Source code for libpyhat.transform.dim_reductions.mnf

import numpy as np
import pandas as pd
from pysptools.noise import Whiten
from sklearn.decomposition import PCA


# adapted from pysptools to take n_components


def mnf(M, n_components):
    w = Whiten()
    wdata = w.apply(M)
    h, w, numBands = wdata.shape
    X = np.reshape(wdata, (w * h, numBands))
    pca = PCA(n_components=n_components)
    mnf = pca.fit_transform(X)
    mnf = np.reshape(mnf, (h, w, n_components))
    return mnf





class MNF:
    def __init__(self, n_components=4):
        self.n_components = n_components



    def fit_transform(self, data):
        """
        Description: Minimum Noise Fraction (MNF) wrapper for pysptools
        implementation
        Rationale: Removes noise while preserving information
        """
        # Convert Series to ndarray
        if issubclass(type(data), pd.Series) or issubclass(type(data), pd.DataFrame):
            np_data = data.to_numpy()
        elif issubclass(type(data), np.ndarray):
            np_data = data
        else:
            raise ValueError("Input for MNF must inherit from pd.Series or np.ndarray")

        cube_data, num_dimensions = make_3d(np_data)

        # check data shape
        # if cube_data.shape[0]*cube_data.shape[1]>cube_data.shape[2]:

        mnf_result = mnf(cube_data, self.n_components)

        # Return result in dimensionality of input
        if num_dimensions == 2:
            return np.squeeze(mnf_result)
        else:
            return mnf_result


        # else:
        #     raise ValueError("For MNF, # of samples must be greater than #
        #     of spectral channels")
        #




def make_3d(data):
    # Ensure 3 dimensional input for MNF
    num_dimensions = len(data.shape)
    if num_dimensions == 2:
        # Add arbitrary 3rd dimension
        cube_data = np.expand_dims(data, axis=0)
    elif num_dimensions == 3:
        cube_data = data
    else:
        raise ValueError("Input must be 2 or 3 dimensional")
    return cube_data, num_dimensions