Method, device and medium for rapid detection of morning and evening fog based on adaptive threshold visual background extractor

Abstract

The present invention discloses a method, device, and medium for rapid detection of dawn and dusk fog based on an adaptive threshold visual background extractor. The method comprises: obtaining H8 / AHI data and ground station fog observation marker data during the dawn and dusk periods and preprocessing them; constructing a time series detection dataset based on the H8 / AHI data; constructing an adaptive threshold visual background extractor based on the ViBe model, denoted as the ST-ViBe model; introducing an LBSP texture feature description operator when initializing the background model, and introducing a SILTP texture feature when establishing a pixel foreground detection parameter set; adaptively adjusting model parameters based on the input time series detection dataset; using the ST-ViBe model, obtaining preliminary fog detection results based on the time series detection dataset; and optimizing the preliminary fog detection results using traditional cloud removal and post-processing methods to remove residual clouds and afterimages. The present invention can accurately, efficiently, and rapidly detect dawn and dusk land fog.

Description

Technical Field

[0001] The present invention relates to the fields of environmental monitoring and weather forecasting, and in particular to a method, device and medium for quickly detecting morning and evening fog based on an adaptive threshold visual background extractor (ST-ViBe). Background Art

[0002] Fog is a common weather phenomenon, often occurring during peak travel times, such as dawn and dusk. It can significantly impact respiratory health and traffic safety, making it a key monitoring target for meteorological and environmental agencies. Traditional fog detection, primarily through the deployment of observation stations, is inadequate for capturing the spatiotemporal patterns and trends of fog formation and evolution across a large region. With the rapid development of meteorological satellite remote sensing technology, remote sensing, with its advantages of large observation range, high temporal resolution, and low cost, has become a primary data source for large-scale fog detection research.

[0003] Traditional remote sensing fog detection relies on differences in reflected radiation and texture characteristics between fog and background information such as the ground surface in the visible and infrared bands, focusing primarily on daytime and nighttime fog detection. Daytime fog detection primarily relies on the spectral and textural characteristics of fog. For example, fog's reflectivity during the day is lower than that of mid- and high-altitude clouds, but higher than that of water and the ground. For example, the texture of fog tops is relatively smooth and uniform, while the texture of mid- and high-altitude clouds is relatively rough. These unique spectral and textural differences in daytime fog provide a physical basis for daytime fog detection. Nighttime fog remote sensing detection primarily relies on differences in its emission characteristics. The emissivity of fog in the 3.5μm-3.8μm mid-infrared band (MIR) is lower than its emissivity in the 8.5μm-13μm thermal infrared band (TIR). Based on this characteristic, establishing a dual-channel brightness temperature difference (BTD) between MIR and TIR is a very effective method for nighttime fog detection. However, detecting morning and evening fog is extremely difficult. The main reason is that the solar altitude is low at this time, resulting in a small difference in reflectance between fog and background in the visible light band. In the mid- and far-infrared bands, due to fog grounding, their spectral characteristics are also close to those of the ground surface, making it very difficult to detect morning and evening fog based on a single remote sensing image. However, the dual-channel brightness temperature difference (BTD) between fog and the ground surface varies significantly over time due to the influence of solar altitude. Nighttime fog, due to its inherent radiation characteristics, has a BTD far below zero. As the solar altitude increases, the BTD increases from negative to larger positive values. While the ground surface experiences similar changes, the magnitude of the changes is much smaller than that of fog. The opposite is true at dusk. Considering the significant differences in fog BTD between night and day, some researchers have used different thresholds to detect daytime, nighttime, and morning and evening fog separately. However, due to algorithmic and data limitations, it is not possible to fully extract morning and evening fog. Summary of the Invention

[0004] The present invention provides a method, device and medium for rapid detection of dawn and dusk fog based on an adaptive threshold visual background extractor (ST-ViBe), which can accurately, efficiently and rapidly realize dawn and dusk land fog detection.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A fast detection method for morning and evening fog based on an adaptive threshold visual background extractor, comprising:

[0007] Step 1: Obtain H8 / AHI data during the morning and evening hours and preprocess them;

[0008] Step 2: Construct a time series detection dataset for the morning and evening periods based on the H8 / AHI data;

[0009] Step 3: construct an adaptive threshold visual background extractor based on the ViBe model, denoted as the ST-ViBe model; wherein, the LBSP texture feature description operator is introduced in the background model initialization stage, and the SILTP texture feature is introduced in the pixel foreground detection parameter set establishment stage; and the parameters of the ST-ViBe model, including the minimum correlation number and the distance measurement threshold, are adaptively adjusted according to the input time series detection dataset;

[0010] Step 4: Use the ST-ViBe model to obtain preliminary fog detection results for the time series detection dataset;

[0011] Step 5: Use traditional cloud removal methods to optimize the detection results of the ST-ViBe model and remove the residual clouds in the fog detection results;

[0012] Step 6: Use post-processing methods to optimize the fog detection results after traditional cloud removal to remove the residual image in the fog detection results.

[0013] Furthermore, the specific process of constructing the time series detection dataset is as follows:

[0014] Step B1, subtract band 7 from band 14 in the H8 / AHI data to obtain a BTD image;

[0015] In step B2, the data of the H8 / AHI data with wavelengths of 8.7 μm, 10.8 μm, and 11.2 μm are fused with the BTD image data, a total of four channels, to construct a time series detection dataset.

[0016] Furthermore, in the background model initialization stage, the LBSP operator of the preset size window is selected as the calculation window of the neighborhood mean and variance, and the background model BM(I)1 based on BTD and the background model BM(I)2 based on the LBSP neighborhood brightness temperature mean and variance are established, where: BM(I)1 = {I1, I2, ... I n}, I t is the BTD value of the tth background pixel sample, and n is the background sample parameter; is the mean and variance of the LBSP operator brightness temperature difference of the t-th background pixel, and n is the background sample parameter.

[0017] Furthermore, the step of establishing the pixel foreground detection parameter set includes:

[0018] Step C2.1, calculate pixel I (x,y) The SILTP texture feature encoding is calculated as follows:

[0019]

[0020] Where: I c Pixel I (x,y) BTD value, I k Pixel I (x,y) The BTD value of pixel k in the eight-neighborhood, τ is the allowed fluctuation parameter; s τ (I c ,I k ) is the neighborhood pixel k and the center pixel I (x,y) Grayscale difference coding factor; Indicates sequential coding, pixel I (x,y) Grayscale difference coding factor s of all neighboring pixels τ (I c ,I k ) in sequence to form a sixteen-bit binary code SILTP (x, y);

[0021] Step C2.2: Count the number of times NUM_SILTP 1 appears in the binary code SILTP(x,y). (x,y) , defined as SILTP texture features, coupling pixel brightness temperature difference BTD and SILTP texture features to construct scene factor L c , calculated as follows:

[0022] L c =I c / NUM_SILTP (x,y)

[0023] Step C2.3, the center pixel I (x,y) BTD value I c , mean brightness temperature difference of LBSP neighborhood m c and variance SILTP texture feature NUM_SILTP (x,y) , scene factor L c Composition pixel foreground detection parameter set

[0024] Furthermore, the method for adaptively adjusting the minimum correlation number and the distance measurement threshold is:

[0025] Step C4.1, use the initialized background model to preliminarily adjust the parameters Min and R: calculate the sample mean in the background model BM(I)2 and variance If the current frame pixel I (x,y) BTD value I c In range If the pixel is within the range of 0, it indicates that the probability that the pixel is background is high. According to the following formula, the minimum intersection number Min is reduced and the distance measurement threshold R is increased:

[0026]

[0027] Step C4.2: divide the time into two different periods: morning and evening, and calculate the pixel I of the current frame. (x,y) The scene factor L c and SILTP texture feature NUM_SILTP (x,y) , adjust the distance measurement threshold R, specifically:

[0028] The adjustment method of the distance measurement threshold R in the morning scene is:

[0029]

[0030] Among them, R and R dawn are the distance measurement thresholds before and after adjustment in the morning period, respectively;

[0031] The adjustment method of the distance measurement threshold R in the dusk scene is:

[0032]

[0033] Where R and R dusk are the distance measurement thresholds before and after adjustment during the dusk period, respectively.

[0034] Furthermore, empirical values ​​are used as the initial values ​​of the minimum intersection number Min and the distance measurement threshold R of the ST-ViBe model.

[0035] Furthermore, when the ST-ViBe model is used for fog detection in step 4, if a pixel is detected as a background point, then: half of the samples of the background models BM(I)1 and BM(I)2 are extracted and replaced with the corresponding BTD value of the pixel detected as the background point and the mean and variance of the brightness temperature difference of the LBSP operator.

[0036] Furthermore, step 5 specifically includes:

[0037] Step E1.1, ice cloud removal: Based on the low brightness temperature of ice clouds in the 11.2 μm band of the H8 / AHI data, and based on experience and existing research, a threshold of 256 K is selected for removal.

[0038] Step E1.2, thin cirrus removal: thin cirrus clouds are removed based on the brightness temperature difference between the 8.7μm and 10.8μm bands of the H8 / AHI data being greater than 0K.

[0039] Step 6 is as follows: take the union of the pixels detected as clouds in the remote sensing image at the previous moment, and take the intersection with the fog detection result at the current moment. The intersecting pixels are false cloud detections and are removed.

[0040] An electronic device includes a memory and a processor, wherein the memory stores a computer program. When the computer program is executed by the processor, the processor implements the method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor as described in any of the above technical solutions.

[0041] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor as described in any of the above technical solutions.

[0042] Beneficial effects

[0043] Based on the ViBe model, this paper proposes a morning fog detection model, ST-ViBe, that couples the LBSP texture feature descriptor and the SILTP texture feature. This model implements morning fog detection based on an adaptive threshold visual background extractor. Compared to existing technologies, the ST-ViBe model offers the following advantages: Compared to the ViBe algorithm, GMM-ViBe algorithm, ViBeBgR, and SILTP-ViBe algorithm, the ST-ViBe model can fully detect fog over a wide area even in the presence of sudden morning and evening illumination changes and heavy cloud interference. Verification results using multiple days of ground observation data demonstrate that the ST-ViBe model has high overall accuracy for morning fog detection, with correct detection rates of 81.1% and 73.7% for the morning and evening periods, respectively, and false detection rates of 23.5% and 24.8%, respectively. Results from time-series imagery fog detection demonstrate that the algorithm can fully detect fog around and after the morning line during the morning period, and that the algorithm can fully detect fog on both sides of the evening line during the evening period. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flowchart of the fast detection algorithm for morning and evening fog based on the adaptive threshold visual background extractor;

[0045] Figure 2 These are the fog-brightness temperature difference images at 7:30 am (a) and 16:40 pm (b) on November 30, 2015. DETAILED DESCRIPTION

[0046] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present invention.

[0047] This embodiment provides a method for rapid detection of morning fog and dusk fog based on an adaptive threshold visual background extractor. Figure 1 As shown, the following steps are included:

[0048] Step 1: Raw data preprocessing: preprocessing the H8 / AHI data and ground station data used;

[0049] Step A1: H8 / AHI data is remote sensing image data acquired by the Advanced Himawari Imager (AHI) sensor aboard the Himawari-8 geostationary meteorological satellite. The preprocessing steps for H8 / AHI data are as follows:

[0050] Step A1.1: Radiation calibration. There are 16 bands in the original AHI data, of which bands 1-6 are calibrated as reflectance and bands 7-16 are calibrated as brightness temperature.

[0051] Step A1.2: Projection conversion: convert the calibrated AHI data to the WGS-84 coordinate system and resample the data spatial resolution to 2 km.

[0052] Step A2: Ground station data refers to the full-element Micaps-diamond1 mapping data recorded by ground meteorological stations. Ground stations with visibility less than or equal to 10,000 meters and relative humidity greater than 90% are marked as foggy stations. Among them, stations with visibility between 0 and 200 meters are marked as dense fog stations, stations with visibility between 200 and 500 meters are marked as moderate fog stations, stations with visibility between 500 and 10,000 meters are marked as light fog stations, and stations with visibility greater than 10,000 meters are marked as fog-free stations. This data is used for field verification of satellite data detection results.

[0053] Step 2: Construct a time series detection dataset for dawn and dusk.

[0054] Step B1: Construct the BTD brightness temperature difference image at dawn and dusk. There are 16 bands in the H8 / AHI data. Subtract band 7 (3.90 μm) from band 14 (11.2 μm) to obtain the BTD brightness temperature difference image. The BTD brightness temperature difference image is as follows: Figure 2 shown.

[0055] Among them, the solar zenith angle information of the image can be directly obtained in the H8 / AHI data, and the solar zenith angle at dawn and dusk is defined as (65°, 90°).

[0056] Step B2: Construct a time series detection dataset. The data from the four channels (BTD, 8.7μm, 10.8μm, and 11.2μm) are fused to construct a time series detection dataset.

[0057] Step 3: Build an adaptive threshold visual background extractor based on the ViBe model, denoted as the ST-ViBe model.

[0058] Based on the ViBe model, the LBSP (Local Binary Simolarity Patterns) texture feature description operator is introduced during the background model initialization phase. Based on the BTD brightness temperature difference data, a background model based on the LBSP domain mean and variance is established. Incorporating texture information into the algorithm can reduce the probability of complex textured clouds being detected as fog, improving the accuracy of foreground and background detection. The SILTP (Scale Invariant Local Ternary Pattern) texture feature is introduced during the establishment of the pixel foreground detection parameter set to address the problem of large variations in fog BTD values ​​with illumination within a single image, enabling adaptive determination of the foreground detection threshold. The specific process is as follows:

[0059] Step C1: Background model initialization. Select the LBSP operator of the 5*5 window as the calculation window of the neighborhood mean and variance, and establish two background models BM(I)1 and BM(I)2 of the mean and variance of the BTD and LBSP neighborhood brightness temperature differences, where: BM(I)1 = {I1, I2, ... I n}, I t is the brightness temperature difference of the tth background pixel, n is the background sample parameter, that is, the number of background samples; is the mean and variance of the LBSP operator brightness temperature difference of the tth background pixel, and n is the background sample parameter. Initialize the sample mean of the background model BM(I)2 and variance It is used to preliminarily adjust the parameters of the minimum intersection number Min and the distance measurement threshold R, reduce the probability of complex texture clouds being detected as fog, and improve the accuracy of foreground and background detection.

[0060] Step C2: Establish the pixel foreground detection parameter set FG(x,y). The steps of the pixel foreground detection parameter set FG(x,y) are as follows:

[0061] Step C2.1: Calculate pixel I (x,y) The SILTP texture feature encoding is calculated as follows:

[0062]

[0063] Where: I c Pixel I (x,y) BTD value, I k Pixel I (x,y) The BTD value of the pixels in the eight neighborhoods, τ is the allowed fluctuation parameter. According to the target spectrum variation characteristics in the present invention, τ is set to 0.3; s τ (I c ,I k ) is the neighborhood pixel k and the center pixel I (x,y) Grayscale difference encoding factor. Indicates sequential coding, pixel I (x,y) Grayscale difference coding factor s of all neighboring pixels τ (I c ,I k ) in sequence to form the sixteen-bit binary code SILTP(x,y).

[0064] Step C2.2: Define the parameter NUM_SILTP (x,y) The number of times 1 appears in the SILTP (x, y) code (between 0 and 8). When the solar zenith angle is greater than 90°, the fog edge and the ground NUM_SILTP (x,y) Between 5 and 8, when the solar zenith angle is small, the fog inside and the ground NUM_SILTP (x,y) Between 0 and 2. The scene factor L is constructed by coupling the pixel brightness temperature difference and SILTP texture features. The calculation method is as follows:

[0065] L c =I c / NUM_SILTP (x,y)

[0066] The pixel BTD(I c ), mean variance of brightness temperature difference in LBSP neighborhood SILTP texture features (NUM_SILTP (x,y) ), scene factor (L c ) constitutes the pixel foreground detection parameter set

[0067] Step C3: Determining the initial values ​​of the model parameters. This embodiment uses empirical values ​​as the initial values ​​of the ST-ViBe model parameters, where: background sample parameter n=20, minimum intersection number Min=4, and distance measurement threshold R=3k.

[0068] Step C4: Adaptively adjust the distance measurement threshold. Adaptively adjusting the distance measurement threshold R facilitates fog detection at different solar zenith angles, improving fog detection accuracy and fog area integrity. The foreground detection parameter set FG(x,y) is used together with the background models BM(I)1 and BM(I)2 to determine the distance measurement threshold R. The specific steps are as follows:

[0069] Step C4.1: Use the initialized background model to preliminarily adjust the parameters Min and R. Calculate the sample mean in the background model BM(I)2 and variance If the current frame pixel I (x,y) BTD value I c In range If the pixel is within the range of 0, it indicates that the probability of the pixel being the background is high. Reducing the minimum intersection number Min and increasing the distance measurement threshold R can increase the possibility of detecting it as the background:

[0070] Step C4.2: Divide into two different time periods, morning and evening, and calculate the pixel I of the current frame. (x,y) The scene factor L c and SILTP texture feature NUM_SILTP (x,y) , adjust the distance measurement threshold R, specifically:

[0071] The adjustment method of the distance measurement threshold R in the morning scene is:

[0072]

[0073] Among them, R dawn This is the adjusted detection threshold for morning hours. When the scene factor is less than 5, and the pixel is before or during the morning line, reducing the distance measurement threshold can ensure complete fog edge detection. When the scene factor is greater than 5, and the pixel is within uniformly textured fog or after the morning line, increasing the distance measurement threshold can reduce false detections of clouds and ground.

[0074] The adjustment method of the distance measurement threshold R in the dusk scene is:

[0075]

[0076] Where R dusk is the detection threshold at dusk. When the scene factor is less than 0, the pixel brightness temperature difference is less than 0, which means it is the area after the dusk line passes. dusk Set to 1k. When the scene factor is greater than 10, the area is before the dusk line passes, R dusk Set to 2k; when the scene factor is between 0-10, the area is near the dusk line, R dusk Set to 1.5k.

[0077] Step C5: Foreground detection and background update. Using the adaptive distance measurement threshold R, the foreground is detected and the background is updated according to the traditional VIBE algorithm's foreground and background determination rules. The ST-ViBe algorithm rapidly updates the background by increasing the background update amount and rate. All pixels detected as background points undergo a rapid background update. Half (n / 2) samples of the background models BM(I)1 and BM(I)2 are randomly extracted and replaced with the current pixel values, ensuring that the background models contain more of the latest background values.

[0078] Step 4: Input the time series detection dataset into the ST-ViBe model to obtain preliminary fog detection results.

[0079] Step D1: Before fog detection, the newly acquired data is preprocessed in step A1 and the time series detection dataset construction steps B1 and B2 are performed. After completion, the image is input into the ST-ViBe model to obtain preliminary fog detection results.

[0080] Step 5: Use traditional cloud removal methods to optimize the detection results of the ST-ViBe algorithm and remove the residual clouds in the detection results.

[0081] Step E1: Compared to conventional high-frequency time-series imagery, the dawn and dusk periods are shorter, resulting in only a dozen frames of H8 / AHI time-series video. The ViBe algorithm is suited for high-frame video object detection, so it cannot remove very fast-moving clouds. Therefore, traditional cloud removal methods are used, combined with the unique spectral and textural characteristics of mid- and high-altitude clouds in the H8 data, to sequentially remove ice clouds and thin cirrus clouds from the imagery, removing any residual clouds from the ST-ViBe algorithm's initial fog detection results. The specific steps are as follows:

[0082] Step E1.1: Ice cloud removal. Ice clouds have a low brightness temperature in the 11.2 μm band of the H8 / AHI data. Based on experience and existing research, a threshold of 256 K is selected for removal.

[0083] Step E1.2: Thin cirrus clouds are removed. Thin cirrus clouds are removed if the brightness temperature difference between the 8.7μm and 10.8μm bands of the H8 / AHI data is greater than 0K. This is used as the threshold.

[0084] Step 6: Use post-processing methods to optimize the fog detection results after traditional cloud removal to remove the residual image in the detection results.

[0085] Step F1: Removing afterimages. Afterimages are traces left by fast-moving clouds. Clouds are affected by atmospheric motion and move quickly. Therefore, clouds moving several moments before the current moment are prone to forming afterimages and being mistakenly detected as fog. This algorithm takes the union of pixels detected as clouds in the remote sensing imagery at previous moments and intersects them with the fog detection result at the current moment. Intersecting pixels are false cloud detections and are removed. For example, the fog detection result at 5:00 PM may contain afterimages of clouds at 4:30 PM, 4:40 PM, and 4:50 PM. Using the traditional cloud removal method in step E1, clouds at these moments are detected, taken from the union, and intersected with the fog detection result at 5:00 PM. The intersecting pixels are then removed from the fog detection result at 5:00 PM. Furthermore, due to the presence of noise in the detection results due to phenomena such as different objects with the same spectrum, the algorithm uses a 3x3 median window to denoise the detection results to obtain the final fog detection result.

[0086] Next, we verify the accuracy of the ST-ViBe algorithm detection results of the present invention:

[0087] Ground observation data at 8:00 in the morning and 17:00 in the evening were selected to conduct qualitative and quantitative evaluation of the satellite fog detection results at the same time from November 27 to November 30, 2015 to verify the reliability of the algorithm.

[0088] Comparison of the detection results with ground observations shows that all sites with dense and moderate fog are within the algorithm's detection results. Most sites with light fog agree with the detection results, while sites without fog are largely absent. The algorithm effectively captures the wide-scale distribution of fog at dawn and dusk, and is generally spatially consistent with ground observations. However, there are some false positives. For example, in the mornings of November 27 and 29, significant fog was detected in the northern part of the study area, even though ground observations indicated no fog. The algorithm mistakenly detected low clouds as fog. Satellite detection results from the evenings of November 27 and 29 also showed significant noise in the northern part of the study area, indicating that fast-moving clouds were mistakenly detected as fog. H8 / AHI false-color imagery and satellite fog detection results from 7:10-8:10 and 16:10-17:10 on November 30, 2015, demonstrate that the ST-ViBe algorithm effectively detects fog near the terminator, at times similar to the surface brightness temperature difference. As the terminator moves forward in the morning, the extent of the fog detected within the study area gradually increases. Fog is detected in the daytime low-sun-angle region to the right of the morning line. However, the algorithm is unable to detect fog in the nighttime region before the morning line passes. During dusk, the detection range shifts with the dusk line, effectively detecting fog within a certain range to both left and right of the dusk line. This demonstrates that the algorithm has a reasonable ability to detect fog in the low-sun-angle region and in the early nighttime region at dusk. Some fog detection results at dusk exhibit slight afterimages, a sign of rapidly moving clouds. Fog detection results at dusk also exhibit slight afterimages, a sign of very fast-moving clouds. Despite the ST-ViBe algorithm's rapid background updates and afterimage post-processing, it is unable to remove these fast-moving clouds.

[0089] The ST-ViBe algorithm has varying adaptability at night and in the morning and evening. This is primarily due to the fact that the area before the morning line passes remains dark, with minimal fog BTD variation and no discernible rate of change. This is considered background fog during time-series image modeling and therefore undetectable. While the area after the evening line passes remains dark, some of the previous timeframe was daytime, resulting in a significant rate of change compared to the current pixel. Compared to the established background model, this larger value is considered foreground fog, making the evening algorithm suitable for low sun altitudes and the beginning of night.

[0090] To test the effectiveness of the ST-ViBe algorithm, a qualitative comparison of the fog detection results of the ST-ViBe algorithm, the traditional ViBe algorithm (R = 3k), and the ViBeBgR algorithm (R = 3k) at 7:30 and 16:50 was performed. The ViBe algorithm was unable to detect fog near the morning line whose spectral characteristics were consistent with those of the ground surface. The ViBeBgR algorithm achieved similar detection results to the VIBE algorithm. The ST-ViBe algorithm fully detected fog near the morning line whose spectral characteristics were similar to those of the ground surface, as well as fog in the area where the morning line passed, with complete detection results. The fog area detected by the ViBe algorithm was larger than the actual fog area, and some cloud areas were falsely detected. The ViBeBgR algorithm performed well, but had some noise and cloud area false detection issues. The ST-ViBe algorithm fully detected fog on both sides of the evening line, but a small amount of cloud was falsely detected as fog.

[0091] Comparison of the detection results of three different algorithms, ST-ViBe algorithm, ViBe algorithm and ViBeBgR algorithm, at morning and evening time shows that the ST-ViBe algorithm can more completely detect fog in a larger range at morning and evening time, and the fog interior is intact, with good detection results.

[0092] Ground observation data is used to quantitatively verify the detection accuracy of the algorithm. The commonly used indicator evaluation system is used to evaluate the detection results, including the probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI). These indicators are defined as:

[0093]

[0094] Where: N X is the number of detections, and its subscript X is the type of detection index, including H, M, and F. (H indicates that the satellite detection result is consistent with the ground observation result, that is, a correct detection; M indicates that the satellite detection result shows no fog but the ground data shows fog, that is, a false detection; F indicates that the satellite detection result shows fog but the ground data shows no fog, that is, a false detection). The above detection indices all range from 0 to 1. A larger POD indicates a higher detection accuracy; a larger CSI indicates a more effective method.

[0095] Tables 1 and 2 show the fog detection accuracy at 8:00 AM and 5:00 PM over four days from November 27 to 30, 2015. As shown in Tables 1 and 2, the average accuracy of satellite fog detection at 8:00 AM over the four days was 72.9%, the average error rate was 12.7%, and the average reliability factor was 66%. The average accuracy of satellite fog detection at 5:00 PM over the four days was 68.9%, the average error rate was 14.5%, and the average reliability factor was 61.7%. The algorithm has a high accuracy rate, a slightly lower error rate, and a reliability factor greater than 60%. Fog detection accuracy during dusk is lower than during the morning.

[0096] Further analysis of the distribution of ground observations in satellite fog detection results shows that the missed points in Tables 1 and 2 are mostly light fog. Because it is thin, its information is relatively weak in remote sensing images, and the sensor used in this paper does not have obvious characteristic bands to capture it. The false detection points are mostly located in the remote sensing fog detection results. Because the algorithm extracts information based on spectral features, the sensor receives information from the top of the target. Radiation fog is mostly surface-shaped water vapor condensation, so the remote sensing detection results are surface-shaped. In addition to the influence of water vapor, the actual weather conditions on the ground are also affected by surface cover type, surface wind speed, and topography. Fog may be dissipating or not grounded as low cloud, resulting in false detections. In the future, we will focus on how to combine these influencing factors to further screen satellite detection results in order to obtain more accurate detection results. In addition to false detections in satellite detection results, ground observation data can also cause certain errors due to recording rules. For example, when weather phenomena such as haze, cloudy, rain, and snow occur simultaneously with fog, ground observation stations preferentially record them as rain and snow. The blurred distinction between haze and fog also affects the accuracy of detection results.

[0097] Table 1. Fog detection accuracy at 08:00 in the study area

[0098]

[0099] Table 2. Fog detection accuracy at 17:00 in the study area

[0100]

[0101] Data from 12 months in 2015 were selected to verify the algorithm detection effect.

[0102] Tables 3 and 4 demonstrate the applicability of the ST-ViBe algorithm for all 12 months of the year. The validation results for morning fog show that the overall accuracy of the POD, FAR, and CSI are 72.5%, 18.5%, and 62.4%, respectively. Compared to morning fog, the performance at dusk is relatively weak, with POD, FAR, and CSI indicators for the entire study area being 70.6%, 33.6%, and 52.3%, respectively.

[0103] Table 3. Fog detection accuracy at 08:00 in the study area

[0104]

[0105] Table 4. Fog detection accuracy at 17:00 in the study area

[0106]

[0107] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A fast detection method for morning and evening fog based on adaptive threshold visual background extractor, characterized in that: include: Step 1: Obtain H8 / AHI data during the morning and evening hours and preprocess them; Step 2: Construct a time series detection dataset for the morning and evening periods based on the H8 / AHI data; Step 3: construct an adaptive threshold visual background extractor based on the ViBe model, denoted as the ST-ViBe model; wherein, the LBSP texture feature description operator is introduced in the background model initialization stage, and the SILTP texture feature is introduced in the pixel foreground detection parameter set establishment stage; and the parameters of the ST-ViBe model, including the minimum correlation number and the distance measurement threshold, are adaptively adjusted according to the input time series detection dataset; Among them, the method for adaptively adjusting the minimum correlation number and distance measurement threshold is: Step C4.1, use the initialized background model to preliminarily adjust the parameters Min and R: calculate the sample mean in the background model BM(I)2 and variance If the current frame pixel I (x,y) BTD value I c In range If the pixel is within the range of 0, it indicates that the probability that the pixel is background is high. According to the following formula, the minimum intersection number Min is reduced and the distance measurement threshold R is increased: Step C4.2: divide the time into two different periods: morning and evening, and calculate the pixel I of the current frame. (x,y) The scene factor L c and SILTP texture feature NUM_SILTP (x,y) , adjust the distance measurement threshold R, specifically: The adjustment method of the distance measurement threshold R in the morning scene is: Among them, R and R dawn are the distance measurement thresholds before and after adjustment in the morning period, respectively; The adjustment method of the distance measurement threshold R in the dusk scene is: Where R and R dusk are the distance measurement thresholds before and after adjustment during dusk period, respectively; Step 4: Use the ST-ViBe model to obtain preliminary fog detection results based on the time series detection dataset; Step 5: Use traditional cloud removal methods to optimize the detection results of the ST-ViBe model and remove the residual clouds in the fog detection results; Step 6: Use post-processing methods to optimize the fog detection results after traditional cloud removal to remove the residual image in the fog detection results.

2. The method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor according to claim 1 is characterized in that: The specific process of constructing a time series detection dataset is as follows: Step B1, subtract band 7 from band 14 in the H8 / AHI data to obtain a BTD image; In step B2, the data of the H8 / AHI data with wavelengths of 8.7 μm, 10.8 μm, and 11.2 μm are fused with the BTD image data, a total of four channels, to construct a time series detection dataset.

3. The method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor according to claim 1, characterized in that: In the background model initialization stage, the LBSP operator of the preset size window is selected as the calculation window of the neighborhood mean and variance, and the background model BM(I)1 based on BTD and the background model BM(I)2 based on the LBSP neighborhood brightness temperature mean and variance are established, where: BM(I)1 = {I1, I2, ... I n }, I t is the BTD value of the tth background pixel sample, and n is the background sample parameter; is the mean and variance of the LBSP operator brightness temperature difference of the t-th background pixel, and n is the background sample parameter.

4. The method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor according to claim 1, characterized in that: The steps in establishing the pixel foreground detection parameter set include: Step C2.1, calculate pixel I (x,y) The SILTP texture feature encoding is calculated as follows: Where: I c Pixel I (x,y) BTD value, I k Pixel I (x,y) The BTD value of pixel k in the eight-neighborhood, τ is the allowed fluctuation parameter; s τ (I c ,I k ) is the neighborhood pixel k and the center pixel I (x,y) Grayscale difference coding factor; Indicates sequential coding, pixel I (x,y) Grayscale difference coding factor s of all neighboring pixels τ (I c ,I k ) in sequence to form a sixteen-bit binary code SILTP (x, y); Step C2.2: Count the number of times NUM_SILTP 1 appears in the binary code SILTP(x,y). (x,y) , defined as SILTP texture features, coupling pixel brightness temperature difference BTD and SILTP texture features to construct scene factor L c , calculated as follows: 50 c =I c / NUM_SILTP (x,y) Step C2.3, the center pixel I (x,y) BTD value I c , mean brightness temperature difference of LBSP neighborhood m c and variance SILTP texture feature NUM_SILTP (x,y) , scene factor L c Composition pixel foreground detection parameter set 5. The method for rapid detection of morning and evening fog based on adaptive threshold visual background extractor according to claim 1, characterized in that: The empirical values ​​are used as the initial values ​​of the minimum intersection number Min and the distance measure threshold R of the ST-ViBe model.

6. The method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor according to claim 1, characterized in that: When using the ST-ViBe model for fog detection in step 4, if a pixel is detected as a background point, then: extract half of the samples of the background models BM(I)1 and BM(I)2 respectively, and replace them with the corresponding BTD value of the pixel detected as the background point and the mean and variance of the brightness temperature difference of the LBSP operator.

7. The method for rapid detection of morning and evening fog based on an adaptive threshold visual background extractor according to claim 1, characterized in that: Step 5 specifically includes: Step E1.1, ice cloud removal: Based on the low brightness temperature of ice clouds in the 11.2 μm band of the H8 / AHI data, and based on experience and existing research, a threshold of 256 K is selected for removal. Step E1.2, thin cirrus removal: thin cirrus clouds are removed based on the brightness temperature difference between the 8.7μm and 10.8μm bands of the H8 / AHI data being greater than 0K. Step 6 is as follows: take the union of the pixels detected as clouds in the remote sensing image at the previous moment, and take the intersection with the fog detection result at the current moment. The intersecting pixels are false cloud detections and are removed.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images