KR20230018624A - Method of Calculating Day and Night Total Cloud Cover using Photographing Image Ground Camera-based and Support Vector Machine Algorithm - Google Patents

Abstract

The present invention relates to a calculating method of day and night total cloud amounts using a support vector machine algorithm and an image photographed based on a ground camera, which uses a support vector machine algorithm using statistical property data of images photographed by a ground camera, time information, and the like as input data to calculate a total cloud amount, and converts each pixel of an image photographed by a camera equipped with a fisheye lens into RGB light and shade to remove surrounding covers and light sources and perform distortion correction to improve the reliability of calculation results. The calculating method of day and night total cloud amounts using a support vector machine algorithm and an image photographed based on a ground camera comprises: a preprocessing step of calculating statistical properties of RGB light and shade for past data photographed by a ground camera equipped with a fisheye lens, and classifying inputted past data into training data and verification data; a support vector machine building step of building a model capable of calculating an optimal total cloud amount through classification/regression analysis for statistical properties of preprocessed image data; a step of optimizing a support vector machine through a method of using data classified as training data in the preprocessing step to train the support vector machine and using data classified as verification data to verify the support vector machine; and a total cloud amount calculation step of inputting real-time data into the support vector machine to calculate a total cloud amount, and storing the data as a database.

Description

Method of Calculating Day and Night Total Cloud Cover using Photographing Image Ground Camera-based and Support Vector Machine Algorithm}

The present invention relates to a method for calculating total cloudiness during day and night using an image captured based on a terrestrial camera and a support vector machine algorithm, and more particularly, to statistical characteristics of images captured to calculate total amount of cloudiness from a camera-based observation device on the ground. Calculate the total amount of cloudiness using a support vector machine algorithm that takes data and time information as input data, but converts each pixel of the image captured by a camera equipped with a fisheye lens into RGB contrast to remove surrounding shields and light sources and distort it. A method for calculating total day and night cloudiness using a support vector machine algorithm and an image taken based on a terrestrial camera capable of improving the reliability of the calculation result by performing correction.

Clouds occupy about 70% of the global area and change the radiant energy balance every moment. In addition, interaction with aerosols changes the physical properties of clouds, causing global cooling and greenhouse effect, and changes in precipitation efficiency affect vegetation changes and securing water resources. Therefore, understanding the mechanism of cloud-aerosol-precipitation is essential for forecasting changes in weather and climate.

In particular, total cloudiness represents not only the weather forecast but also the meteorological condition of the runway for air navigation, and is used as an important input data for predicting changes in various weather-related factors. However, since the uncertainty of predicted meteorological variables increases according to the accuracy and precision of total cloudiness data, production of high frequency and high accuracy total cloudiness data is required.

Observation of total cloudiness on the ground is made through the human eye according to the normalized synoptic observation regulations of the World Meteorological Organization (WMO) from the past to the present, and total cloudiness is “tenth (1/10)” Or, it is written as “okta (1/8)”.

However, the measurement of total cloudiness through the human eye lacks consistency and objectivity of observation depending on the observer's condition or observation period.

In addition, despite the importance of continuous observation of total cloud cover during the day and night, the continuity of the data is insufficient because the observation is performed at a longer period in the case of the night, unlike the daytime.

In addition, it is difficult to construct a dense cloud observation network in an observation environment that is difficult for humans to access, such as at the top of a mountain. Therefore, many studies have been conducted and utilized to automatically capture (photograph) clouds and calculate total cloudiness through camera-based observation equipment on the ground that can replace these problems and continuously monitor clouds.

In this regard, many studies using camera-based observation equipment have been attempted to automatically observe clouds on the ground and calculate total cloudiness. These results can be used for numerical meteorological analysis and forecasting, and are ideal for cloud monitoring for local areas and are very economical.

In general, the amount of cloudiness can be calculated using the brightness of red, green, blue (RGB) colors of an image captured by a camera. In other words, the cloud is detected from the ratio or difference between the ratios or differences of RGB and the total amount of cloudiness is calculated by using the characteristics of different contrasts of RGB according to the scattering of light of the sky and clouds. For example, if RBR (red-blue ratio), the ratio of R to B, is more than 0.6 or if the difference between R and B (RBD, red-blue difference) is less than 30, classify pixels as cloud pixels and calculate total amount of cloudiness. .

However, these empirical methods are difficult to classify the sky and clouds under various meteorological conditions. That is, since the colors of the sky and clouds vary in accordance with the state of the atmosphere and the position of the sun, boundary conditions may change every moment.

Accordingly, recently, methods of detecting clouds and calculating total cloudiness by learning various images using a machine learning method, rather than such an empirical method, have been used. In other words, the calculation of the total amount of cloudiness using images taken from camera-based cloud observation equipment is calculated by constructing a model that predicts real conditions by learning data for which the correct answer is already known from a supervised learning-based machine learning model.

In addition, camera-based cloud observation on the ground is observed by attaching a fisheye lens with a viewing angle of 180° to the camera, so the size of the object becomes relatively smaller in the edge area of the image, and all surroundings, including the sky and clouds, are observed by the fisheye lens. Since the object is photographed, it is necessary to remove surrounding shields and light sources that brightly contaminate the image.

On the other hand, as a result of investigating the prior art related to the present invention, a number of patent documents were searched, and some of them are introduced as follows.

Patent Document 1 includes a first process of loading a first image obtained through a fish-eye lens of a wide-angle camera; a second process of constructing an image coordinate system for the first image; a third process of converting the image coordinate system into an image sensor coordinate system; a fourth process of calculating a zenith angle (θ) and an azimuth angle (φ) of a point in the image sensor coordinate system; And a fifth process of constructing a three-dimensional virtual space by calculating the coordinates P (x, y, z) on the Cartesian coordinate system for the one point, and displaying the two-dimensional image with distortion corrected in real time on a monitor. Disclosed is a correction method for a distorted image obtained using a fisheye lens that can be used in various ways such as indoor and outdoor security, surveillance, crime prevention, and a rear camera of a car, and an image display system for realizing the same.

Patent Document 2, from the sky image data taken through the sky view device, shielding removal in the image, image classification according to GBR pixel distribution, cloud pixel classification considering RBR boundary value, sunlight removal from classified cloud pixels, cloud pixel After the validation process is performed, the total cloudiness is calculated, and the total cloudiness calculation method and system using the RGB color of the all-sky image data are systematic and scientific, as well as satisfying both accuracy and objectivity. are doing

Patent Document 3 is a pre-processing unit that performs a function of creating input data of a machine learning model in which clouds misrecognized pixels such as snowfall are separated using brightness temperature and reflectance transmitted after observation from an observation satellite; a cloudiness determining unit that performs a function of determining total cloudiness by sequentially driving machine learning models trained for each cluster through the machine learning model input data created in the preprocessing unit; and an image display and point-by-point aggregation unit that performs a function of expressing the total cloudiness result calculated by the cloudiness determination unit as an image and point-by-point aggregate information, including objective total cloudiness data and sky conditions at the level observed on the ground. By expressing the image based on , the user can more accurately and conveniently use and judge the total cloudiness information for the desired area, and the satellite-based total cloudiness calculation method using machine learning to utilize the information for weather forecasting is starting

Patent Document 4 includes the steps of collecting original image data of the entire sky taken from the ground; setting the center coordinates and radius of the original image data, the viewing angle, and the size of the output image; converting each pixel of the collected original image data into brightness according to RGB color; removing the occlusion pixels by inputting an occlusion image corresponding to the original image data; performing distortion correction on the all-sky pixels from which the shielding pixels are removed; Including, producing an all-sky image using all-sky pixels for which distortion correction has been completed, and classifying pixels that cannot be distinguished as sky or clouds from image data of the entire sky taken from the ground as shielded images, and then actually photographed original images. Distorted image pre-processing for ground-based automatic calculation of total cloudiness by removing shielded images from image data and producing image data similar to what a person actually sees through distortion correction for images below the viewing angle and automatically calculating total cloudiness method is initiated.

KR 10-2014-0090775 A KR 10-1709860 B1 KR 10-1855652 B1 KR 10-2209866 B1

The present invention has been devised to solve the above problems of the prior art, and a support vector machine that uses statistical characteristics data and time information of images taken as input data to calculate total cloudiness from camera-based observation equipment on the ground. The total amount of cloudiness is calculated using an algorithm, but each pixel of the image captured by a camera equipped with a fisheye lens is converted to RGB contrast to remove surrounding shields and light sources and perform distortion correction, thereby improving the reliability of the calculation result. The purpose of this study is to provide a method and system for calculating day and night total cloudiness using an image captured based on a terrestrial camera and a support vector machine algorithm.

The present invention for achieving the above object includes a preprocessing step of calculating statistical characteristics of RGB contrast for past data captured by a terrestrial camera equipped with a fisheye lens and classifying the input past data into training data and verification data; A support vector machine configuration step of constructing a model capable of calculating an optimal amount of cloudiness through classification/regression analysis on statistical characteristics of preprocessed image data; optimizing the support vector machine by learning the support vector machine using data classified as training data in the preprocessing step and verifying the support vector machine using data classified as verification data; and a total cloudiness calculation step of inputting real-time data to the support vector machine to calculate total cloudiness and storing the data in a database.

In addition, according to the method for calculating total day and night cloudiness using a support vector machine algorithm and an image captured by a terrestrial camera of the present invention, the preprocessing step includes inputting past image data for a certain period of time; Classifying the input image data for each pixel and converting them into RGB (red, green, blue) contrast; removing a shielded area by objects shielding the sky and clouds from image data photographed with a fisheye lens, and performing orthogonal projection distortion correction to make the relative object sizes of the center and edge areas of the image the same; removing a light source in the screen by removing a corresponding pixel when the average of RGB contrast exceeds 240; Calculating statistical characteristics of RGB contrast for image data that has undergone shielding, light source removal, and distortion correction; and classifying past data as training data for use as learning data for the support vector machine and verification data for verifying the predictive performance of the support vector machine, and extracting them at a predetermined ratio.

In addition, according to the method for calculating total day and night cloudiness using a support vector machine algorithm and an image captured based on a terrestrial camera of the present invention, the support vector machine construction step is a hyperparameter according to classification/regression analysis of the relationship between input past data. A support vector machine is constructed through the setting of (hyper-parameters), and when the verification data is applied to the model constructed in the process of repetitive training for hyper-parameter search, the bias and average of the total predicted and observed total cloudiness It is characterized by constructing an optimal support vector machine by searching for a set value having the smallest square root error (RMSE) and the highest correlation coefficient (R).

In addition, according to the method for calculating total day and night cloudiness using a support vector machine algorithm and an image taken based on a terrestrial camera of the present invention, the support vector machine can classify the maximum margin for the distance between vectors. After finding the hyperplane consisting of , search for the optimal hyperplane by obtaining the parameters for which the mapping function is minimized, determine the boundary of the margin, and set the hyperparameter to allow the error of the margin to search the hyperplane and the margin for high-dimensional mapping. It is characterized in that the model is optimized by adjusting the error and constraint cost settings.

In addition, according to the method for calculating total day and night cloudiness using a support vector machine algorithm and an image taken based on a terrestrial camera of the present invention, when setting the hyper parameter, the set value is searched at dense intervals within a certain set value range, If there are several, it is characterized in that the optimal setting value of the hyperparameter is searched for by the number of iterations.

The method for calculating day and night total cloudiness using a support vector machine algorithm and an image captured by a ground camera according to the present invention inputs real-time data consisting of an image taken by a ground camera to a support vector machine to calculate total day and night cloudiness, For optimization of the vector machine, each pixel of the image captured by a camera equipped with a fisheye lens is converted to RGB contrast, surrounding shields and light sources are removed, and past data that has been preprocessed to perform distortion correction are used for daytime as well as nighttime data. It is possible to calculate the total amount of cloud in real time, and the reliability of the calculated result of total amount of cloud is improved.

1 is a flow chart illustrating a method for calculating the amount of total day and night cloudiness using a support vector machine algorithm and an image captured based on a terrestrial camera according to the present invention.
2 is a flowchart illustrating a preprocessing process of a method for calculating total day and night cloudiness using a support vector machine algorithm and an image captured based on a terrestrial camera according to the present invention.

Hereinafter, with reference to the accompanying drawings, a description will be given of a method for calculating total day and night cloudiness using an image captured based on a terrestrial camera and a support vector machine algorithm of the present invention.

The terms used in the present invention are terms defined in consideration of functions in the present invention, which may vary according to the intention or custom of the user or operator, so the definitions of these terms are meanings consistent with the technical details of the present invention. and should be interpreted as a concept.

In addition, the embodiments of the present invention do not limit the scope of the present invention, but are only exemplary of the components presented in the claims of the present invention, are included in the technical spirit throughout the specification of the present invention, and cover the scope of the claims. It is an embodiment including components that can be replaced as equivalents in components.

In addition, optional terms in the following embodiments are used to distinguish one element from other elements, and elements are not limited by the terms.

Therefore, in describing the present invention, detailed descriptions of related known technologies that may unnecessarily obscure the gist of the present invention will be omitted.

To explain a preferred embodiment of the present invention, FIG. 1 is a flow chart showing a method for calculating total day and night cloudiness using an image captured based on a terrestrial camera and a support vector machine algorithm according to the present invention, and FIG. 2 is a flowchart according to the present invention. It is a flow chart showing the preprocessing process of total day/night cloud cover calculation method using support vector machine algorithm and image captured by ground camera.

As shown in FIG. 1, the method for calculating total day and night cloudiness using an image captured by a terrestrial camera and a support vector machine algorithm according to the present invention is a statistical method of RGB contrast for past data photographed by a terrestrial camera equipped with a fisheye lens. A pre-processing step of calculating characteristics and classifying input past data into training data and verification data; A support vector machine configuration step of constructing a model capable of calculating an optimal amount of cloudiness through classification/regression analysis on statistical characteristics of preprocessed image data; optimizing the support vector machine by learning the support vector machine using data classified as training data in the preprocessing step and verifying the support vector machine using data classified as verification data; and a total cloudiness calculating step of inputting real-time data to the support vector machine to calculate total cloudiness and storing the data in a database.

Here, the pre-processing step includes inputting past image data of a certain period as shown in FIG. 2; Classifying the input image data for each pixel and converting them into RGB (red, green, blue) contrast; removing a shielded area by objects that shield the sky and clouds from image data photographed with a fisheye lens, and performing orthogonal projection distortion correction to make the relative object sizes of the center and edge regions of the image equal; removing a light source in the screen by removing a corresponding pixel when the average of RGB contrast exceeds 240; Calculating statistical characteristics of RGB contrast for image data that has undergone shielding, light source removal, and distortion correction; and classifying past data as training data for use as learning data for the support vector machine and verification data for verifying the predictive performance of the support vector machine, and extracting them at a predetermined ratio.

In the step of inputting the video data, past data at a point in time before real-time data is input is input to perform model optimization through training and verification of the support vector machine, and past data for a certain period is input. The input data includes images taken by a camera on the ground, total visual cloud data observed by a professional weather observer, recording time (Julian day, hour), and solar zenith angle (SZA) data. In addition, it is preferable to input data of at least several weeks to one year for the period, and to improve the learning performance of the support vector machine, it is preferable to set the period of data for a long period of time.

The step of converting to RGB contrast conversion is a process of converting the color of each pixel in image data stored in the form of a picture that we commonly see into red, green, blue (RGB) contrast. At this time, it is converted into R, G, and B contrast for each pixel along the x-axis and y-axis of the image, and each has an integer value between 0 and 255.

The shielding removal and fisheye distortion correction steps are processes of removing the shielding material covering the sky in the image and correcting the distortion of the image. That is, since the image captured by the fisheye lens captures all objects within the angle of view of the fisheye lens, it is necessary to remove the shielding area for objects such as buildings, trees, and other equipment that block the sky and clouds for calculation of total cloudiness. will be. In particular, since a solar zenith angle (SZA) of 80° or more permanently blocks the sky due to surrounding shielding materials, it is necessary to permanently remove the corresponding area when calculating total cloudiness. In addition, due to the nature of the image captured with the fisheye lens, since the imaged object is relatively small at the edge of the image, orthogonal projection distortion correction is performed to make the relative object size of the center and edge of the image the same. Correct distortion.

The step of removing the light source in the image is a process of removing bright pixels in order to prevent difficulty in distinguishing the sky and clouds due to the light source. Normally, when there is a light source in an image, the pixel where the light source is located (pixels close to RGB 255, 255, 255, white) or the pixels around it appear bright, which affects the RGB contrast of the actual sky and clouds. makes it difficult to discern. Therefore, bright pixels around the light source and the light source need to be removed. When the average of RGB contrast exceeds 240, the corresponding pixel is removed.

The step of calculating the statistical characteristics of the image is a process of calculating the statistical characteristics of RGB contrast of the image data that has undergone the process of removing shielding and light sources and correcting distortion so that it can be used as input data for the support vector machine. For most image data, the RGB contrast ratio and difference and luminance (Y) vary according to daytime, nighttime, time of day, season, degree of clouds in the image, weather conditions, etc., and the frequency distribution of the values also shows different characteristics. . Therefore, RBR (red-blue ratio), the ratio of R and B for each image data, BRD (blue-red difference), the difference between B and R, and the average, standard deviation, mode, and mode of the frequency distribution of luminance (Y) The frequency, kurtosis, skewness, and quartiles (Q0~Q4: 0%, 25%, 50%, 75%, 100%) are calculated and used as input data. At this time, the size of the class interval of each frequency distribution of RBR, BRD, and Y is set as a multiple of 0.01, 1, and 1 (eg, 0.02, 2, and 2 intervals), and it is better to ignore classes whose frequency of a specific class is less than 100. desirable. This is because statistical characteristics such as kurtosis, skewness, and quartiles can be sensitively changed by classes with small frequencies.

The data classification step is a process of classifying each data into training data and verification data so that they can be used for optimization of the support vector machine. That is, the average, standard deviation, mode, frequency of the mode, kurtosis, skewness, quartile of the RBR, BRD, Y frequency distribution, and the time (Julian day, hour) input in the image data input process, and SZA data and support vectors The visual cloud data for machine model learning is classified into training data and verification data, respectively. Training data and verification data are randomly extracted at a ratio of 8:2 based on the collected data (100%). Accordingly, the prediction performance of the support vector machine configured in the process of learning the training data can be performed through the verification data.

In addition, the support vector machine construction step is a step of constructing a support vector machine through setting of hyper-parameters according to classification/regression analysis on the relationship between input past data.

The support vector machine finds a hyperplane made up of support vectors that can classify the maximum margin for the distance between vectors, finds a parameter for which the mapping function is minimized, searches for an optimal hyperplane, and searches for the boundary of the margin. It is configured in such a way that the model is optimized by adjusting the margin, error, and constraint cost settings for hyperplane search and high-dimensional mapping by determining and setting hyperparameters to allow margin errors.

That is, in order to construct a support vector machine, after finding a hyperplane composed of support vectors capable of classifying the maximum margin for the distance between vectors, a parameter that minimizes the mapping function in the optimal hyperplane is obtained. . Thereafter, a radial basis function (RBF) kernel is applied to perform high-dimensional mapping. At this time, the boundary of the margin is determined, and a slack variable to allow margin error and a constraint cost that can violate these conditions are set. Therefore, by setting hyperparameters such as epsilon (ε), gamma (γ), and cost (C), the margin, error, and constraint cost settings for hyperplane search and high-dimensional mapping are adjusted, and the model is optimized through this.

When setting the hyper parameter, it is preferable to search for the set value at dense intervals within a certain set value range, and when there are several hyper parameters, the optimal set value of the hyper parameter is searched for as many times as the number of iterations. For example, when searching for epsilon (ε) at 0.05 intervals in the range of 0.05 to 0.50, gamma (γ) at 0.01 intervals in the range of 0.01 to 0.10, and cost (C) at 1 intervals in the range of 1 to 10, a total of 10 × It is to search for the optimal hyperparameter setting value through a process of 10 × 10 iterations. This process is used semi-permanently if it is performed only once for the first time for model construction, but it is desirable to update it periodically.

In addition, when the verification data was applied to the model constructed in the process of repetitive training for hyperparameter search, the bias and root mean square error (RMSE) of the predicted and observed total cloudiness were the smallest and the correlation coefficient (R) was The highest setting value is searched for and an optimal support vector machine is constructed.

The total cloud amount calculation step is a step in which the total amount of cloud is calculated by inputting real-time data after optimization of the support vector machine is completed, and the result is stored in a database.

The method for calculating the amount of day and night total cloudiness using the support vector machine algorithm and the image taken by the ground camera of the present invention configured as described above is real-time data composed of the image taken by the ground camera on the support vector machine constructed through past data for a certain period of time. By inputting , it is possible to calculate the amount of total cloud cover during the day as well as at night.

Although the above has been described and illustrated in relation to several embodiments for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as described above, and the scope of the technical idea described in the description of the invention It will be readily apparent to those skilled in the art that many changes and modifications can be made to the present invention without departing from this. Accordingly, all such appropriate changes and modifications and equivalents should be regarded as falling within the scope of the present invention.

Claims (5)

A pre-processing step of calculating statistical characteristics of RGB contrast for past data captured by a terrestrial camera equipped with a fisheye lens and classifying the input past data into training data and verification data;
A support vector machine configuration step of constructing a model capable of calculating an optimal amount of cloudiness through classification/regression analysis on statistical characteristics of preprocessed image data;
optimizing the support vector machine by learning the support vector machine using data classified as training data in the preprocessing step and verifying the support vector machine using data classified as verification data; and
A total cloudiness calculation step of inputting real-time data into a support vector machine to calculate total cloudiness and storing the data in a database; How to calculate cloudiness.
According to claim 1,
The preprocessing step is
inputting past image data of a certain period of time; Classifying the input image data for each pixel and converting them into RGB (red, green, blue) contrast;
removing a shielded area by objects shielding the sky and clouds from image data photographed with a fisheye lens, and performing orthogonal projection distortion correction to make the relative object sizes of the center and edge areas of the image the same;
removing a light source in the screen by removing a corresponding pixel when the average of RGB contrast exceeds 240;
Calculating statistical characteristics of RGB contrast for image data that has undergone shielding, light source removal, and distortion correction; and
Classifying past data as training data for use as learning data for the support vector machine and verification data for verifying the predictive performance of the support vector machine and extracting them at a certain ratio; A method for calculating total day and night cloud cover using a supported image and a support vector machine algorithm.
According to claim 1,
In the support vector machine construction step, a support vector machine is configured through the setting of hyper-parameters according to classification / regression analysis of the relationship between input past data,
When the verification data was applied to the model constructed in the process of repetitive training for hyperparameter search, the bias and root mean square error (RMSE) of the predicted and observed total cloudiness were the smallest and the correlation coefficient (R) was the highest. A method for calculating total day and night cloudiness using an image captured based on a ground camera and a support vector machine algorithm, characterized in that a support vector machine is configured by searching for set values.
According to claim 1,
The support vector machine finds a hyperplane made up of support vectors that can classify the maximum margin for the distance between vectors, finds a parameter for which the mapping function is minimized, searches for an optimal hyperplane, and searches for the boundary of the margin. , and setting hyperparameters to allow margin errors to optimize the model by adjusting margins, errors, and constraint cost settings for hyperplane search and high-dimensional mapping. A method for calculating total day and night cloud cover using image and support vector machine algorithm.
According to claim 4,
When setting the hyper parameters, the set values are searched at dense intervals within a certain set value range, but when there are several hyper parameters, the optimal setting values of the hyper parameters are searched for as many times as the number of repetitions. A method for calculating total day and night cloud cover using image and support vector machine algorithm.

Images