pktools 2.6.7
Processing Kernel for geospatial data
pkfsann

feature selection for artificial neural network classifier pkann

SYNOPSIS

Usage: pkfsann -t training -n number

Options:

Advanced options:

Description

Classification problems dealing with high dimensional input data can be challenging due to the Hughes phenomenon. Hyperspectral data, for instance, can have hundreds of spectral bands and require special attention when being classified. In particular when limited training data are available, the classification of such data can be problematic without reducing the dimension.

The utility pkfsann implements a number of feature selection techniques, among which a sequential floating forward search (SFFS). Also consider the SVM classifier implemented in pksvm, which has been shown to be more robust to this type of problem than others.

Options

  • use either -short or --long options (both --long=value and --long value are supported)
  • short option -h shows basic options only, long option --help shows all options
    short long type default description
    i input std::string input test set (leave empty to perform a cross validation based on training only)
    t training std::string training vector file. A single vector file contains all training features (must be set as: B0, B1, B2,...) for all classes (class numbers identified by label option). Use multiple training files for bootstrap aggregation (alternative to the bag and bsize options, where a random subset is taken from a single training file)
    n nf unsigned short 0 number of features to select (0 to select optimal number, see also ecost option)
    tln tln std::string training layer name(s)
    label label std::string label identifier for class label in training vector file.
    balance unsigned int 0 balance the input data to this number of samples for each class
    random random bool true in case of balance, randomize input data
    min min int 0 if number of training pixels is less then min, do not take this class into account
    b band unsigned short band index (starting from 0, either use band option or use start to end)
    sband startband unsigned short Start band sequence number
    eband endband unsigned short End band sequence number
    offset offset double 0 offset value for each spectral band input features: refl[band]=(DN[band]-offset[band])/scale[band]
    scale scale double 0 scale value for each spectral band input features: refl=(DN[band]-offset[band])/scale[band] (use 0 if scale min and max in each band to -1.0 and 1.0)
    a aggreg unsigned short 0 how to combine aggregated classifiers, see also rc option (0: no aggregation, 1: sum rule, 2: max rule).
    sm sm std::string sffs feature selection method (sffs=sequential floating forward search,sfs=sequential forward search, sbs, sequential backward search ,bfs=brute force search)
    ecost ecost float 0.001 epsilon for stopping criterion in cost function to determine optimal number of features
    cv cv unsigned short 2 n-fold cross validation mode
    c class std::string list of class names.
    r reclass short list of class values (use same order as in classname opt.
    n nneuron unsigned int 5 number of neurons in hidden layers in neural network (multiple hidden layers are set by defining multiple number of neurons: -n 15 -n 1, default is one hidden layer with 5 neurons)
    connection float 1 connection reate (default: 1.0 for a fully connected network)
    w weights float 0 weights for neural network. Apply to fully connected network only, starting from first input neuron to last output neuron, including the bias neurons (last neuron in each but last layer)
    l learning float 0.7 learning rate (default: 0.7)
    maxit unsigned int 500 number of maximum iterations (epoch) (default: 500)

Usage: pkfsann -t training -n number