0

Use of Evidence Likelihood transformation

zushika

June 12, 2017 02:08

None

Follow

Hello,

I am going through the tutorial EXERCISE 4-2 of LCM and have a question on the use of data transformed by Evidence Likelihood transformation as a driver variable in LCM (esp. page 324). I understood that Evidence Likelihood transformation derives relative frequency with which different land cover categories occurred within the areas that transitioned from 1986 to 1994. The, we use it as one of driver variables to train MLP Neural Network submodel.

My colleague mentioned that it does not seem to be appropriate to include variable transformed with Evidence Likelihood transformation in the training of MLP Neural Network Sub-model because the transformed variable is similar to what the Sub-model is trying to predict, and looks tautological process. 

I found a couple of journal articles that utilized LCM with a transformed driver variable through Evidence Likelihood transformation, which suggested the process is methodologically acceptable. However, the point my colleague raised sounds also reasonably.

I am still looking for the scientific explanation that justifies the use of a driver variable, transformed with Evidence Likelihood transformation, in the training of MLP Neural Network Sub-model.

 

 

 

 

• Facebook
• Twitter
• LinkedIn

2 comments

• 0
  
  Clark Labs June 16, 2017 19:13

  This is an interesting question. MLP is modeling transitions, and not land covers (in fact, the land cover prediction part is achieved with a separate allocation procedure in LCM). The output from MLP is a set of transition potential maps – one for each transition being modeled. It models on the basis of samples of pixels that went through the transition in question and samples that were eligible to go through the transition, but did not. Note also that the modeled transition potentials are masked to force all pixels that are not in the correct “from” class to have a transition potential of 0. For example, if the transition being modeled is forest to cropland, any pixel that is not in forest in the later land cover image will have a transition potential of 0. Thus each transition potential model knows very explicitly which land cover it is coming from and expresses the potential of transitioning to a single land cover in the future.

   

  Consider again the example of modeling the transition from forest to cropland. If you model your transitions separately (e.g., forest to cropland independently of other transitions), then an evidence likelihood map based on land cover will have no value since EVERY pixel that is used for training will have the same evidence likelihood value. You will notice that adding this variable will have no impact on the skill – it has no useful information.

   

  However, if you create a sub-model with several transitions that involve multiple “from” classes (e.g., forest to cropland and grassland to cropland), then you may notice a difference. The evidence likelihood values will most likely differ between the different “from” classes. This is information that may prove useful both independently and in interaction with other variables (MLP is very powerful in modeling interactions by means of its hidden layer).

   

  Ron Eastman

• 0
  
  zushika June 16, 2017 22:48

  Thank you very much for the detailed explanation.  I understood that Evidence Likelihood transformation does not produce tautological data in training MLP sub-models and it is effective in modeling land use change from multiple classes.