Weather forecast apparatus and method based on recognition of echo patterns of radar images

• US 5,796,611 A
• Filed: 10/03/1995
• Issued: 08/18/1998
• Est. Priority Date: 10/04/1994
• Status: Expired due to Fees

• Current Assignee: Nippon Telegraph and Telephone Corporation
• Original Assignee: Nippon Telegraph and Telephone Corporation
• Inventors: Suzuki, Hideto, Ochiai, Keihiro, Sonehara, Noboru
• Primary Examiner(s): Cosimano, Edward R.
• Application Number:
  US08/538,723
• Time in Patent Office: 1,050 Days
• Field of Search: 73/170.16, 340/601, 364/420, 342/26, 395/1, 395/10, 395/20, 395/21, 395/23, 395/50
• US Class Current: 702/3
• CPC Class Codes: G01S 13/95 for meteorological use
• CPC Class Codes: G01S 7/417 involving the use of neural networks
• CPC Class Codes: G01W 1/10 Devices for predicting weather conditions computers per se G06; display devices G09
• CPC Class Codes: Y02A 90/10 Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
• CPC Class Codes: Y10S 706/931 Weather

Abstract

The present invention provides a weather forecast apparatus and a method for the same, to systematically classify a measured radar image based on results of pattern classification of past radar images so as to use the classified radar image. In the present invention, rapid forecasting is possible by making the FNN model previously learn based on data of each class (and indexes for forecast times) obtained by classification of past weather data for every resembling pattern. In addition, a calculation procedure for improving the classifying ability of patterns can be established by varying the procedure for calculating feature quantities with regard to the radar image by using the learning of the TNN model. Furthermore, systematic classification of a pre-learned image can be realized by performing self organization with regard to compound feature quantities extracted from a radar image in the PNN model, a typical example of which is a competitive learning model.

Claims

1. A weather forecast apparatus that provides a weather radar image to an FNN model, which is a neural network model for learning and forecasting, so that the FNN model learns weather dynamics and forecasts weather after the learning, comprising:

pattern recognition means, including a PNN model, which is a neural network model for pattern classification, and is prepared based on classification criteria according to past radar images, for pattern-recognizing a measured radar image by providing the radar image to the PNN model to pattern-classify the radar image;

memory means for storing plural weights for the FNN model, which are determined based on past plural radar images, the memory means having clusters corresponding to each past radar image;

learning means, including the FNN model, for selecting a weight that corresponds to a pattern resembling the pattern of the measured radar image from among those stored in the memory means in accordance with the pattern-recognized result obtained by the pattern recognition means;

setting the selected weight as an initial value of the FNN model; and

making the FNN model re-learn; and

forecast means for fixing the weights of the FNN model after the re-learning, and for forecasting and outputting a future radar image,

wherein the learning means includes means for causing the FNN model to re-learn by providing the radar image, which was pattern-classified by the pattern recognition means, to the FNN model when the radar image belongs to an existing cluster, while generating a new cluster and weights which are most suitable for forecasting a weather pattern belonging to the new cluster when the radar image does not belong to an existing cluster.

2. A weather forecast apparatus as claimed in claim 1, wherein:

the memory means comprises a database for registering and managing the weights for the FNN model, which are generated by the learning means based on the pattern-recognized result obtained by the pattern recognition means, and the radar images used for the learning, and the size of the database is automatically enlarged or reduced by adding a cluster or unifying the clusters.

3. A weather forecast apparatus as claimed in any one of claims 1 or 2, further comprising:

means for providing indexes representing plural forecast times for the FNN model and making the FNN model learn so that the FNN model could output each radar image which corresponds to each time represented by each index; and

means for outputting a radar image corresponding to any possible forecast time obtained by any combination of the plural forecast times, for providing radar images that correspond to each forecast time comprising the combination for the FNN model in sequence.

4. A weather forecast method for providing a weather radar image to an FNN model, which is a neural network model for learning and forecasting, so that the model learns weather dynamics and forecasts weather by using the FNN model after the learning, the method performing pattern recognition of the radar image by using a TNN model that is a neural network model for transforming feature quantities which were input into the TNN model, comprising the steps of:

calculating and extracting one or more feature quantities from a radar image;

providing the feature quantities extracted in the calculating and extracting step for the TNN model as input and teacher data, and learning identical transformation from the feature quantities to the feature quantities;

calculating an estimated value based on a criterion for recognition from the output feature quantities that were identically transformed and output from the TNN model;

adding perturbation to the output feature quantities;

performing re-learning of the TNN model such that the estimated value is maximized or minimized;

pattern-recognizing a pre-learned image by providing feature quantities calculated and extracted from the pre-learned image for the TNN model after the re-learning; and

forecasting and outputting a future radar image based on result of the pattern recognition of the pre-learned image.

5. A weather forecast method as claimed in claim 4, wherein the criterion is such as to estimate a cluster of the radar image based on a covariance matrix.

8. A weather forecast method as claimed in one of claims 4 or 6, wherein the feature quantities are selected from the group consisting of energy, entropy, correlation, local homogeneity, and inertia based on a co-occurrence matrix;

contrast, angular second moment, entropy, mean, and inverse difference moment based on a GLDM; and

short run emphasis, long run emphasis, gray level distribution, run length distribution, and run percentage based on a GLRLM.

6. A weather forecast method for providing a weather radar image to an FNN model, which is a neural network model for learning and forecasting, so that the model learns weather dynamics and forecasts weather by using the FNN model after the learning, the method performing systematic classification of the radar image by using a PNN model, which is a neural network model for pattern classification, and selects reference vectors based on input feature quantities, comprising the steps of:

calculating and extracting plural feature quantities from the radar image;

providing the feature quantities extracted in the calculating and extracting step for the PNN model, and performing learning of the PNN model by finding a minimum value of an object function such that the quantization error of each reference vector which belongs to each output unit of the PNN model is minimized;

classifying a pre-learned image by providing the feature quantities calculated and extracted from the pre-learned image for the PNN model after the learning; and

forecasting and outputting a future radar image based on results of the classification of the pre-learned image.

7. A weather forecast method as claimed in claim 6, wherein the PNN model is a competitive learning model.

9. A weather forecast apparatus for providing a weather radar image to an FNN model, which is a neural network model for learning and forecasting, so that the model learns weather dynamics and forecasts weather by using the FNN model after the learning, comprising:

feature quantity calculating means for calculating and extracting one or more feature quantities from a radar image;

a TNN model, which is a neural network model, for transforming and outputting the feature quantities that are input from the feature quantity calculating means;

perturbation generating means for generating an amount of perturbation;

means for calculating an estimated value from the feature quantities output from the TNN model, based on a criterion for recognition;

learning means for providing the feature quantities extracted by the feature quantity calculating means for the TNN model as input and teacher data, for performing identical transformation from the feature quantities to the feature quantities, for adding the amount of perturbation to the output feature quantities after the transformation, and for performing re-learning of the TNN model such that the estimated value is maximized or minimized;

recognizing means for pattern-recognizing a pre-learned image according to the estimated value that is obtained when feature quantities calculated and extracted from the pre-learned image are input into the TNN model after the re-learning; and

forecast means for forecasting and outputting a future radar image based on result of the pattern recognition of the pre-learned image.

10. A weather forecast apparatus for providing a weather radar image to an FNN model, which is a network model for learning and forecasting, so that the model learns weather dynamics and forecasts weather by using the FNN model after the learning, comprising:

feature quantity calculating means for calculating and extracting plural feature quantities from a radar image;

a PNN model, which is a neural network model for pattern classification, for selecting reference vectors based on the feature quantities that are input from the feature quantity calculating means;

learning means for providing the feature quantities extracted by the calculating and extracting means for the PNN model, and for performing learning of the PNN model by finding a minimum value of an object function such that the quantization error of each reference vector which belong to each output unit of the PNN model is minimized;

classifying means for classifying a class of the image according to the reference vectors that were selected by the PNN model, wherein systematic classification of a pre-learned image is performed by inputting the feature quantities calculated and extracted;

from the pre-learned image into the PNN model after the learning and

forecast means for forecasting and outputting a future radar image based on result of the classification of the pre-learned image.

11. A weather forecast apparatus as claimed in one of claims 9 or 10, wherein the feature quantities are selected from the group consisting of energy, entropy, correlation, local homogeneity, and inertia based on a co-occurrence matrix;

contrast, angular second moment, entropy;

mean, and inverse difference moment based on a GLDM; and

short run emphasis, long run emphasis, gray level distribution, run length distribution, and run percentage based on a GLRLM.

12. A weather forecast apparatus as claimed in claim 10, wherein the PNN model is a competitive learning model.