JPH09318745A - Device and method for mutually correlational rainzone prediction - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speedily and easily predict the time series change of rainzone based on weather radar images by estimating a local displacement vector by way of an elastic equation and repeatedly moving the rainzone image near the starting point of the displacement vector whereon Kalman filter is operated to the direction of the displacement vector and for that distance. SOLUTION: An image input part 100 is input by radar rainzone images. Time series radar rainzone images are accumulated in an image accumulation part 110. A correlation energy calculating part 120 calculates, by wavelet- transforming the both images with local correlation coefficient between two frames, their energy value of inner product. A moving vector calculating part 130 estimates a local displacement vector by way of an elastic equation. A rainzone moving part 140 predicts the change of rainzone by repeatedly moving the rainzone image near the start point of the displacement vector whereon Kalman filter is operated to the direction of the displacement vector and for that distance.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention calculates a cross-correlation of two or more types of time-series data, and calculates a cross-correlation coefficient relating to a time-series change in a rainy area in a narrow area based on a weather radar image. The present invention relates to an apparatus and a method for predicting a rain region with high accuracy so that various economic activities can be smoothly performed in a narrow area that is sensitively affected by a change in weather by performing a short-term prediction.

[0002]

2. Description of the Related Art Conventionally, when performing weather forecasting, the Japan Meteorological Agency uses various physical parameters such as two-dimensional and three-dimensional temperature, pressure, dew point and wind vector obtained from AMeDAS and meteorological satellites. It predicts changes in meteorological phenomena across Japan over several tens of hours or days through physical equations.

On the other hand, in recent years, there has been a demand for a weather forecast in a narrower area in a short time on the order of minutes. Therefore, it is essential to make a forecast based on rain zone change information from a locally arranged weather radar device. Has become. The change of the rain area means the change of the area in the rainy state or the snowy state, and it looks like a cloud. The rain area as an image is expressed as a combination of various shapes and gray values, and it is difficult to clearly describe a feature amount. Therefore, it is not easy to detect the motion of the rain area based on an image processing method. Therefore, even now, subjective readings from changes in rainy areas based on human rules of thumb are performed and utilized.

In order to make the rain area change detection automatic and objective, it has been attempted to perform the rain area change detection on a computer by the spatiotemporal correlation method. In this method, after calculating the similarity in gray value between f (t) and f (t + 1) (t is a time step) between two image frames, it is assumed that the points having the highest correlation value correspond to each other. The amount of movement in the rain area is estimated above, and the rain area is translated in parallel based on the estimated amount of movement.

By repeating the estimated next step rain image, it is possible to predict up to a specified number of steps. Moreover, the movement vector of one rainy area is calculated from two frames.

[0006]

The spatio-temporal correlation method applied so far estimates one motion vector from two consecutive images. The spatiotemporal correlation method has a problem that the reliability of the correlation value becomes low as it is unless the object to be handled changes slightly between consecutive frames.

In other words, the gray value of the radar image has a one-to-one correspondence between the gray value of the previous frame and the gray value of the next frame because the rain zone is continuously generated and disappears. Therefore, there is a problem in that the reliability of the correlation value based on the gray value is not always high and the correlation value is unstable.

Further, in the case of a diffuse rainy area image for which a correlation coefficient cannot be obtained, the motion vector can hardly be estimated, and no interpolation method between them has been found at present.

In the actual rainy region, local flow exists in addition to the large flow of the whole. Since the conventional correlation method has been a method of obtaining only one representative vector from two consecutive frames, there is no information about the local flow, and the prediction accuracy is reduced accordingly.

If a local flow component is to be found, the image is divided into small blocks, and the spatial and spatial correlation methods are independently used for these small blocks to estimate the local component. However, it is difficult to calculate the motion component near the small block boundary, and there is a limit to the estimation of the local motion vector.

Furthermore, although the correlation coefficient itself is assumed to be a normal signal as a statistical signal to be handled, in the case of a radar image, the gray value used as a corresponding value is rarely a normal distribution. Is. Therefore, it is necessary to derive a more flexible correlation coefficient based on the assumption of non-normal distribution.

The present invention solves the above problems, and provides an apparatus and method for estimating a rain zone change from a radar image capable of stably estimating a more local motion vector and greatly reducing the calculation cost. Aim to achieve.

[0013]

A cross-correlation rain area prediction apparatus of the present invention is a cross-correlation rain area prediction apparatus for predicting a time series change of a rain area, which is a rainfall area, quickly and easily based on a weather radar image. Therefore, image input means for inputting a radar image in the rainy area, image storage means for storing the input image, and wavelet transform of both images by a local correlation coefficient between the two images, Correlation energy calculating means for calculating the energy value of the inner product of, the movement vector calculating means for estimating the local displacement vector via the elasticity equation, and the rain area image near the starting point of the displacement vector on which the Kalman filter is applied. It is characterized by further comprising: a rain area transfer means for predicting a change in the rain area by repeatedly moving the area by the direction of the displacement vector and its distance.

In this case, the rain zone moving means may include means for obtaining a displacement vector and then predicting the movement vector repeatedly with respect to the movement vector via a linear equation.

Further, in the correlation energy calculating means,
Gabor conversion and spline conversion may be included.

The cross-correlation rain area prediction method of the present invention is a cross-correlation rain area prediction method for quickly and easily predicting a time series change of a rain area which is a rainfall area based on a weather radar image. Wavelet transform is applied to both images by the local correlation coefficient between them, the energy value of the inner product of these is calculated, the local displacement vector is estimated through the elastic equation, and the displacement is applied by the Kalman filter. It is characterized by predicting a change in the rain area by repeatedly moving the rain area image near the start point of the vector in the direction of the displacement vector and its distance.

(Operation) When the cross-correlation coefficient between two frames is calculated, the statistical property of the signal of each frame is not necessarily a normal distribution. Therefore, when the formula for calculating the correlation coefficient is applied as it is. , The reliability of the calculated correlation coefficient is low. In the present invention, in order to improve its reliability and at the same time reduce the influence of signal disturbance due to noise, wavelet bases are applied in advance to both frames, so that the local property of this base is Thus, the coefficient value most suitable for the signal is calculated. At this time, a local displacement vector between frames can be calculated under the correlation coefficient energy with improved reliability by an image processing technique via an elasticity equation.

By repeatedly performing prediction through a linear equation or a Kalman filter, statistically the maximum likelihood estimated / predicted motion vector can be obtained for a long time, and highly accurate prediction can be achieved.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a diagram showing the configuration of an embodiment of the present invention. In the embodiment shown in FIG. 1, an image input unit 100 for inputting radar rain area images, an image storage unit 110 for storing time-series radar rain area images, and a local correlation coefficient between two frames are A correlation energy calculation unit 120 that performs a wavelet transform on both images to calculate the energy value of these inner products, a movement vector calculator 130 that estimates a local displacement vector via an elasticity equation, and a Kalman filter are operated. The rain zone image near the starting point of the displacement vector is repeatedly moved in the direction of the displacement vector and its distance, and the rain zone moving means 140 for predicting the change in the rain zone and the output section 150 for calculating the result. It consists of

FIG. 2 is a diagram showing the distribution of correlation energy between two consecutive frames.

Based on the local similarity between the two frames 200, the vector when the similarity is calculated between the frames via the elasticity equation is regarded as the local motion vector.

The applied elasticity equation is given as follows.

[0024]

[Equation 1] In the above equation, the left sides C ₁ and C ₂ of the equation (1) represent the elastic coefficient related to isotropic property and the elastic coefficient related to torsion, respectively. Indicates the external force exerted. Therefore, equation (1) represents the equilibrium equation for force.

Expression (2) represents the addresses of pixels in the x and y directions of the two-dimensional images, and Expression (3) represents the energy gradient value related to the correlation coefficient (similarity) expressed by Expression (4). Indicates that it is equal to the external force of (1).

The expression (5) is a discretized expression of the expression (1) according to the difference method, and in particular, the expression (1) includes an approximation term obtained by Taylor-expanding the expression (3) to a second order. Equation (6)
Represents the similarity (normalized correlation coefficient) of Expression (4), and is the normalized correlation coefficient in the local crowd of images of two consecutive frames t and s. This size is 3
It is determined from about × 3 to about 21 × 21 depending on the size of the input image and the size of the motion per target frame.

Equation (5) is a simultaneous linear equation concerning the local image size. Therefore, a solution can be obtained while accelerating the convergence according to the acceleration relaxation method such as the SOR method. The solution to be obtained is the equation (2), and this solution becomes a local displacement vector between two frames. In principle, after calculating the potential field related to the similarity between two frames, the slope of the potential peak, which is a scalar quantity, that is, the vector quantity becomes the displacement vector.

By the way, even if the correlation coefficient of the equation (6) is applied to the image as it is, a correct result is output to some extent.
Radar images do not always output stably because the generation and disappearance are changed gently between frames. Therefore, assuming that the gray value of the image changes due to noise, a method of suppressing the influence of such noise as much as possible is essential. As one of the measures, as shown in Expression (7),
After simplifying equation (6), a wavelet base is applied to each frame s, t as in equation (8).
In equations (7) and (8), <,> are inner products, uppercase letters T, S
Represents the Fourier transform.

[0029]

[Outside 1] Represents convolution.

[0030]

[Equation 2] By applying a wavelet basis to each image in advance as in Expression (8), it is possible to locally suppress noise and enhance the local gray value of the image. The reliability of the correlation coefficient is inevitably higher when the equation (8) is used than when the equation (6) is used. From these, a graph 210 showing a state in which the correlation value U of two frames is taken in the height direction and a graph 220 showing the gradient value and the direction two-dimensionally are obtained.

Various kinds of wavelet bases such as those by Gabor transformation or spline transformation are conceivable, but a function convenient for time-frequency analysis may be selected.
There is no particular limitation.

FIG. 3 is a diagram showing an example of a wavelet base. FIG. 3A shows the scaling function φ (x) of the cubic spline basis, and FIG. 3B shows the mother basis ψ (x). The mother basis ψ (x) is derived from the scaling function φ (x). When calculating the correlation coefficient, such a mother basis ψ (x) is applied to the image.

FIG. 4 shows that the rain area is translated in the next frame according to the local displacement vector of the image. In the figure, it is assumed that the displacement vector from the white circle to the black circle is obtained. At this time, the gray value (rain area) near the white circle is moved as it is according to the displacement vector. Only when this process is repeated, the rain area can be predicted.

FIG. 5 is a diagram showing changes in the displacement vector when the Kalman filter is applied, and FIG. 5A shows the three local vectors (V1 to V3) that have been subjected to the maximum likelihood estimation by the Kalman filter. Show changes. In the case of the conventional cross-correlation method, only one motion vector is obtained from two frames, but according to the present invention, a plurality of motion vectors, that is, local motion vectors can be estimated.

FIG. 5 (b) is a diagram corresponding to FIG. 5 (a), in which the difference in prediction error is compared only in two cases, and the wavelet base is not applied (W).
(No T) and a Kalman filter (KF + WT) with a wavelet base acting.

From FIG. 5 (b), it can be seen that the post-prediction is reduced by operating the wavelet base and the Kalman filter.

In the present invention, the Kalman filter is used when predicting the repeated displacement vector. Also,
When calculating the cross-correlation coefficient, the wavelet base is applied to two frames in advance. The difference between the effects of the prediction error with and without the application of the wavelet base becomes clear with time. At this time, the Kalman filter has an action of reducing noise included in the displacement vector. Therefore, it can be said that the method that uses the wavelet base and the Kalman filter together can predict the rain area with maximum likelihood in a stable and accurate manner.

An outline of the maximum likelihood estimation / prediction method of the runoff angle in the rain area by the Kalman filter is shown below.

Here, with respect to the local displacement vector in the rain region, the acceleration change is constant. This indicates that the time series changes in the rainy area show relatively moderate changes. The updated position (movement vector, angle) is the current position

[0040]

[Outside 2] As a Taylor expansion approximate expression for

[0041]

[Outside 3] It is expressed as a function having a variable as follows (Equations 9 and 10). It is assumed that the system dynamics regarding the movement vector and the angle can be described by a state space equation such as equation (11). Equation (12) is an observation equation.

[0042]

(Equation 3) The matrices Φ, Γ, and H are the state transition
Since it is an observation matrix and a linear equation is derived,
The coefficients of equations (11) and (12) are uniquely determined.

All of the above matrices are stable,
Moreover, it is clear that the {Φ, Γ} and {Φ, H} pairs are controllable and observable. In step k,

[0044]

[Outside 4] Represents a two-dimensional position and velocity state vector,

[0045]

[Outside 5] Represents an observation vector related to the position.

Vector

[0047]

[Outside 6] Is the Gaussian / random acceleration component,

[0048]

[Outside 7] Is the random observation noise. It is assumed that these two vectors are independent of each other. Q and R are marked with the Kronecker symbol, vector

[0049]

[Outside 8] Represents the covariance matrix of. E [•] represents an ensemble average.

The algorithm of the stationary Kalman filter is expressed by the equations (13) and (14), and the maximum likelihood estimation value and the maximum likelihood prediction value are statistically obtained from each of them. The matrix included in the following equations is uniquely determined from the state space equation and the observation equation.

[0051]

(Equation 4)

[0052]

As described above, according to the present invention,
When the cross-correlation coefficient is calculated, the wavelet base is applied to the two images in advance, so that the noise is locally suppressed when the change in the density of the rain area is regarded as noise, and at the same time, the output value of the correlation coefficient Improves reliability. as a result,
The accuracy of the local displacement vector in the rain area is improved.

Further, the Kalman filter is used for the displacement vector, which has the effect of predicting the rainfall region continuously and accurately even when the rain region is in a diffused state.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an exemplary embodiment of the present invention.

FIG. 2 is a diagram showing the distribution of correlation energy between two consecutive frames.

FIG. 3 is a diagram showing an example of a wavelet basis,
(A) is a scaling function φ of cubic spline basis
(X) is shown, (b) shows mother basis (psi) (x).

FIG. 4 is a diagram showing local parallel movement of a rain area.

FIG. 5 is a diagram showing a change of a displacement vector when a Kalman filter is applied. FIG.
V3) is shown, and (b) is a diagram in which the difference in prediction error corresponding to (a) is compared only in two cases.

[Explanation of symbols]

 100 Image Input Unit 110 Image Storage Unit 120 Correlation Energy Calculation Unit 130 Movement Vector Calculation Unit 140 Rain Area Movement Unit 150 Output Unit

Front page continued (72) Inventor Hideto Suzuki 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo Inside Nippon Telegraph and Telephone Corporation

Claims (4)

[Claims] 1. A cross-correlation rain area prediction device for quickly and easily predicting a time-series change of a rain area, which is a rainfall area, based on a weather radar image, comprising image input means for inputting a radar image of the rain area. , An image accumulating means for accumulating the input images, a correlation energy calculating means for performing a wavelet transform on both images by a local correlation coefficient between the two images, and calculating an energy value of an inner product of these images, A movement vector calculating means for estimating a local displacement vector through an equation, and a rainwater image near the starting point of the displacement vector, which is operated by a Kalman filter, is repeatedly moved to the direction of the displacement vector and its distance. Therefore, a cross-correlation rain area prediction device comprising: a rain area transfer means for predicting a change in rain area. 2. The rain zone moving means, after obtaining the displacement vector, predicts the repeated movement vector with respect to the movement vector via a linear equation.
The cross-correlation rain area prediction device described in. 3. The correlation energy calculating means is characterized by performing Gabor conversion and spline conversion.
The cross-correlation rain area prediction device described in. 4. A cross-correlation rain area prediction method for quickly and easily predicting a time-series change of a rain area, which is a rainfall area, based on a weather radar image, the method comprising a local correlation coefficient between two images. Wavelet transform is applied to both images, the energy value of these inner products is calculated, the local displacement vector is estimated through the elastic equation, and the Kalman filter is applied to the rain area image near the starting point of the displacement vector. A cross-correlation rain area prediction method, characterized in that rain area changes are predicted by repeatedly moving to the direction of the displacement vector and its distance.