def gmrx_anomaly(hsi_img, n_comp, mask=None):
"""
Gaussian Mixture RX Anomaly Detector
fits GMM assuming entire image is background
assigns pixels to highest posterior probability mixture component
computes pixel Mahlanobis distance to component mean
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
n_comp - number of Gaussian components to use
Outputs:
gmrx_out - detector output image
8/7/2012 - Taylor C. Glenn - [email protected]
5/5/2018 - Edited by Alina Zare
11/2018 - Python Implementation by Yutai Zhou
"""
gmm_out, kwargsout = img_det(gmrx_helper,
hsi_img,
None,
mask,
n_comp=n_comp)
return gmm_out
def ace_rt_max_detector(hsi_img, tgt_sig, mask = None, mu = None, sig_inv = None): """ Adaptive Cosine/Coherence Estimator given Multiple Target Signatures. Confidence value is the max ace score over all target signatures. Inputs: hsi_image - n_row x n_col x n_band hyperspectral image tgt_sig - target signature (n_band x n_sig - column vector) mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used mu - background mean (n_band x 1 column vector) siginv - background inverse covariance (n_band x n_band matrix) Outputs: ace_out - detector image mu - mean of input data siginv - inverse covariance of input data 8/8/2012 - Taylor C. Glenn 6/2/2018 - Edited by Alina Zare 11/2018 - Python Implementation by Yutai Zhou """ if tgt_sig.ndim == 1: tgt_sig = tgt_sig[:, np.newaxis] ace_rt_max_out, kwargsout = img_det(ace_rt_max_helper, hsi_img, tgt_sig, mask, mu = mu, sig_inv = sig_inv) return ace_rt_max_out, kwargsout['mu'], kwargsout['sig_inv']
def ccmf_detector(hsi_img, tgt_sig, mask = None, n_comp = 5, gmm = None): """ Class Conditional Matched Filters inputs: hsi_image - n_row x n_col x n_band hyperspectral image tgt_sig - target signature (n_band x 1 - column vector) mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used n_comp - number of Gaussian components to use gmm - optional mixture model structure from previous training data outputs: ccmf_out - detector image gmm - mixture model learned from input image 8/8/2012 - Taylor C. Glenn 6/2/2018 - Edited by Alina Zare 12/2018 - Python Implementation by Yutai Zhou """ if tgt_sig.ndim == 1: tgt_sig = tgt_sig[:, np.newaxis] ccmf_out, kwargsout = img_det(ccmf_helper, hsi_img, tgt_sig, mask, n_comp = n_comp, gmm = gmm) return ccmf_out, kwargsout['gmm']
def cem_detector(hsi_img, tgt_sig, mask=None):
"""
Constrained Energy Minimization Detector
solution to filter with minimum energy projected into background space
Ref: J. C. Harsanyi, ?Detection and classification of subpixel spectral signatures in hyperspectral image sequences,? Ph.D. dissertation, University of Maryland Baltimore County, 1993.
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sigs - target signatures (n_band x n_sigs)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
cem_out - detector image
w - cem filter
8/8/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
cem_out, kwargsout = img_det(cem_helper, hsi_img, tgt_sig, mask)
return cem_out, kwargsout['w']
def palm_detector(hsi_img, tgt_sig, mask=None, n_comp=5):
"""
Pairwise Adaptive Linear Matched Filter
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
n_comp - number of Gaussian components to use
Outputs:
palm_out - detector image
8/8/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
palm_out, kwargsout = img_det(palm_helper,
hsi_img,
tgt_sig,
mask,
n_comp=n_comp)
return palm_out
def smf_max_detector(hsi_img, tgt_sig, mask=None, mu=None, sig_inv=None):
"""
Spectral Matched Filter, Max over targets
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sigs - target signatures (n_band x n_signatures)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
smf_out - detector image
8/8/2012 - Taylor C. Glenn
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
n_sig = tgt_sig.shape[1]
n_row, n_col, n_band = hsi_img.shape
sig_out = np.zeros((n_row, n_col, n_sig))
for i in range(n_sig):
sig_out[:, :, i], kwargsout = img_det(smf_det_array_helper,
hsi_img,
tgt_sig[:, i][:, np.newaxis],
mask,
mu=mu,
sig_inv=mu)
smf_out = np.max(sig_out, 2)
return smf_out
def spsmf_detector(hsi_img, tgt_sig, mask = None, mu = None, sig_inv = None): """ Subpixel Spectral Matched Filter matched filter derived from a subpixel mixing model H0: x = b H1: x = alpha*s + beta*b Ref: formulation from Eismann's book, pp 664 Inputs: hsi_image - n_row x n_col x n_band hyperspectral image tgt_sig - target signature (n_band x 1 - column vector) mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used mu - background mean (n_band x 1 column vector) siginv - background inverse covariance (n_band x n_band matrix) Outputs: spsmf_out - detector image 8/8/2012 - Taylor C. Glenn 6/2/2018 - Edited by Alina Zare 12/2018 - Python Implementation by Yutai Zhou """ if tgt_sig.ndim == 1: tgt_sig = tgt_sig[:, np.newaxis] spsmf_out, kwargsout = img_det(spsmf_helper, hsi_img, tgt_sig, mask, mu = mu, sig_inv = sig_inv) return spsmf_out
def ace_ss_detector(hsi_img, tgt_sig, mask=None, mu=None, sig_inv=None):
"""
Adaptive Cosine/Coherence Estimator - Subspace Formulation
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sigs - target signatures (n_band x M - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
ace_ss_out - detector image
8/8/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
11/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
ace_ss_out, kwargsout = img_det(ace_ss_helper,
hsi_img,
tgt_sig,
mask,
mu=mu,
sig_inv=sig_inv)
return ace_ss_out
def ha_detector(hsi_img, tgt_sig, ems, mask=None, n_comp=2):
"""
Hybrid Abundance Detector
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
ems - background endmembers
n_comp - number of mixture components for abundance mixtures
Outputs:
ha_out - detector image
8/27/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
ha_out, kwargsout = img_det(ha_helper,
hsi_img,
tgt_sig,
mask,
ems=ems,
n_comp=n_comp)
return ha_out
def fcbad_anomaly(hsi_img, n_cluster, mask=None):
"""
Fuzzy Cluster Based Anomaly Detection (FCBAD)
Ref: Hytla, Patrick C., et al. "Anomaly detection in hyperspectral imagery: comparison of methods using diurnal and seasonal data." Journal of Applied Remote Sensing 3.1 (2009): 033546
This algorithm requires skfuzz! https://pythonhosted.org/scikit-fuzzy/install.html
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
n_cluster - number of clusters to use
Outputs:
fcbad_out - detector output image
cluster_img - cluster label image
8/8/2012 - Taylor C. Glenn
5/5/2018 - Edited by Alina Zare
11/2018 - Python Implementation by Yutai Zhou
"""
fcbad_out, kwargsout = img_det(fcbad_out_helper,
hsi_img,
None,
mask,
n_cluster=n_cluster)
cluster_img = kwargsout['idx']
return fcbad_out, cluster_img
def mtmf_statistic(hsi_img, tgt_sig, mask=None):
"""
Mixture Tuned Matched Filter Infeasibility Statistic
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
mtmf_out - MTMF infeasibility statistic
alpha - matched filter output
8/12/2012 - Taylor C. Glenn - [email protected]
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
mnf_img, n_dim, mnf_vecs, mnf_eigvals, mnf_mu = mnf(hsi_img, 1)
# tgt_sig = tgt_sig[:n_dim,0][:,np.newaxis]
s = mnf_vecs @ (tgt_sig - mnf_mu)
mtmf_out, kwargsout = img_det(mtmf_helper,
mnf_img,
s,
mnf_eigvals=mnf_eigvals)
return mtmf_out, kwargsout['alpha']
def knn_classifier(hsi_img, train_data, K, mask = None): """ A simple K nearest neighbors classifier Inputs: hsi_img - hyperspectral data cube (n_rows x n_cols x n_bands) train_data - numpy void structure containing training data train_data['Spectra'][0, i]: matrix containing training data from class i train_data['name'][0, i]: matrix containing name of class i mask - binary image indicating where to apply classifier K - number of neighbors to use during classification 10/31/2012 - Taylor C. Glenn 05/12/2018 - Edited by Alina Zare 10/2018 - Python Implementation by Yutai Zhou """ knn_out, kwargsout = img_det(knn_cfr, hsi_img, train_data, mask = mask, K = K); return knn_out
def md_anomaly(hsi_img, mask=None):
"""
Mahalanobis Distance anomaly detector
uses global image mean and covariance as background estimates
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
dist_img - detector output image
8/7/2012 - Taylor C. Glenn
5/5/2018 - Edited by Alina Zare
11/2018 - Python Implementation by Yutai Zhou
"""
dist_img, kwargsout = img_det(md_helper, hsi_img, None, mask)
return dist_img
def beta_anomaly(hsi_img, mask):
"""
Beta Distribution Anomaly Detector
fits beta distribution to each band assuming entire image is background
computes negative log likelihood of each pixel in the model
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
beta_out - detector output image
8/24/2012 - Taylor C. Glenn
5/5/2018 - Edited by Alina Zare
11/2018 - Python Implementation by Yutai Zhou
"""
beta_out, kwargsout = img_det(beta_helper, hsi_img, None, mask)
return beta_out
def gmm_anomaly(hsi_img, n_comp, mask = None): """ Gaussian Mixture Model Anomaly Detector fits GMM assuming entire image is background computes negative log likelihood of each pixel in the fit model Inputs: hsi_image - n_row x n_col x n_band hyperspectral image mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used n_comp - number of Gaussian components to use Outputs: gmm_out - detector output image 8/7/2012 - Taylor C. Glenn 5/5/2018 - Edited by Alina Zare 11/2018 - Python Implementation by Yutai Zhou """ gmm_out, kwargsout = img_det(gmm_helper, hsi_img, None, mask, n_comp = n_comp) return gmm_out
def cbad_anomaly(hsi_img, n_cluster, mask = None): """ Cluster Based Anomaly Detection (CBAD) Ref: Carlotto, Mark J. "A cluster-based approach for detecting man-made objects and changes in imagery." IEEE Transactions on Geoscience and Remote Sensing 43.2 (2005): 374-387. Inputs: hsi_image - n_row x n_col x n_band hyperspectral image mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used n_cluster - number of clusters to use Outputs: cbad_out - detector output image cluster_img - cluster label image 8/7/2012 - Taylor C. Glenn 5/5/2018 - Edited by Alina Zare 11/2018 - Python Implementation by Yutai Zhou """ cbad_out, kwargsout = img_det(cbad_out_helper, hsi_img, None, mask, n_cluster = n_cluster) cluster_img = kwargsout['idx'] return cbad_out, cluster_img
def smf_detector(hsi_img, tgt_sig, mask = None, mu = None, sig_inv = None): """ Spectral Matched Filter Inputs: hsi_image - n_row x n_col x n_band hyperspectral image tgt_sig - target signature (n_band x 1 - column vector) mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used mu - (optional) mean for filter (if not provided, computed from image) siginv - (optional) inverse covariance for filter (if not provided, computed from image) Outputs: smf_out - detector image mu - mean of background sig_inv - inverse covariance of background 8/8/2012 - Taylor C. Glenn 6/2/2018 - Edited by Alina Zare 10/2018 - Python Implementation by Yutai Zhou """ smf_out, kwargsout = img_det(smf_det_array_helper, hsi_img, tgt_sig, mask, mu = mu, sig_inv = sig_inv) return smf_out, kwargsout['mu'], kwargsout['sig_inv']
def sam_detector(hsi_img, tgt_sig, mask=None):
"""
Spectral Angle Mapper
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
Outputs:
sam_out - detector image
8/8/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
sam_out, kwargsout = img_det(sam_helper, hsi_img, tgt_sig, mask)
return sam_out
def amsd_detector(hsi_img, tgt_sig, mask=None, n_dim_tgt=1, n_dim_bg=5):
"""
Adaptive Matched Subspace Detector
Reference:
Hyperspectral subpixel target detection using the linear mixing model (article)
Manolakis, D. and Siracusa, C. and Shaw, G.
Geoscience and Remote Sensing, IEEE Transactions on
2001 Volume 39 Number 7 Pages 1392 -1409 Month jul
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sigs - target signature(s) (n_band x n_sig - column vectors)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
n_dim_tgt - number of dimensions to use for target subspace,
if argument is 0, use the target sigs themselves
n_dim_bg - number of dimensions to use for background subspace
Outputs:
amsd_out - detector image
8/22/2012 - Taylor C. Glenn
6/02/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
amsd_out, kwargsout = img_det(amsd_helper,
hsi_img,
tgt_sig,
mask,
n_dim_tgt=n_dim_tgt,
n_dim_bg=n_dim_bg)
return amsd_out
def osp_detector(hsi_img, tgt_sig, mask = None, n_dim_ss = 2): """ Orthogonal Subspace Projection Detector Inputs: hsi_image - n_row x n_col x n_band hyperspectral image tgt_sig - target signature (n_band x 1 - column vector) mask - binary image limiting detector operation to pixels where mask is true if not present or empty, no mask restrictions are used n_dim_ss - number of dimensions to use in the background subspace Outputs: osp_out - detector image 8/8/2012 - Taylor C. Glenn 6/2/2018 - Edited by Alina Zare 11/2018 - Python Implementation by Yutai Zhou """ if tgt_sig.ndim == 1: tgt_sig = tgt_sig[:, np.newaxis] osp_out, kwargsout = img_det(osp_helper, hsi_img, tgt_sig, mask, n_dim_ss = n_dim_ss) return osp_out
def abd_detector(hsi_img, tgt_sig, ems, mask=None):
"""
Abundance of Target when unmixed with background endmembers
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
ems - background endmembers
Outputs:
abd_out - detector image
8/19/2012 - Taylor C. Glenn
6/2/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
abd_out, kwargsout = img_det(abd_helper, hsi_img, tgt_sig, mask, ems=ems)
return abd_out
def hua_detector(hsi_img, tgt_sig, ems, mask=None, n_comp=2, sig_inv=None):
"""
Hybrid Unstructured Abundance Detector
Ref:
Hybrid Detectors for Subpixel Targets
Broadwater, J. and Chellappa, R.
Pattern Analysis and Machine Intelligence, IEEE Transactions on
2007 Volume 29 Number 11 Pages 1891 -1903 Month nov.
Inputs:
hsi_image - n_row x n_col x n_band hyperspectral image
tgt_sig - target signature (n_band x 1 - column vector)
mask - binary image limiting detector operation to pixels where mask is true
if not present or empty, no mask restrictions are used
ems - background endmembers
siginv - background inverse covariance (n_band x n_band matrix)
Outputs:
hua_out - detector image
8/19/2012 - Taylor C. Glenn
6/3/2018 - Edited by Alina Zare
12/2018 - Python Implementation by Yutai Zhou
"""
if tgt_sig.ndim == 1:
tgt_sig = tgt_sig[:, np.newaxis]
hua_out, kwargsout = img_det(hua_helper,
hsi_img,
tgt_sig,
mask,
ems=ems,
n_comp=n_comp,
sig_inv=sig_inv)
return hua_out
