-
Notifications
You must be signed in to change notification settings - Fork 4
/
tikorder.py
397 lines (336 loc) · 16.3 KB
/
tikorder.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
# David R. Thompson
# Copyright 2019, by the California Institute of Technology. ALL RIGHTS
# RESERVED. United States Government Sponsorship acknowledged. Any commercial
# use must be negotiated with the Office of Technology Transfer at the
# California Institute of Technology. This software is controlled under the
# U.S. Export Regulations and may not be released to foreign persons without
# export authorization, e.g., a license, license exception or exemption.
from spectral.io import envi
import argparse, json
import scipy
import scipy.interpolate
from collections import OrderedDict
import numpy as np
import pylab as plt
import logging
import pylab as plt
import multiprocessing as mp
import ray
# Suppress warnings that don't come from us
import warnings
warnings.filterwarnings("ignore")
def srf(x, mu, sigma):
"""Spectral Response Function """
u = (x-mu)/abs(sigma)
y = (1.0/(np.sqrt(2.0*np.pi)*abs(sigma)))*np.exp(-u*u/2.0)
return y/y.sum()
def spectrumResample(x, wl, wl2, fwhm2=10, fill=False):
"""Resample a spectrum to a new wavelength / FWHM.
I assume Gaussian SRFs"""
return np.array([x[np.newaxis,:] @ srf(wl, wi, fwhmi/2.355)[:,np.newaxis]
for wi, fwhmi in zip(wl2, fwhm2)]).reshape((len(wl2)))
class ConstraintViolation(Exception):
def __init__(self, message):
self.msg = message
class LibrarySpectrum():
def __init__(self, config):
"""Parse a configuration object, get spectrum and constraints"""
self.wl, self.rfl = np.loadtxt(config['reflectance_file']).T
if all(self.wl<100):
self.wl = self.wl * 1000.0 # convert to nm
self.rfl = self.rfl.reshape((len(self.rfl),))
self.group = config['group']
self.name = config['name']
self.prior = np.array(config['prior'])
self.mixing = np.array(config['mixing_fraction'])
self.continuua = []
self.cr = [] # container for the actual library cr
for feature in config['features']:
self.continuua.append(feature['continuum'])
self.n_intervals = len(self.continuua)
for constraint in ['constraint_ctm_slope', 'constraint_ctm_left',
'constraint_ctm_right','constraint_ctm']:
setattr(self, constraint, [])
for feature in config['features']:
if constraint in feature:
getattr(self,constraint).append(feature[constraint])
else:
getattr(self,constraint).append([-9999.0, 9999.0])
def resample(self, wl, fwhm):
"""Resample the spectrum to a new wavelength grid"""
rfl = self.rfl.copy().reshape(self.wl.shape)
self.rfl = spectrumResample(rfl, self.wl, wl, fwhm)
self.wl = wl.copy()
def ctmrm(self, rfl, plot=False):
"""Remove the continuua from a reflectance spectrum, raising
exceptions if any constraint is violated.
Input:
rfl - Reflectance spectrum. Should have the same
wavelength sampling and size as this library
spectrum.
Returns:
refl_ctmrm - Continuum-removed reflectance across N features.
The format is a C x N array, where C is the total
number of wavelengths across all feature intervals
(concatenated one after the other) and N is the
number of feature intervals. Each interval's
continuum-removed reflectance appears in the
relevant column, and the reflectances of all other
intervals in that column (representing other
features) are set to zero.
ctm - The continuum estimate in the original reflectance
units. The format is a C x N array with te same
block-nonzero structure described above.
ivals - A binary matrix of size C x N (see above) with
ones in the "active" channels for each feature's
column, zeros elsewhere.
idx - An integer matrix representing the indices of
all feature intervals, concatentated together, in
the original reflectance spectrum. In other words,
the mapping from channels in rfl_ctmrm to channels
in the original input, rfl.
"""
rfl_ctmrm = [[] for q in range(self.n_intervals)]
ctms = [[] for q in range(self.n_intervals)]
ivals = [[] for q in range(self.n_intervals)]
idx_ctmrm = []
for i in range(self.n_intervals):
# Locate continuum
rctma, rctmb, lctma, lctmb = self.continuua[i]
in_rct = np.logical_and(self.wl>rctma, self.wl<rctmb)
in_lct = np.logical_and(self.wl>lctma, self.wl<lctmb)
rctm_idx = int(np.where(in_rct)[0].mean())
lctm_idx = int(np.where(in_lct)[0].mean())
idx_ctmrm.extend(range(rctm_idx,lctm_idx+1))
rct = rfl[in_rct].mean()
lct = rfl[in_lct].mean()
ct = (rct+lct)/2.0
slope = rct/lct
# Check constraints
if not (rct > self.constraint_ctm_right[i][0] and \
rct < self.constraint_ctm_right[i][1]):
raise ConstraintViolation(f'{self.name}, rct = {rct}')
if not (lct > self.constraint_ctm_left[i][0] and \
lct < self.constraint_ctm_left[i][1]):
raise ConstraintViolation(f'{self.name}, lct = {lct}')
if not (slope > self.constraint_ctm_slope[i][0] and \
slope < self.constraint_ctm_slope[i][1]):
raise ConstraintViolation(f'{self.name}, rct/lct = {slope}')
if not (ct > self.constraint_ctm[i][0] and \
ct < self.constraint_ctm[i][1]):
raise ConstraintViolation(f'{self.name}, ct = {ct}')
# divide by local continuum across this interval
rfl_ival = rfl[rctm_idx:(lctm_idx+1)]
n_channels = len(rfl_ival)
p = scipy.interpolate.interp1d([0, n_channels-1], [rfl_ival[0], rfl_ival[-1]])
ctm = p(range(n_channels))
if plot:
plt.plot(self.wl, rfl)
plt.plot(self.wl[rctm_idx:(lctm_idx+1)], p(range(n_channels)))
plt.show(block=True)
for j in range(self.n_intervals):
if i==j:
rfl_ctmrm[j].extend(rfl_ival/ctm)
ctms[j].extend(ctm)
ivals[j].extend(np.ones(n_channels))
else:
rfl_ctmrm[j].extend(np.zeros(n_channels))
ctms[j].extend(np.zeros(n_channels))
ivals[j].extend(np.zeros(n_channels))
return (np.array(rfl_ctmrm).T, np.array(ctms).T, np.array(ivals).T,
idx_ctmrm)
def fit (self, rfl, uncert, plot=False):
# Set up matrices
obs_noise = pow(uncert.copy(), 2)
if uncert.ndim < 2 or (uncert.shape[0] != uncert.shape[1]):
obs_noise = np.diag(obs_noise)
rfl_ref, ctm_ref, ivals, idx = self.cr
rfl_test, ctm_test, ivals, idx = self.ctmrm(rfl)
rfl_test = rfl_test.sum(axis=1)
ctm_test = ctm_test.sum(axis=1)
# State vector has multipliers, offsets in that order (one per feature)
K = np.concatenate((rfl_ref, ivals), axis=1)
offs_prior = np.eye(ivals.shape[1])
S_a = scipy.linalg.block_diag(self.prior, offs_prior)
x_a = np.zeros(S_a.shape[0])
# Input uncertainty treats continuum removal, a linear transformation
# It currently ignores uncertainty in the continuum placement itself!
S_e = np.array([obs_noise[i, idx] for i in idx])
Q = np.eye(len(ctm_test)) * (1.0/ctm_test)
S_e = Q.T @ S_e @ Q
iS_e, iS_a = scipy.linalg.inv(S_e), scipy.linalg.inv(S_a)
# Tikonov solution provides true posteriors
x = x_a + scipy.linalg.inv(K.T @ iS_e @ K + iS_a) @ (K.T @ iS_e @ (rfl_test - K @ x_a))
S_hat = scipy.linalg.inv(K.T @ iS_e @ K + iS_a)
rfl_hat = K @ x
residual = rfl_hat - rfl_test
sign, logdet = np.linalg.slogdet(S_e)
Z = len(residual) * np.log(2*np.pi) + logdet
nll = 0.5 * (residual @ iS_e @ residual + Z)
corr = pow(np.corrcoef(rfl_hat, rfl_test)[0, 1], 2)
# Depth estimate is taken from the most certain measurement
# In the future consider a Kalman-like update
coeffs = x[:self.n_intervals]
uncerts = np.sqrt(np.diag(S_hat))[:self.n_intervals]
best = np.argmin(uncerts)
depth = coeffs[best]
post = uncerts[best]
return depth, post, nll, corr
class Library():
def __init__(self, config):
self.lib = OrderedDict()
for fn in config:
for source in config['sources']:
src = LibrarySpectrum(source)
if src.group in self.lib:
self.lib[src.group].append(src)
else:
self.lib[src.group] = [src]
c, self.wl, self.fwhm = np.loadtxt(config['wavelength_file']).T
if all(self.wl < 100):
self.wl = self.wl * 1000 # convert to nm
self.fwhm = self.fwhm * 1000
self.resample(self.wl, self.fwhm)
self.cr() # calculate cr for lib instances
def resample(self, wl, fwhm):
for grp in self.lib:
for i in range(len(self.lib[grp])):
self.lib[grp][i].resample(wl, fwhm)
self.wl = wl
self.fwhm = fwhm
def cr(self):
for group in self.lib:
for i in range(len(self.lib[group])):
rfl = self.lib[group][i].rfl
self.lib[group][i].cr = self.lib[group][i].ctmrm(rfl)
def fit(self, data, plot=False):
rfl = data[:, 0]
uncert = data[:, 1]
group_depths, group_posts, group_nlls, group_corrs = [], [], [], []
for group, spectra in self.lib.items():
depths, posts, nlls, models, corrs = [],[],[],[],[] # negative log likelihood of fitted model
for spectrum in spectra:
try:
depth, post, nll, corr = spectrum.fit(rfl, uncert, plot)
corrs.append(corr)
nlls.append(nll)
depths.append(depth * spectrum.mixing)
posts.append(post * spectrum.mixing)
models.append(spectrum.name)
except ConstraintViolation as v:
continue
if len(nlls) == 0:
group_depths.append(0)
group_posts.append(-9999)
group_nlls.append(999999)
group_corrs.append(0)
continue
best_fit_idx = np.argmin(nlls)
group_depths.append(depths[best_fit_idx])
group_posts.append(posts[best_fit_idx])
group_nlls.append(nlls[best_fit_idx])
group_corrs.append(corrs[best_fit_idx])
return (np.array(group_depths), np.array(group_posts),
np.array(group_nlls), np.array(group_corrs))
def nchan(self):
return len(self.lib)
def groups(self):
return self.lib.keys()
@ray.remote
def run_one_row(r, lib, reflectance_hdr, uncertainty_hdr, depth_hdr, posterior_hdr, likelihood_hdr, corr_hdr):
reflectance_ds = envi.open(reflectance_hdr)
uncertainty_ds = envi.open(uncertainty_hdr)
depth_ds = envi.open(depth_hdr)
posterior_ds = envi.open(posterior_hdr)
likelihood_ds = envi.open(likelihood_hdr)
corr_ds = envi.open(corr_hdr)
#logging.info('Row %i'%r)
print(f'Row {r}')
# We delete the old objects to flush everything to disk, empty cache
reflectance_mm = reflectance_ds.open_memmap(interleave="bip", writable=False)
uncertainty_mm = uncertainty_ds.open_memmap(interleave="bip", writable=False)
depth_mm = depth_ds.open_memmap(interleave="source", writable=True)
posterior_mm = posterior_ds.open_memmap(interleave="source", writable=True)
likelihood_mm = likelihood_ds.open_memmap(interleave="source", writable=True)
corr_mm = corr_ds.open_memmap(interleave="source", writable=True)
# Get reflectance subframe
sub_rfl = np.array(reflectance_mm[r,:,:], dtype='float32')
# Get input uncertainty
sub_uncert = np.array(uncertainty_mm[r,:,:], dtype='float32')
# Set-up for parallel. By convention, we exclude final state
# vector uncertainties which are related typically to atmosphere
nrfl = sub_rfl.shape[1]
sub_data = np.stack((sub_rfl, sub_uncert[:, 0:nrfl]), axis=2)
results = []
for c in range(sub_data.shape[0]):
temp = lib.fit(sub_data[c,:,:])
results.append(temp)
results = np.asarray(results)
# Write to output file
depth_mm[r,:,:] = results[:, 0, :].copy()
posterior_mm[r,:,:] = results[:, 1, :].copy()
likelihood_mm[r,:,:] = results[:, 2, :].copy()
corr_mm[r,:,:] = results[:, 3, :].copy()
del reflectance_mm, uncertainty_mm, depth_mm, posterior_mm, likelihood_mm, corr_mm
def main():
# Parse command line
description = 'Spectroscopic Surface & Atmosphere Fitting'
parser = argparse.ArgumentParser()
parser.add_argument('config_file')
parser.add_argument('--level', default='INFO')
parser.add_argument('--ip_head',default=None, help='ray-specific argument')
parser.add_argument('--redis_password',default=None, help='ray-specific argument')
parser.add_argument('--ray_temp_dir',default=None, help='ray-specific argument')
parser.add_argument('--n_cores', type=int, default=-1,help="number of cores to run on. -1 for all, 1 for debug mode")
args = parser.parse_args()
logging.basicConfig(format='%(message)s', level=args.level)
# Load a parallel Pool
rayargs = {'address': args.ip_head, 'redis_password': args.redis_password,
'local_mode': args.n_cores == 1}
if args.n_cores != -1:
rayargs['num_cpus'] = args.n_cores
if args.ray_temp_dir is not None:
rayargs['temp_dir'] = args.ray_temp_dir
ray.init(**rayargs)
# Load the configuration file.
config = json.load(open(args.config_file, 'r'))
logging.info('Loading library')
lib = Library(config['library'])
# Get image and wavelengths
logging.info('Opening input data')
reflectance_input_header = str(config['input_reflectance']+'.hdr')
uncertainty_input_header = str(config['input_uncertainty']+'.hdr')
depth_output_header = str(config['output_depths']+'.hdr')
posterior_output_header = str(config['output_posterior']+'.hdr')
likelihood_output_header = str(config['output_likelihood']+'.hdr')
corr_output_header = str(config['output_corr']+'.hdr')
reflectance_ds = envi.open(reflectance_input_header)
meta = reflectance_ds.metadata.copy()
uncertainty_ds = envi.open(uncertainty_input_header)
# Now that the input images are available, resample wavelengths
if 'wavelength' in meta:
wl = np.array([float (w) for w in meta['wavelength']])
if all(wl<100): wl = wl * 1000.0
else:
wl = lib.wl.copy()
if 'fwhm' in meta:
fwhm = np.array([float (f) for f in meta['fwhm']])
if all(fwhm<0.1): fwhm = fwhm * 1000.0
else:
fwhm = np.ones(wl.shape) * (wl[1]-wl[0])
lib.resample(wl, fwhm)
# Create output images
meta['bands'] = lib.nchan()
if 'wavelength' in meta: del meta['wavelength']
if 'fwhm' in meta: del meta['fwhm']
meta['band names'] = lib.groups()
meta['data type'] = 4
meta['interleave'] = 'bip'
depth_ds = envi.create_image(depth_output_header, meta, force=True, ext="")
posterior_ds = envi.create_image(posterior_output_header, meta, force=True, ext="")
likelihood_ds = envi.create_image(likelihood_output_header, meta, force=True, ext="")
corr_ds = envi.create_image(corr_output_header, meta, force=True, ext="")
ids = [run_one_row.remote(r, lib, reflectance_input_header, uncertainty_input_header, depth_output_header, posterior_output_header, likelihood_output_header, corr_output_header) for r in range(reflectance_ds.shape[0])]
ret = [ray.get(id) for id in ids]
if __name__ == "__main__":
main()