stable visibility estimation result.

[0250] In step S11, the data collection interval of the fixed camera is every 5 minutes or 10 minutes.

[0251] In step S12, the monitoring site data also includes meteorological data, which is used to assist in analyzing visibility changes.

[0252] In step S13, the resolution range of the satellite-inverted visibility data is from one hundred meters to nine kilometers.

[0253] In step S33, data enhancement is achieved by scaling and transposing the visual image and satellite data.

[0254] In step S34, the model fine-tuning function uses manual visual methods or mobile visibility monitoring equipment to measure the visibility value of the model application location, accumulates certain data, and fine-tunes the basic model to optimize the model effect.

[0255] The visibility analysis method based on satellite observation and continuous images proposed in the present invention is realized based on satellite observation and continuous shooting at fixed intervals by a fixed camera. By acquiring continuous image data of a fixed camera, hourly visibility observation data of a ground monitoring station and satellite inversion visibility data, and storing them in a system database after preprocessing, the data are classified by region and time. Not only the original data is stored and retained for reprocessing during subsequent algorithm optimization, but also each processed image data of a fixed camera, visibility observation data of a ground monitoring station and satellite inversion visibility data are provided with a timestamp and a spatial label to ensure consistency with the input of the model.

[0256] After the data is processed and stored, multi-source data is dynamically fused in space. Multi-source data spatial fusion is one of the key steps in the visibility analysis method. By introducing a meshed 3x3 grid model, the three types of data from fixed cameras, ground monitoring sites and satellite observations are spatially fused and weighted. This process requires the use of appropriate interpolation algorithms according to different data sources, and dynamically adjusts the contribution weight of each data source based on the data distribution density to ensure that the visibility model has accurate prediction capabilities in space.

[0257] After achieving dynamic fusion of multi-distance data space, the model is built, trained and fine-tuned to make the method more adaptable. In actual use, a certain degree of fine-tuning is required. The visibility value of the model application location can be measured by manual visual method or mobile visibility monitoring equipment, and a certain amount of data can be accumulated to fine-tune the basic model to optimize the model effect.

[0258] The model can be used after it is completed. In the model application stage, it no longer relies on ground monitoring site data and satellite data, but uses real-time images from fixed cameras for visibility recognition. Unlike traditional single image recognition, this method can make full use of camera images at continuous moments by introducing a time weighting mechanism, improve the stability and accuracy of visibility prediction, and avoid the noise or outliers at a single moment from having a significant impact on the results. For the target moment t, the system extracts image data of 2n+1 moments before and after from the fixed camera image sequence. For example, when n=2, images of five moments t-2, t-1, t, t+1, and t+2 are extracted, and visibility recognition is performed at each moment. Then, based on the time difference between the fixed camera image shooting moment and the target moment, different weights are assigned to each moment. Finally, the visibility value at moment t is obtained by weighted calculation.

[0259]

[0260] Where V(t+i) is the estimated visibility at time t+i; ω i is a weight inversely proportional to the time difference |i|. The smaller the time difference, the greater the weight.

[0261] Weight explanation: The weight ω corresponding to the current time t 0 Maximum, ensuring that the model is dominated by the observations at the current moment; the data before and after the moment are used as auxiliary, and the weight ω -2 ,ω -1 ,ω 1 ,ω 2 It gradually decreases with the increase of time distance.

[0262] For example, if the weights are calculated according to ω 0 =0.4,ω-1 =ω 1 =0.2,ω -2 =ω 2 =0.1, the calculation formula is:

[0263] V(t)=0.4×V(t)+0.2×(V(t-1)+V(t+1))+0.1×(V(t-2)+V(t+2))

[0264] This time-weighted processing can fully refer to the image data at adjacent moments based on the use of the current moment image, thereby reducing the impact of noise or instantaneous fluctuations in a single-frame image on the result and obtaining a more stable visibility estimation result.

[0265] 1. The present invention has higher-precision visibility observation: by constructing a visibility recognition system based on satellites and visual images and combining different observation methods with artificial intelligence methods, the accuracy of visibility observation can be improved.

[0266] 2. The present invention has a wider range of visibility observation: the automated observation of visibility is no longer limited to expensive observation equipment with certain construction requirements. Satellite data has global coverage, and combined with a large number of camera resources, visibility observation can be completed more widely.

[0267] 3. The present invention has the ability to improve forecasting capabilities: For weather forecasting technology, better restoration of the current state of the atmosphere is a key factor in improving the accuracy of deducing future weather trends. High-precision, high-density visibility observations can better help restore the current state of the atmosphere, thereby improving weather forecasting capabilities.

[0268] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A visibility analysis method based on satellite observation and continuous images, characterized in that: The steps include: S1. Data preparation: Build a system database to collect continuous image data from fixed cameras, hourly visibility observation data from ground monitoring stations, and satellite inversion visibility data. S11, fixed camera continuous image data: The continuous image data of fixed cameras is provided by fixed cameras. The data of fixed cameras comes from traffic cameras, urban security cameras or special cameras deployed by other meteorological departments. The cameras continuously collect images at fixed intervals to form a time-series image data set. Data preprocessing for the time series image dataset: S111, Image size unification: All images are scaled to a uniform size, and the image size is uniformly interpolated using bilinear interpolation or cubic spline interpolation algorithms. I resize =resize(I original ,800,600) Among them I resize is the processed image, I original is the original image; S112, Color channel standardization: In order to meet the input requirements of the neural network, the three RGB color channels of the image need to be normalized separately. The normalization formula is as follows: Where I(x, y, c) is the value of the color channel c of the pixel (x, y), μ c and σ c are the mean and standard deviation of color channel c, respectively. S113, denoising and enhancement: Use median filtering or bilateral filtering to denoise the image: I denoise =medianFilter(I norm ,k) Where k is the filter window size, The image is enhanced by contrast stretching and brightness adjustment. S114, Data Enhancement: The image data is enhanced by random rotation, cropping, and flipping: I aug =randomAugment(I denoise ) The data-augmented images will be used together with the original images for visibility regression model training. S115, Time Alignment: Time alignment is performed through interpolation algorithms to ensure that the timestamp of the image is consistent with the hourly visibility observation data of the ground monitoring station and the visibility data inverted by the satellite. If the camera image is missing at a certain moment, it can be filled by interpolating the images at adjacent moments: This ensures that there is continuous image data at all time points for subsequent analysis; S12. Hourly visibility observation data of ground monitoring stations: The hourly visibility observation data of ground monitoring stations are provided by the meteorological department. The meteorological department monitors visibility through ground monitoring stations in various cities. The data of ground monitoring stations include hourly visibility data and meteorological data. The meteorological data is used to assist in analyzing visibility changes. Data preprocessing for hourly visibility observation data: S121, Time Alignment: Temporally interpolate the hourly visibility observation data to make it consistent with the camera image data collected by the fixed camera: Among them, V(t) is the visibility value at time t, t0 and t1 are two adjacent hourly moments, S122. Missing value processing: If the data of ground monitoring stations are missing at a certain moment, they can be supplemented by interpolation or using interpolation methods based on historical data; S13. Satellite inversion visibility data: The satellite-derived visibility data is obtained through polar-orbiting satellites or geostationary satellites, and its resolution depends on the corresponding satellite system. S131. Remove outliers: In the preprocessing stage, outliers in the satellite-retrieved visibility data are eliminated, and error data are detected through upper and lower limit constraints and statistical methods; S132, Time Alignment: The time gaps in the satellite-inverted visibility data are filled by linear interpolation to ensure that the camera image data, the hourly visibility observation data of the ground monitoring stations and the satellite-inverted visibility data correspond to each other at the same time. S14. All processed data will be stored in the system database and classified and managed by region and time. Data management includes: storage of raw data and storage of processed data. Each processed camera image data, hourly visibility observation data and satellite-derived visibility data will be time-stamped and spatially labeled to ensure consistency with the input of the visibility regression model; S2. Dynamic fusion of multi-source data space: By introducing the mesh model, the three types of data, namely camera image data, hourly visibility observation data and satellite inversion visibility data, are spatially fused and weighted as follows: The corresponding interpolation algorithm is used according to different data sources, and the contribution weight of each data source is dynamically adjusted based on the data distribution density to ensure that the visibility regression model has accurate prediction capabilities in space. Based on the highest resolution grid of satellite inversion visibility data, the target area is gridded in geographic space. The grid division is based on satellite resolution. The grid where the fixed camera is located is used as the core grid, and eight surrounding grids are expanded outward to form a 3x3 nine-square grid structure. Among these nine grids, the data of the core grid is integrated with fixed cameras, ground monitoring stations and satellite systems, while the peripheral grids are used for interpolation of satellite data and ground station data. The spatial data fusion method is adjusted according to the distribution of ground monitoring stations within the grid, and is divided into four cases: Case 1: No site There are no ground monitoring stations in the core grid and the eight surrounding grids. Due to the lack of ground monitoring station data, the visibility of the core grid is completely dependent on the satellite system. The calculation process does not require an interpolation algorithm. The visibility value is determined only based on the observation data of the satellite system. The data weight distribution is as follows: Satellite data weight within the core grid: ω sat =1 Ground station data weight: ω site =0 Core grid visibility algorithm formula: V core =V sat (core) Among them, V core is the visibility of the core grid after data fusion, V sat (core) is the satellite observation visibility data within the core grid, Case 2: Clustered Sites There is at least one ground monitoring station in the core grid, but no station in the surrounding eight grids. In this case, it is necessary to combine the station data in the core grid with the satellite data. If there is only one ground monitoring site in the core grid, the data of the ground monitoring site is directly used as the site visibility V of the core grid. site (core), If the core grid has more than one ground monitoring station, interpolation processing is required between the ground monitoring stations: Among them, V(x i ) represents the visibility observation value of site i, d(x0, x i ) is the distance between the camera position x0 and site i, p is the distance attenuation factor, Data weight distribution: Satellite data weight within the core grid: ω sat =0.5 Data weight of ground monitoring stations within the core grid: ω site =0.5 Taking into account the satellite data and site data of the core grid, the visibility calculation formula of the core grid is: V core =ω site ×V site (core)+ω sat ×V sat (core) Case 3: Sparse sites There is no ground monitoring station in the core grid, but there is at least one ground monitoring station in the surrounding 8 grids. In this case, the visibility of the core grid needs to be interpolated using the data of the surrounding ground monitoring stations. If there is only one ground monitoring station in the surrounding grid, the data of the ground monitoring station is directly used as the ground monitoring station visibility V of the surrounding grid. site (out), If there are multiple ground monitoring stations in the surrounding grid, use the Kriging interpolation method to interpolate the ground monitoring station data of the surrounding grid into the core grid: Among them, V(x i ) is the visibility observation value of the surrounding ground monitoring station i, λ i is the interpolation weight, determined based on geographic spatial correlation, satisfying Data weight distribution: Satellite data weight for the core grid: ω sat =0.8 Ground monitoring station data weight: ω site =0.2 The interpolated ground monitoring station data is combined with the satellite data of the core grid, and the visibility calculation formula of the core grid is: V core =ω site ×V site (out)+ω sat ×V sat (core) Case 4: Multisite There are ground monitoring stations in the core grid and the 8 surrounding grids. At this time, the ground monitoring station data of the core grid and the surrounding grids need to be processed separately. The ground monitoring station data in the core grid is interpolated using the inverse distance weighted method, and the ground monitoring station data in the surrounding grid is interpolated to the core grid using the Kriging interpolation method. The interpolation algorithm for ground monitoring stations within the core grid is as follows: The interpolation algorithm for the ground monitoring sites of the surrounding grid is: Data weight distribution: The ground monitoring station data weight of the core grid: ω site (core) = 0.6 Satellite data weight for the core grid: ω sat =0.3 The data weight of the ground monitoring stations in the surrounding grid: ω site (out) = 0.1 Combining the interpolation data of ground monitoring stations of the core and surrounding grids, and the satellite data of the core grid, the visibility calculation formula of the core grid is: V core =ω site (core)×V site (core)+ω sat ×V sat (core)+ω site (oμt)×V site (out): S3. Visibility regression model training: S31. Visual image feature extraction: The backbone network is used to extract features from visual images taken by fixed cameras to obtain high-dimensional features of the images. The backbone network is used to extract features from satellite data acquired by the satellite system to obtain high-dimensional features of the satellite data. The satellite data is standardized to obtain satellite values. S32. Establish a fully connected regression model: A fully connected neural network regression model is established by combining the high-dimensional features of the image extracted from the visual image and the high-latitude features or satellite values ​​extracted from the satellite data with the visibility label data; S33, model training: use multi-region training samples with long time scales to perform model training; S34, Model fine-tuning: In model training, the model trained with samples from multiple locations and long time scales needs to be fine-tuned when used in the field. In the model application, a model fine-tuning function is set to achieve fine-tuning of the model; S4. Visibility identification: In the application stage of the visibility regression model, we no longer rely on ground monitoring station data and satellite inversion visibility data, but use real-time images from fixed cameras for visibility recognition. By introducing a time weighting mechanism, we can make full use of the real-time images of fixed cameras at consecutive moments. For the target time t, the image data of 2n+1 moments before and after are extracted from the image sequence taken by the fixed camera, and visibility recognition is performed at each moment. Then, different weights are assigned to each moment based on the time difference between the image capture moment and the target moment; Finally, the visibility value at time t is obtained by weighted calculation method. Where V(t+i) is the estimated visibility at time t+i; ω i is a weight inversely proportional to the time difference i. The smaller the time difference, the greater the weight. Weight explanation: The weight ω0 corresponding to the current moment t is the largest, ensuring that the visibility regression model is dominated by the observation value at the current moment; the data at the previous and next moments are used as auxiliary, and the weight ω -2 ,ω -1 , ω1, ω2 gradually decrease with the increase of time distance.

2. The visibility analysis method based on satellite observation and continuous images according to claim 1, characterized in that: In step S11, the data collection interval of the fixed camera is every 5 minutes or 10 minutes.

3. The visibility analysis method based on satellite observation and continuous images according to claim 1, characterized in that: In step S12, the monitoring site data also includes meteorological data, which is used to assist in analyzing visibility changes.

4. The visibility analysis method based on satellite observation and continuous images according to claim 1, characterized in that: In step S13, the resolution of the satellite-inverted visibility data ranges from one hundred meters to nine kilometers.

5. The visibility analysis method based on satellite observation and continuous images according to claim 1, characterized in that: In step S33, data enhancement is achieved by scaling and transposing the visual image and satellite data.

6. The visibility analysis method based on satellite observation and continuous images according to claim 1, characterized in that: In step S34, the model fine-tuning function uses manual visual methods or mobile visibility monitoring equipment to measure the visibility value of the model application location, accumulates certain data, and fine-tunes the basic model to optimize the model effect.

Images