JP3375038B2 - Precipitation pattern change prediction method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for short-term prediction of time series changes in a rainy area in a narrow area based on a weather radar echo image, and more particularly to a method for predicting precipitation pattern changes in which topographic effects are added. Regarding the device.

[0002]

2. Description of the Related Art Conventionally, when performing weather forecasts, the Meteorological Agency, meteorological system-related companies, airline companies, etc., have various two-dimensional temperature, pressure, dew point, wind vector, etc. obtained from meteorological satellites and ground observation. Using physical parameters, we have been predicting changes in meteorological phenomena nationwide tens of hours ahead or days ahead through mathematical and physical equations. On the other hand, in recent years, in a small area, weather forecast in a short time of the order of minutes,
In particular, since information on rainfall and snowfall is desired, it is becoming important to predict changes using the information on changes in the rain range from locally located weather radar devices. Such a prediction method is particularly called “nowcasting”. Meteorologists using weather radar echoes
Its mission is to constantly monitor changes in the situation 24 hours a day to prevent disasters in advance.

A change in the rain area means a temporal change in an area in which it is raining or snowing, and the area has an apparent cloud shape. The gray value is almost proportional to the precipitation amount. The rain area as an image is expressed as a combination of various shapes and gray values, and it is difficult to describe a clear feature amount. Therefore, automatic detection of rain area motion based on image processing methods is not possible. It's not easy. That is, the tracking method required for the precipitation pattern is different from the tracking of the movement of a rigid body such as an automobile, or the tracking of the movement of an elastic body in an echo image of the heart.
It is impossible to realize automatic prediction without tracking while creating and disappearing.

Therefore, even now, subjective reading is performed from changes in precipitation patterns based on human experience.
That information is used.

On the other hand, on the other hand, in order to automatically and objectively change the rain area on a computer, a method of detecting the rain area change by the spatiotemporal correlation method and linearly extrapolating the same precipitation pattern ( The linear extrapolation method) has been tried. This method
After calculating the similarity of the gray value between two image frames [f (t) and f (t + 1): t is a time step), it is assumed that the point having the highest correlation value corresponds to the movement amount. In the above, the amount of movement in the rainy area is estimated. The rain area is translated based on the estimated amount of movement. By repeating the rain image of the next step estimated in this way, it is possible to predict up to a specific number of steps ahead.

Attempts have been made to estimate changes in the precipitation area several minutes to several hours ahead from a two-dimensional image of the precipitation area at a certain time.
During the prediction, it was not possible to predict the degree of change in the precipitation area because the same image was translated in parallel. Since it is a natural phenomenon, the size, shape, and gray value of the precipitation pattern change non-steadily, but such expressions are not reflected in the prediction method. In addition, the effect of topography is not considered. That is, the two-dimensional information called an image does not include, as an expression, a series of physical processes in the original precipitation area, such as springing, sucking, dissipating, growing, and declining.

In the linear extrapolation method, an image can be divided into a plurality of regions to estimate movement vectors for a plurality of precipitation patterns, but no rule for changing the gray value has been introduced. In addition, the estimated movement vectors have also faced various directions, so if the prediction time is advanced a little, the movement vectors will come closer to each other and intersect. At the same time, the precipitation patterns unnaturally overlap in the vicinity of the intersection, which causes a problem that a large amount of rainfall is predicted due to the overlap. Therefore, there was a clear limit to improving the accuracy of the prediction algorithm based on the linear extrapolation method.

There is a prediction method based on a neural network model, and attempts have been made to acquire the mapping relationship between time-series two-dimensional images as a weighting coefficient of the neural network. However, the structure of the optimal neural network cannot be easily estimated, and the predicted image only gives the average of the learned patterns. In other words, the gray value of the precipitation pattern is constant, and only the prediction that the center of gravity of the pattern does not move has been obtained, and it can be said that the prediction of the precipitation pattern by the neural network model is impossible at present.

[0009]

The spatiotemporal correlation method applied so far estimates one motion vector from two consecutive images. However, the gray value of the image does not always have a one-to-one correspondence between the gray value of the previous frame and the gray value of the next frame because the rain area is continuously generated and extinguished. However, there is a problem in that the reliability of the correlation value based on the gray value is not necessarily high, but rather unstable. Furthermore, in the case of a diffuse rainy area image where the correlation coefficient cannot be obtained, the motion vector can hardly be estimated, and no concrete method has been found for the interpolation method between them. is there.

In the correlation method, it is necessary to appropriately prepare a large window function when the amount of movement per unit time is large and a small window function when the amount of movement is small. Is not easy to determine, and at the same time, if it is attempted to be determined, there arises a problem that the operation time required for it is increased by several to ten times.

When the rain area is stagnant, the change in the gray value of the rain area is more severe than the change in the contour shape of the rain area. In such a rainy area, if the correlation method is applied as it is, there is a change in the density of the rainy area even if there is no movement of the entire rainy area.
Some kind of motion vector was obtained, and the rainy area was moved by the motion vector. Therefore, there is a problem that the more the prediction progresses, the more the rainy region deviates.

Since the linear extrapolation method is a simple model in which a strong assumption that the change in the precipitation area is constant is made, a prediction method considering the degree of change in the precipitation area due to convection has not been attempted. It was At least the rapid development of the precipitation area,
Or, regarding the temporary development of the precipitation area during the decline process, because of lack of statistical knowledge, no prediction has been attempted that corresponds to the development of the precipitation area.

In the linear extrapolation method, an image can be divided into a plurality of regions to estimate movement vectors for a plurality of precipitation patterns, but no rule for changing the gray value has been introduced. Further, since the estimated movement vector also faces various directions, the movement vectors come close to each other and cross each other if the prediction time is advanced a little. Each movement vector was treated independently. At the same time, the problem that precipitation patterns overlap unnaturally near the intersection will occur.

The precipitation pattern may be affected by topography such as mountains and the sea, depending on the location. However, the linear extrapolation method can be incorporated at most as a prediction method by changing the movement vector between the sea and the ground. Even in a neural net model, it is not clear how much the terrain effect is added, and the mathematical expression ability of how to reflect physical effects that cannot be learned in the model. Lacks flexibility.

In the linear extrapolation method, with respect to two-dimensional information called an image, the diffusion, advection,
A series of physical processes such as springing, inhaling, dissipating, growing, and declining cannot be expressed in a unified manner.

The present invention solves the above problems and takes into consideration the effect of topography to develop the precipitation area in radar images.
It is an object of the present invention to provide a precipitation pattern change prediction method and device for predicting decline with high accuracy in the form of a two-dimensional image.

That is, the present invention predicts a time series change of a rainy region in a narrow area for a short time based on a weather radar echo image, and smoothly conducts various economic activities sensitive to the change of weather. As shown in Fig.
A method and device for predicting the development and decline of quantity in the form of images in a two-dimensional manner, and at the same time, supporting a specialist who is engaged in weather radar echo images for 24 hours as a system.

[0018]

A precipitation pattern change prediction method of the present invention has a calculation means, and includes an image input unit, an image storage unit, an image processing unit, a terrain effect introduction unit, an advection / diffusion equation calculation unit, and a prediction. Is a method of predicting a change in a precipitation pattern using a precipitation pattern change predicting apparatus including an output unit and an output unit. The radar image of the rain area to be predicted is input using the image input unit, and the past time-series images input by the image input unit are stored in the image storage unit. The image processing unit calculates the amount of change in precipitation and the image feature amount of the edge gradient distribution between two or more continuous two-dimensional images extracted from the image storage unit for calculation. Advection vector when the predicted topographical location of the precipitation area is on the sea
Is set to increase the size and decrease the diffusion coefficient.
Stored and stored in advance, the location of the location of the predicted precipitation area
If the shape is land, the magnitude of the advection vector becomes smaller
It is set and stored so that the diffusion coefficient becomes large . Advection
In the diffusion equation calculation unit, various image features calculated by the image processing unit are used as initial values, and the spatiotemporal change in precipitation is regarded as a variable. The diffusion coefficient and advection vector stored in the terrain effect introduction unit are also used. Numerical calculation based on the mathematical expression of the advection / diffusion equation. The prediction unit predicts the precipitation pattern with the topographic effect based on the numerically calculated result, and the output unit outputs the precipitation area as a time-series image from the result predicted by the prediction unit.

The calculation of the image feature amount of the edge gradient distribution in the image processing section may include a process of dividing the target region into a region having a large gradient value and a region having a small gradient value.

Advection / advection used in the diffusion equation calculator
Diffusion equations include time terms, advection
Term, diffusion term, sou
rce), sink (sink), dissipation (disp)
If the amount of precipitation is compared with the gray value on the image, the gray value at each pixel may be given as a variable of the equation.

The diffusion coefficient in the diffusion term of the advection / diffusion equation should be changed according to the rate of change of the one-dimensional area calculated from the number of pixels in which the gray value of the precipitation pattern occupies one frame of the image. It may include a process of dividing one frame into a plurality of small blocks under a predetermined condition and calculating the number of pixels of the precipitation pattern in the same procedure for each small block.

In calculating the one-dimensional area change rate for changing the diffusion coefficient, when the area change is obtained from a plurality of past frames, the time-series data of the past area change is determined according to a predetermined condition. Two-dimensional function and sin, cos
Approximate fitting to any of the non-linear functions including etc. by the method of least squares, non-linear extrapolation about the change of the area is performed up to a predetermined prediction time, and the one-dimensional non-linear prediction of the area change is performed. It may include a process of interlocking and changing the corresponding diffusion coefficient.

The source and sink terms of the advection / diffusion equation may be given values in positive and negative regions, respectively, when the difference in the precipitation pattern between consecutive frames is taken. When the source and sink terms are obtained from the image data in, the source and sink image areas are set to either the center of gravity movement vector or the estimated advection vector according to the prescribed conditions for a predetermined prediction time. It may include a process of moving in correspondence with each other.

The advection vector used in the advection-diffusion equation is obtained by extracting the change in the center of gravity of the region with a higher edge gradient in the gray value of the precipitation pattern, and when the advection property is stronger, the advection vector with a lower edge gradient is given. The direction of the maximum similarity point found between the image frames based on the change of the center of gravity extracted from the region and the cross-correlation method, and the movement vector of the precipitation pattern by giving either one of the intervals, the direction of the maximum similarity point based on the correlation method, To calculate the interval, one frame is divided into a plurality of small blocks, and the movement vector of the precipitation pattern is obtained by the cross-correlation method for each same small block between frames. It may include the process of estimating.

When estimating the advection vector in the image from the precipitation pattern, the contour line of the precipitation pattern is extracted at a certain interval, and the displacement of the contour line is used as a rough advection vector, which is predicted by the moving average filter. When the movement vector estimated by the cross-correlation method is used, including the process of iteratively propagating the advection vector for all pixels in the image with the number of iterations increased or decreased in proportion to the time length It may include a process of propagating the movement vector over the entire image by the iterative method using a plurality of movement vectors estimated from the intervals as an initial vector.

When extracting the contour line from the precipitation pattern of the image for estimating the advection vector, binarization, labeling processing, isolated point removal, expansion / contraction processing to the precipitation pattern containing large and small cloud blocks are performed. After performing a basic pre-processing including, the cloud of a specific number is extracted, the boundary line search algorithm extracts the boundary line of the cloud cloud, and the discretized contour line. For each point sequence of, the search may be performed along the normal direction of the contour line in the past frame until it intersects with the contour line in the current frame, and the displacement of the contour line between two frames may be obtained. .

When the precipitation pattern moves between the sea and the ground during the target time for which the calculation for prediction is performed by using the advection-diffusion equation, a predetermined value is calculated between the sea and the ground. The size of the advection vector may be increased or decreased, and at the same time, the size of the diffusion coefficient may be increased or decreased to include the terrain effect in the calculation, or the extinction coefficient above the sea than above the ground may be increased.

The advection / diffusion equation may include a process of performing discretization in a calculation space and solving it by using any one of a difference method, a finite element method and a spectrum method according to a predetermined condition.

When the calculation for prediction is performed using the advection / diffusion equation, the region of the precipitation pattern is divided into a region having a large edge gradient (first differential value) and a region having a small edge gradient by image processing based on the grayscale distribution of the image. It is also possible to divide by the value, obtain the center of gravity of a large area, weight the suction amount according to the distance from the center of gravity, and adjust the extinction coefficient to increase as the distance increases.

When carrying out the calculation for prediction using the advection-diffusion equation, in addition to inputting the gray value of the precipitation pattern into the advection-diffusion equation as a variable, carbon monoxide and nitrogen emitted from vehicles and factories The concentration values of oxides and sulfur oxides may be input to the advection / diffusion equation, and the effect of the concentration values of chemical substances on the observation area may be applied to the prediction calculation.

When performing a calculation for prediction using the advection / diffusion equation, when integrating the advection / diffusion equation with time,
The prediction result may be obtained by selecting a time width smaller than the observed time and repeating the time integration.

A precipitation pattern change predicting apparatus of the present invention is a precipitation pattern change predicting apparatus having a calculating means for predicting a precipitation pattern change, and for inputting a radar image of a rain range to be predicted. An image input unit, an image storage unit that stores past time-series images input by the image input unit, and a precipitation amount between two or more consecutive two-dimensional images extracted from the image storage unit for calculation. The image processing unit that calculates the amount of change and the image feature amount of the edge gradient distribution, and the size of the diffusion coefficient and the size of the estimated advection vector are changed according to the predicted topography of the position where the precipitation area exists. The terrain effect introductory part to be stored as a memory and the various image feature amounts calculated in the image processing part are used as initial values, and the amount of change in precipitation over time and space is regarded as a variable. Advection vector Also, the advection / diffusion equation calculation unit that performs numerical calculation based on the mathematical expression of the advection / diffusion equation, the prediction unit that predicts the precipitation pattern with the topographic effect based on the numerical calculation results, and the prediction unit And an output unit that outputs the precipitation area as a time-series image from the obtained result.

In the present invention, the radar echo image information itself is given as a variable in the advection / diffusion equation, and the precipitation area is automatically predicted based on this equation. Also,
Based on image processing techniques, the components of the various physical effects included in this equation are the time term, advection term, diffusion term,
Each term such as the source term, the suction term, the extinction term, and
The feature is that the diffusion coefficient is sequentially given. further,
Terrain effects can be easily incorporated. For example, changing the advection vector for each image area or pixel unit,
The diffusion coefficient can be changed and given. Regarding the terrain effect, once the observation range is determined, it is sufficient to determine the increase / decrease in the diffusion coefficient and the increase / decrease in the weighting amount for the advection vector at sea and on land.

Since the precipitation phenomenon is approximated by this equation and a plurality of effects can be treated in a unified manner, almost no specialized knowledge is required in application. At the time of prediction, since the initial shape and gray value of the precipitation region are given and the physical equation is used for prediction, it is possible to obtain a highly complex prediction result of the shape and gray value with high accuracy.

When making a prediction, it is sufficient to store only the data of the past 30 minutes for the two-dimensional image data in the memory of the computer, and the one-dimensional area change and intensity change of about one hour. You only need to store the data.

[0036]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is an explanatory diagram of a precipitation pattern change prediction method and apparatus according to an embodiment of the present invention,
(A) is a flowchart of a precipitation pattern change prediction method,
(B) is a block diagram of a precipitation pattern change prediction device. In the figure, reference numeral 11 is an image input unit, 12 is an image storage unit,
13 is an image processing unit, 14 is a terrain effect introduction unit, 15 is an advection / diffusion equation calculation unit, 16 is a prediction unit, 17 is an output unit, S
101 to S110 are each step.

The precipitation pattern change prediction apparatus of the present invention accumulates the image input unit 11 for inputting the radar image of the rain region to be predicted, and the past time series images input by the image input unit 11. An image storage unit 12, and an image processing unit 13 for calculating an amount of change in precipitation amount and an image feature amount of edge gradient distribution between two or more continuous two-dimensional images extracted from the image storage unit 12 for calculation. Calculated by the image processing unit 13 and the terrain effect introduction unit 14 that changes and stores the size of the diffusion coefficient and the estimated size of the advection vector according to the terrain at the position where the precipitation region is predicted in advance. Based on the mathematical expression of the advection-diffusion equation, using various image feature amounts as initial values, the amount of change in precipitation in time and space is regarded as a variable, and the diffusion coefficient and advection vector stored in the terrain effect introduction unit 14 are also used. Numerical counter Advection-diffusion equation calculating section 15 for
And a prediction unit 16 that predicts a precipitation pattern with a topographic effect based on the numerically calculated result, and an output unit 17 that outputs the precipitation area as a time-series image from the result predicted by the prediction unit 16. It

Next, each step of the local rain area prediction method of the present invention will be described with reference to FIG. When the prediction starts (S
101), a radar image of a desired rain area of a desired area is input from an external weather radar echo image using the image input unit 11 (S102) and stored in the image storage unit 12 (S103). The image storage unit 12 stores past time series images.

Next, the present image and the past image to be compared are extracted from the image storage unit 12 (S104), and the amount of change in the precipitation amount and the edge between the two or more consecutive two-dimensional images thus extracted are extracted. The image feature amount of the gradient distribution is calculated (S10
5) In the terrain effect introducing unit 14, the size of the diffusion coefficient and the size of the estimated advection vector are changed and stored according to the terrain at the position where the precipitation region is predicted in advance (S106), and the advection is performed. In the diffusion equation calculation unit 15, various image feature amounts calculated by the image processing unit 13 and the shape / grayscale value information of the initial precipitation area are given as initial values, and the amount of temporal change in precipitation amount is changed. Considering that, the terrain effect introduction section 14
Numerical calculation is performed by using the diffusion coefficient and the advection vector stored in the above, discretizing the equation by the difference method based on the mathematical expression of the advection / diffusion equation, and integrating the equation with respect to time (S107). The prediction unit 16 predicts the precipitation pattern with the topographic effect added based on the numerically calculated result (S108), and the output unit 17 predicts the prediction unit 1
The shape and gray value of the precipitation area are output as a time-series image from the result predicted in 6 (S109), and the process ends (S1).
10).

FIG. 2 is a schematic diagram showing a change between two frames of a precipitation area displayed on an actual radar echo image. FIG. 2A is a schematic diagram showing a change state of one precipitation pattern and an edge gradient value. Graph (b) is a schematic diagram of a state of changes in a plurality of precipitation patterns. In the figure, reference numerals 20 and 25 are precipitation patterns, 21 is a region with a small edge gradient, and 22 is a region with a large edge gradient.

As a method for calculating the feature amount of the edge gradient distribution in the image processing unit 13, here, an edge is extracted from a two-dimensional image by the Sobel operator, and a region 22 having a large gradient is shaded and the rest is left. Is emphasized as a region 21 having a small gradient value. From such image processing, it has been found that the precipitation region has a region with a large amount of precipitation in the center and a region with a small amount of precipitation in the periphery. In all the examples, the precipitation pattern 20 has a double structure, and the cross section of the edge gradient as shown in the edge gradient value graph shown in FIG. 2A always has a mountain-like structure. There are features. In addition, it has been confirmed that the development and decline of the precipitation area always spreads to the weak precipitation area and then dissipates after the strong precipitation area in the shaded area changes first. This is because the center of the convection is located in a region where the precipitation region is strong, and it is considered that the three-dimensional convection intensity appears as such in a two-dimensional image. From this, it is suggested that it is possible to predict the increase or decrease of the area while following the changes in the intensity of the precipitation area.

FIG. 3 is a schematic diagram showing a state in which the difference between two consecutive frames of the precipitation area is extracted, and reference numeral 3 in the figure.
1 is the advanced region, 32 is the past region that does not overlap with the present,
33 is an area where the present and past overlap, 35 is the current area,
36 is a past area.

When the precipitation area moves from the left to the right of the figure, the difference between the two areas (current-past) is taken, and the advanced area 31 (positive) and the non-overlapping area 32 are past areas.
(Negative), a region 33 (zero) in which the both overlap is obtained. As will be described later, the gray value information of these regions is given to the source and sink terms in the advection-diffusion equation.

The advection / diffusion equation used in the advection / diffusion equation calculation unit 15 is as shown in equation (1).

[0045]

[Equation 1] In equation (1),

[0046]

[Equation 2] In addition, the discretization is performed by the difference method. The left side of the equation (1) is a diffusion term, an advection term, a well term, a suction term, and an extinction term from the time term and the first term on the right side. λ is the diffusion coefficient.

[0047]

[Equation 3] In the above formula,

[0048]

[Outer 1] Indicates the gray value at i, j of the nth frame.

The equation (2) is discretized within one frame of the image showing the precipitation area, and a simultaneous linear equation having a size corresponding to the number of pixels of one frame is solved under the initial condition and the boundary condition.
The predicted image is the output result itself of equation (2).

FIG. 4 is a schematic view showing the arrangement of the discretized image and the boundary conditions necessary for solving the equation. In FIG.
The marks indicate points inside the precipitation area, and the ○ marks indicate points outside the precipitation area. Since the precipitation area sees a part of the actual phenomenon, the calculation grid imposes continuous conditions on the inside and outside points. As can be seen from the equation (2), this is the first derivative
When the term of the second derivative is discretized, from the reference point (i, j),
By referring to grid points that are separated by ± 1 in the front and back. The continuous condition, as is often applied in computational numerical mechanics,
For each component of each term in equation (2), the outermost point (circle) of the calculation grid and the point inside one of these points (circle) are simply connected by an equation to calculate the time integration. Every time we do, we refer to the equation.

The initial conditions are the image itself of the precipitation area when the prediction is started, the difference image value, the diffusion coefficient, and the advection vector. The diffusion coefficient is adjusted in advance so that it has a proportional relationship with the magnitude of the change in area. The advection vector is as follows.

FIG. 5 is a schematic graph showing a one-dimensional area prediction method. In the figure, reference numeral 51 is an observed value and 52 is 2
Quadratic function fitting, 53 is the observed area change, 5
4 is the predicted area change. In the calculation of the one-dimensional area change rate for changing the diffusion coefficient, the area change of the actually detected precipitation pattern is obtained by simply counting the number of pixels for each frame by image processing. To be This area change shows a gentle curve when viewed on a scale of several tens of hours. This curve shows global changes in the development and decline of precipitation patterns. Therefore, it is possible to predict the change in area one-dimensionally by performing an approximate fitting of this curve with a non-linear function. Various nonlinear functions can be applied,
Empirically, the quadratic function has been found to be the most suitable. The quadratic function is fitted to the area data about one hour past by shifting the quadratic function little by little by the method of least squares, and after fitting, the area change is extrapolated based on the quadratic function until a predetermined time ahead. By the way, when a higher-order function of the third order or higher is applied by the method of least squares, in most cases, there is a problem that inappropriate plural extreme values appear.

FIG. 6 is a schematic correspondence graph of the area change rate and the diffusion coefficient. In order to change the diffusion coefficient in conjunction with the area change rate, the area change rate is calculated from the precipitation pattern, and the diffusion coefficient is given in a Z shape. That is, when the rate of change of area is positive, the precipitation pattern is in the developing process,
When the diffusion coefficient is increased, on the contrary, when the diffusion coefficient is negative, the diffusion coefficient is decreased because it is in the process of withdrawal. However, the diffusion coefficient is not zero but is asymptotically zero.

FIGS. 7, 8, 9, and 10 are explanatory views of a method for estimating the advection vector from the displacement of the contour line of the precipitation pattern.

FIG. 7 is a schematic plan view for explaining a method of obtaining normals / tangents from individual blocks which are cloud blocks in the rainy area. In the figure, reference numeral 71 is a contour line and 72 is a contour point. ,
73 is a tangent line and 74 is a normal line. Here, it is assumed that the contour line 71 of the rain area has a circular pattern. The circle pattern is discretized by an appropriate number of contour points 72. The direction of the tangent line 73 on the contour point 72 in the figure is a component of the central difference value using the contour points 72 on both sides of the contour point 72 as the center. Therefore, the direction of the normal line 74 is equal to the result obtained so that the inner product with this component becomes zero. Such calculation is performed for all the discretized contour points 72. In addition, in order to apply to a rainy area pattern having an actual complicated contour shape, it is necessary to perform binarization, expansion / contraction, isolated point removal, and labeling as preprocessing.

FIG. 8 is a schematic plan view showing an example in which a contour line is expanded between two image frames of a rainy area block. FIG. 8A shows a state in which the contour line is isotropically expanded. (B)
Indicates a state in which the contour line is anisotropically expanded. Reference numeral 8 in the figure
1, 86 are past contour lines, 84, 87 are normal lines, 85, 8
8 is the current contour line.

(A) and (b) differ from each other in how the small circles are displaced from the great circles. Here, it is assumed that a small circle is a certain rainy area in the past and a large circle is a rainy area of the present size. In the case of (a), the deformation amounts of the two large and small circles are obviously isotropic. On the other hand, in the case of (b), a large amount of deformation is obtained in the upper direction and a small amount of deformation is obtained in the downward direction. As described above, the change of the contour line is usually not uniform.

FIG. 9 is a schematic plan view showing the result of calculating the deformation amount of the contour line in the actual rain area. In the figure, reference numeral 91 is a past contour line, 94 is a normal line, and 95 is a current contour line. It is a line. The detection of the contour line from the precipitation pattern for the advection vector estimation is performed by the basic image processing such as the isolated point removal and the binarization processing, the labeling processing, and the expansion / contraction processing several times each as the preprocessing. Boundary lines are detected by the boundary line search algorithm.

FIG. 9 shows an example in which a block in the rain area shown by small halftone dots spreads a few frames ahead of the image. The relative movement speed is obtained by extending the normal line 94 of the contour line 91 of the halftone dots inward and outward, and the amount of deformation up to the intersection with the larger contour line 95. in this case,
It can be seen that, although the rain area spreads upward as a whole, the lower part of the rain area is deformed into a convex shape and a concave shape because it also moves downward.

FIG. 10 is a schematic plan view showing the result of estimating the overall moving speed from the local deformation amount (moving speed). In the figure, reference numeral 101 is a past contour line, and 104 is a speed vector. , 105 are the current contour lines. By substituting the partially obtained initial moving speed into the Navier-Stokes equation, the total moving speed from the next frame,
That is, the advection speed is predicted. If the time integration is continued, the predicted advection velocity can easily be obtained in the next image frame. Alternatively, by repeating the moving average filter iteratively, the advection vector of the entire image is estimated from the moving vectors that are partially collected.

FIG. 11 is a schematic diagram showing the influence of the topography and the relationship with parameters. (A) is a schematic diagram showing changes in precipitation patterns over time, and (b) shows the magnitude of advection and diffusion coefficient. In the figure, reference numeral 111 is the sea, 112 is land, 113 is a precipitation pattern, 114 is a large advection / small diffusion coefficient state, and 115 is a small advection / diffusion coefficient large state.
As an example of the case where the precipitation pattern moves between the sea and the land, the case where the sea 111 and the land 112 exist in the radar will be described. At the initial stage, a small precipitation pattern 113 on the sea gradually enters the land due to advection.
As it enters the land, the precipitation pattern grows gradually. Fine precipitation patterns are scattered in the rear part. The time when a fine precipitation pattern on the sea is visible to the radar is shorter than that on land. this is,
When the precipitation pattern 113 approaches land, the advection component is weakened due to ground friction, the vertical convection component becomes dominant, and the movement of the precipitation pattern on land continues while the supply of steam from the sea continues. The speed slows down, but the rainy area continues to expand.

In response to the advection / diffusion equation, the influence of the terrain can be easily incorporated by controlling the advection vector and the diffusion coefficient at sea and on land. That is, the advection component may be given a large value on the sea, a little on the land, and the predicted advection vector may be weighted for each pixel, and the diffusion coefficient may be given a little on the sea and a little on the land.

FIG. 12 is a schematic diagram for explaining a method of increasing or decreasing the values of the diffusion coefficient and the extinction coefficient according to the distance from the center of precipitation, in which reference numeral 121 is the sea, 122 is the land, 123 is the center of precipitation.

The center of convection always exists in the precipitation pattern, which corresponds to a region with a high edge gradient value. As the distance from the center of precipitation 123, which is the center of gravity of this region, increases, the diffusion coefficient becomes smaller and the extinction coefficient becomes larger.
Changes in these coefficients may be given during prediction using the advection / diffusion equation. However, the diffusion coefficient is
In Expression (2), λ _{i, j} is changed so as to be given for each pixel.

FIG. 13 is a schematic diagram showing the result of following the change of the center of gravity of the precipitation area by the image processing.
Reference numeral 31 is a region having a small edge gradient, 132 is a region having a large edge gradient, 133 is a precipitation pattern, and 134 is a center of gravity.

This result is the advection vector in equation (2).

[0067]

[Outside 2] Given to. The edge gradient is obtained from the precipitation area image, and the change of the center of gravity 134 over time in the area having a high gradient value is tracked. The direction and magnitude of movement of the center of gravity is assumed as the average advection component occurring in the precipitation area. When a more local advection component is required, the velocity vector obtained as the solution of the Navier-Stokes equation is given to each grid point. In this way, the use of the moving amount of the center of gravity 134 in a region having a high edge gradient value is
It is suitable to use it together with the cross-correlation method in which the corresponding points are not always obtained stably.

FIG. 14 is a schematic diagram showing an example of detecting the amount of movement of a precipitation pattern based on the cross-correlation method.
Reference numeral 1 is a past precipitation pattern, 142 is a present precipitation pattern, and 143 and 144 are movement vectors.

In this example, one frame is divided into nine small blocks, and a cross-correlation coefficient is calculated between the frame and the corresponding small block of the next frame. In each small block, the distance and direction of the point having the highest similarity are obtained, and this is regarded as the movement vector 143 in each small block. However, unlike the conventional method, these movement vectors are not used as advection vectors and the rain area is not linearly extrapolated. Instead, a moving average filter is used to propagate the moving vector to each pixel in the small block to interpolate. The movement vector 144 obtained from this may be given to the advection vector term in the advection / diffusion equation for each pixel.

In the present invention, the magnitude of the diffusion coefficient and the advection vector is changed according to the location of the predicted precipitation area such as the sea and land, and the effect on the precipitation pattern of the terrain is adjusted and predicted. However, when calculating for prediction using the advection-diffusion equation, in addition to inputting the gray value of the precipitation pattern as a variable into the advection-diffusion equation, it is also emitted from vehicles and factories. By inputting the concentration values of carbon monoxide, nitrogen oxides, and sulfur oxides into the advection / diffusion equation, and applying the influence of the concentration values of chemical substances to the observation area to the prediction calculation, the prediction accuracy is further improved. be able to.

FIG. 15 is a comparison graph of the evaluation results of the predictions of the present invention and the conventional method. (A) shows the comparison result by CIS and (b) shows the comparison result by ERR. CSI (Critica), which is widely used as the hit rate of the area
It can be seen that the use of the l Success Index suppresses the decrease in hit rate more than the conventional method (cross-correlation method or neural net model). Further, the gray level value, that is, the predicted value of the precipitation amount is also better in the present invention.

From the above performance evaluation results, it is possible to easily understand the high accuracy of the prediction of the present invention. Also, in the case of stagnant precipitation areas, from the forecast that was until a few hours in the past,
Practical predictive value has been achieved even in the prediction of 5 or 6 hours ahead.

[0073]

As described above, according to the present invention,
By properly connecting time-dependent physical equations and image processing techniques, it is possible to predict the gray value, shape, etc. as the actual precipitation area / precipitation, and by adding the terrain effect, it is possible to achieve a highly accurate hit rate. effective.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of a precipitation pattern change prediction method and apparatus according to an embodiment of the present invention. (A) is a flowchart of a precipitation pattern change prediction method. (B) is a block diagram of a precipitation pattern change prediction device.

FIG. 2 is a schematic diagram showing a change between two frames of a precipitation area displayed on an actual radar echo image. (A) is a schematic diagram of a state of change of one precipitation pattern and a graph of edge gradient values. (B) is a schematic diagram of a change state of a plurality of precipitation patterns.

FIG. 3 is a schematic diagram showing a state in which a difference between two consecutive frames of a precipitation area is extracted.

FIG. 4 is a schematic diagram showing an arrangement of boundary conditions necessary for solving a discretized image and an equation.

FIG. 5 is a schematic graph showing a one-dimensional area prediction method.

FIG. 6 is a schematic correspondence graph of the area change rate and the diffusion coefficient.

[Fig. 7] Normals from individual blocks that are cloud blocks in the rain area
FIG. 6 is a schematic plan view for explaining a method of obtaining a tangent line.

FIG. 8 is a schematic plan view showing an example in which a contour line is widened between frames of two images of a block in a rainy area. (A) shows a state in which the contour line is isotropically expanded. (B) shows a state in which the contour line is anisotropically expanded.

FIG. 9 is a schematic plan view showing a result of calculating a deformation amount of a contour line in an actual rain area.

FIG. 10 is a schematic plan view showing a result of estimating the overall moving speed from a local deformation amount (moving speed).

FIG. 11 is a schematic diagram showing the relationship between the influence of topography and parameters. (A) is a schematic diagram which shows a time series change of a precipitation pattern. (B) is a schematic diagram which shows the magnitude of advection and a diffusion coefficient.

FIG. 12 is a schematic diagram for explaining a method of increasing or decreasing the values of diffusion coefficient and extinction coefficient according to the distance from the center of precipitation.

FIG. 13 is a schematic diagram showing a result of following a change in the center of gravity of a precipitation area by image processing.

FIG. 14 is a schematic diagram showing an example of detecting a movement amount of a precipitation pattern based on a cross-correlation method.

FIG. 15 is a comparison graph of prediction evaluation results according to the present invention and a conventional method. (A) is a comparison result by CIS.
(B) is a comparison result by ERR.

[Explanation of symbols]

11 Image input section 12 Image storage unit 13 Image processing unit 14 Terrain effect introduction section 15 Advection / diffusion equation calculator 16 Predictor 17 Output section 20, 25, 113, 133 precipitation patterns 21 Area with small edge gradient 22 Area with large edge gradient 31 Advanced Area 32 Past areas that do not overlap with the present 33 Areas where the present and the past overlap 35 Current Area 36 Past Area 51 observed values 52 Quadratic function fitting 53 Observed area change 54 Predicted area change 71 outline 72 contour points 73 tangent 74, 84, 87, 94 Normal 81, 86, 91, 101 Past contour lines 85, 88, 95, 105 Current contour line 104 velocity vector 111, 121 sea 112, 122 land 114 State of large advection and small diffusion coefficient 115 State with small advection and large diffusion coefficient 123 Center of precipitation 131 Area with small edge gradient 132 Area with large edge gradient 134 Center of gravity 141 Past precipitation patterns 142 Current precipitation pattern 143, 144 movement vector Steps S101 to S110

Continuation of the front page (56) Reference JP-A-9-61546 (JP, A) JP-A-8-287262 (JP, A) JP-A-5-333160 (JP, A) JP-A-52-9342 (JP , A) JP 62-237379 (JP, A) JP 7-306952 (JP, A) JP 8-271649 (JP, A) JP 57-60275 (JP, A) JP 4-76637 (JP, B2) JP-B-62-40668 (JP, B2) Hidetomo Sakaino, Riki Horikoshi, Satoshi Suzuki, "A method for predicting precipitation pattern changes using advection and nucleic acid equations in radar echo images", Electronic Information Proceedings of IEICE General Conference Information and Systems 2, Japan, The Institute of Electronics, Information and Communication Engineers, March 6, 1997, D-11-197, p. 197 Kazuhiro Otsuka, Riki Horikoshi, Satoshi Suzuki, “Optical flow estimation method based on integration of motion trajectories in spatiotemporal images”, IPSJ research report, Japan, Information Processing Society of Japan, January 23, 1997, Volume 97, No. 10, p. 75-81 (58) Fields investigated (Int.Cl. ⁷ , DB name) G01W 1/00-1/18 G01S 13/95 G06T 1/00 JISST file (JOIS)

Claims (17)

(57) [Claims] 1. A precipitation pattern change prediction apparatus having a calculation means, comprising an image input section, an image storage section, an image processing section, a topographic effect introduction section, an advection / diffusion equation calculation section, a prediction section and an output section. A method of predicting a change in a precipitation pattern by inputting a radar image of a rain region to be predicted using the image input unit, and inputting a radar image in the image storage unit in the past time input by the image input unit. Sequential images are accumulated, and the image processing unit calculates an amount of change in precipitation between two or more consecutive two-dimensional images extracted from the image storage unit for calculation and an image feature amount of edge gradient distribution. However, in the above-mentioned topographic effect introduction part, if the topography at the position where the predicted precipitation area exists is on the sea, the size of the advection vector
Is set and stored so that the diffusion coefficient becomes larger and the diffusion coefficient becomes smaller.
However, the topography of the location where the predicted precipitation area exists is on land.
, The advection vector becomes smaller and the diffusion coefficient
Is set so as to be large , and the advection / diffusion equation operation unit uses various image feature amounts calculated by the image processing unit as initial values, and the amount of temporal change in precipitation as a variable. It is assumed that the diffusion coefficient and the advection vector stored in the topographic effect introduction section are also used to perform a numerical calculation based on the mathematical expression of the advection-diffusion equation, and the prediction section adds the topographic effect based on the result of the numerical calculation. A precipitation pattern change prediction method, comprising: predicting a pattern, and outputting the precipitation area as a time-series image from the result predicted by the prediction unit by the output unit. 2. The calculation of the image feature amount of the edge gradient distribution in the image processing unit includes a step of dividing a target region into a region having a large gradient value and a region having a small gradient value.
Prediction method of precipitation pattern change described in. 3. The advection / diffusion equation used in the advection / diffusion equation calculation unit includes a time term and an advection
on), diffusion (diffusion), source (sink), sink (sink), and dissipation (dispersion) are included, and when precipitation is compared with the gray value on the image, The precipitation pattern change prediction method according to claim 1, wherein the gray value is given as a variable of the equation. 4. The diffusion coefficient in the diffusion term of the advection-diffusion equation changes in accordance with the rate of change of the one-dimensional area calculated from the number of pixels in which the gray value of the precipitation pattern occupies one frame of the image. 4. The method according to claim 3, further comprising the step of dividing one frame into a plurality of small blocks under a predetermined condition and calculating the number of pixels of the precipitation pattern in a similar procedure for each small block. Precipitation pattern change prediction method. 5. When calculating an area change from a plurality of past frames in calculating a one-dimensional area change rate for changing the diffusion coefficient, a predetermined condition is applied to time-series data of the past area change. Depending on the two-dimensional function and sin,
Approximate fitting to any of the non-linear functions including cos etc. by the least squares method, non-linear extrapolation about the change of the area is performed until a predetermined prediction time, and one-dimensional non-linear prediction of the area change is performed. The precipitation pattern change prediction method according to claim 4, including a step of changing corresponding diffusion coefficients in conjunction with each other. 6. The values of positive and negative regions when the difference of the precipitation pattern between consecutive frames is given to the source and suction terms of the advection / diffusion equation, respectively. Prediction method of precipitation pattern change described in. 7. The springing term at the prediction start time,
When the suction term is obtained from the image data, the source and suction image areas are moved corresponding to either the movement vector of the center of gravity or the estimated advection vector according to a predetermined condition for a predetermined prediction time. The precipitation pattern change prediction method according to claim 6, including a step. 8. The advection vector used in the advection-diffusion equation is obtained by extracting the change in the center of gravity of a region having a higher edge gradient of the gray value of the precipitation pattern. The maximum similarity based on the correlation method by giving either the direction or the interval of the maximum similarity found between image frames based on the change of the center of gravity extracted from the lower area and the cross-correlation method to the movement vector of the precipitation pattern. To calculate the direction and interval of points, one frame is divided into a plurality of small blocks, and the movement vector of the precipitation pattern is obtained by the cross-correlation method for each same small block between frames. The precipitation pattern change prediction method according to claim 3, including a step of estimating a movement vector of the precipitation pattern. 9. When estimating the advection vector in the image from the precipitation pattern, the contour line of the precipitation pattern is extracted at a certain interval, and the displacement of the contour line is set as a rough advection vector, and the moving average is calculated. In the case of using a movement vector estimated by the cross-correlation method, including a process of iteratively propagating an advection vector for every pixel in the image with the number of iterations increased or decreased in proportion to the predicted time length in the filter 9. The precipitation pattern change prediction method according to claim 8, further comprising: propagating the movement vector over the entire image by an iterative method using a plurality of movement vectors estimated from the first frames as initial vectors. 10. When extracting a contour line from a precipitation pattern of an image for estimating the advection vector, binarization, labeling processing, isolated point removal of the precipitation pattern including cloud of large and small clouds, After performing basic pre-processing including expansion / contraction processing, a specific number of the cloud lumps are extracted, and a contour line that is a boundary line of the cloud lumps is extracted by a boundary line search algorithm, and discrete. For each point sequence of the contour lines that have been converted, a search is performed along the normal direction of the contour line in the past frame until the contour line in the current frame intersects with the contour line between two frames. The precipitation pattern change prediction method according to claim 9, wherein the displacement of the precipitation pattern is calculated. 11. When a precipitation pattern moves between the sea and the ground during a target time for which a calculation for prediction is performed by using the advection-diffusion equation, the predetermined pattern is calculated between the sea and the ground. 10. The precipitation pattern change prediction method according to claim 1, wherein the size of the advection vector is increased / decreased by the value of, and at the same time, the size of the diffusion coefficient is increased / decreased to include the terrain effect in the calculation. . 12. When the precipitation pattern moves between the sea and the ground during the target time for which the calculation for prediction is performed using the advection-diffusion equation, the dissipation on the sea is higher than on the ground. The precipitation pattern change prediction method according to claim 11, wherein the coefficient is increased. 13. The advection-diffusion equation includes a process of performing discretization in a calculation space and solving the discretization method using a difference method, a finite element method, or a spectrum method according to a predetermined condition. 1. The precipitation pattern change prediction method according to 1. 14. When a calculation for prediction is performed using the advection / diffusion equation, an area of a precipitation pattern is subjected to an edge gradient (first differential value) by image processing from a grayscale distribution of the image.
Is divided by a threshold into a large area and a small area,
The precipitation pattern change according to claim 1, wherein a gravity center of a large area is obtained, a suction amount is weighted according to a distance from the gravity center, and the extinction coefficient is adjusted to increase as the distance increases. Prediction method. 15. When carrying out a calculation for prediction using the advection / diffusion equation, in addition to inputting the gray value of the precipitation pattern as a variable into the advection / diffusion equation, the monoxide emitted from the vehicle or factory The precipitation pattern change according to claim 1, wherein the concentration values of carbon, nitrogen oxides, and sulfur oxides are input to the advection / diffusion equation, and the influence of the concentration value of the chemical substance on the observation area is also applied to the prediction calculation. Prediction method. 16. When performing a calculation for prediction using the advection / diffusion equation, a time width finer than the observed time is selected when the advection / diffusion equation is time-integrated, and the time integration is performed. The precipitation pattern change prediction method according to claim 1, wherein the prediction result is obtained by repeating. 17. A method for predicting precipitation pattern changes,
A precipitation pattern change prediction apparatus having a calculation means, comprising: an image input unit for inputting a radar image of a rain region to be predicted, and an image for accumulating past time series images input by the image input unit. Accumulator and two consecutive images taken from the image accumulator for calculation
An image processing unit that calculates the amount of change in rainfall between two or more two-dimensional images and the image feature amount of the edge gradient distribution, and the terrain at the position where the predicted precipitation region exists is on the sea.
The advection vector is large and the diffusion coefficient is small.
Precipitated area that is set and stored so that it is less likely to be predicted in advance
If the terrain at the location of
Set the size to be smaller and the diffusion coefficient to be larger.
The terrain effect introductory part to be stored as a memory, and various image feature amounts calculated by the image processing part as initial values, the spatiotemporal variation of precipitation is regarded as a variable, and the diffusion stored in the terrain effect introductory part. An advection / diffusion equation calculation unit that performs numerical calculation based on the mathematical expression of the advection / diffusion equation using coefficients and advection vectors, and a prediction unit that predicts the precipitation pattern with the topographic effect based on the numerical calculation results. A precipitation pattern change prediction device, comprising: an output unit that outputs a precipitation area as a time-series image from the result predicted by the prediction unit.