0

MLP in LCM Idrisi Selva 17.00

rumawan

August 20, 2014 10:06

None

Follow

Hello

1. I use LCM in idrisi selva 17:00. I have a variable distance drivers to the road where the closer the value is 0, while the background of the map is worth 0. How to overcome this?

1. MLP on LCM work as regression or classifier?
   
2. when to test Cramer "s V likelihood evidence of residential, why land use class residential have P value = 1.000?

Why are the results of the MLP accuracy rate is high but  prediction accuracy of the results is low and the MLP accuracy rate is low but prediction accuracy of the results is high ?

We are sorry to many question. pleas help me to understanding this

 

thank you 

• Facebook
• Twitter
• LinkedIn

1 comment

• 0
  
  Clark Labs August 25, 2014 13:29

  Hi,

  If you have background, this is reflected in your land cover map as zero and background, and not part of the analysis. For distances of zero within non-background areas, they would be treated correctly.

  MLP in this instance is closer to the classifier concept. As used in LCM it is the same as the stand alone classifier with emphasis on the activation levels.

  Using the Crammer's V is not recommended given the sample size and spatial dependence issues. But a p-value of 1.0 more or less assumes you have an insignificant variable. Since you are using MLP, we strongly recommend using the XML output to assess the quality of your driver variables. It gives you more explicit information on accuracy and skills. 

  As for the skills measure, it is a more accurate predictor. The accuracy rate takes in to account both change and persistence. The skill measure is accuracy minus accuracy of change. So you could have a high accuracy, because of you the models skill at predicting persistence. But with an accompanying lower skill, then you have a potential poorly predicting variable. You may need to rethink your drivers and remove or add new drivers accordingly.

  You can find details of all these concepts in the Help system for each of these topics.

  James Toledano

  Clark Labs