1
$\begingroup$

I'm running a model that uses Landsat spectral values as predictors in a support vector machine to predict percent canopy cover using field data.

The problem I have is that I only have ~70 points from the field representing canopy cover to use as training data. I have run the model using a 20+ band stack of predictors, but I am quite worried that the model will be overfit if I use this many predictor variables. The model results are satisfactory with this group of 20 predictors, but I don't know of a way to see predictor importance in SVM so I can provide fewer predictors to the model with those that are performing the best.

Should I be concerned that the model is overfit and reduce my number of predictors, or does the 1 for 10 rule of thumb not apply as much as it does for a linear model?

Any insight you could provide would be greatly appreciated!

$\endgroup$
0
$\begingroup$

I think your question is very interesting, but I am not familiar with "SVM model selection".

Generally, we can extract the features from the original data instead of using data directly. One way to do that is PCA (principle component analysis).

Compare with the Regression, the regression model selection provides several criterial like AIC, BIC to select the model. I think those criterial are essentially to tradeoff the fit performance and computation(also like overfit).

Hopefully, it can help you.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.