Melvin's digital garden

Groves2003

CREATED: 201004101554 LINK: url:/home/melvin/Modules/Literature/Groves2003.pdf

Propose a method to choose a set of bands and a classifier model to reduce classification error.

Idea is to rank bands according to their information content and evaluate the classification error of the top i bands at a time.

** Unsupervised methods (used for ranking)

  • Entropy
  • First spectral derivative, $D_1(\lambda_i) = \sum_x | I(x, \lambda_i) - I(x, lambda_{i+1}|)$
  • Second spectral derivative, $D_2(\lambda_i) = \sum_x |I(x, \lambda_{i-1}) - 2 I(x, \lambda_i) + I(x, \lambda_{i+1})|$
  • Contrast measure, $C(\lambda) = \sum_{i=1}^m |f_i - E(f)| * f_i$ where f is the histogram of all contrast values across one band computing using the Sobel edge detector, E(f) is the sample mean of f and m is the number of distinct contrast values
  • Spectral ratio measure
  • Correlation measure
  • Principal component analysis ranking
  • Spectral spacing

** Supervised classification methods

  • Naive Bayes
  • Instance based, weighted voting by k nearest neighbours
  • C4.5 decision tree,

Links to this note