Rainfall pattern Recognition / Classification using Neural Network

I am working on assignment titled "Pattern Recognition of Rainfall using ANN". For training and validation proposes, i have time-series rainfall data ranging from 1976 - 2006 of two metro-logical stations.. I have arranged the data in EXCEL from where i import and convert it into matrices . Input matrix contains 480 * 5 . where 480 cloums with 5 rows representing

1. Input= Year, Month, Longitude, Latitude and elevation
2. Target= matrix contains mean rainfall(mm) value of that month with size of 480 * 1.

My questions are:

• What Kind of Neural Network will best to deal with this type of problem?
• Are the above parameters of Inputs and Targets are in correct format and valid for accurate and precise results?

I have used "nprtool" i.e; PattternNet and "lvqnet" Learning vector quantization neural network for this porpose. But my respected teacher suggests me to improve the results as overall regression value (0.25) of mentioned network was not satisfactory.

Thanks for your time....!

Answer 1

[Greg Heath]

This is a timeseries problem. See the help and doc documentation for TIMEDELAYNET, NARNET & NARXNET

help narxnet
doc narxnet

Then search the NEWSGROUP and ANSWERS using

 NARXNET TUTORIAL
 NARNET  TUTORIAL
 TIMEDELAYNET TUTORIAL

Hope this helps.

Greg

[Muhammad Azeem]

Thanks for your suggestions.... NARXNET, TIMEDELAYNET and NARNET deals to predict one time series given past values of the same time series. And here i am working on "Pattern recognition / Classification"... How far these networks are helpful to recognize patterns?

[Greg Heath]

The form of your input/target data is that of an input/output time series (e.g., TIMEDELAYNET and NARXNET). Hopefully, auto and cross-correlations will be sufficient for you to predict the target given the input. How to turn that info into classification requires more information than you have submitted.

[Muhammad Azeem]

@Greg: I appreciate your suggestions. although auto and cross-correlations can be used for predict the target. but problem is remains same. What type of information is needed to turn this info into classification? Thanks for your time...

[Greg Heath]

You have to look at the target data and define which value intervals belong to which classes.

[Muhammad Azeem]

What Kind of Neural Network or other technique will best for this problem?