-
Notifications
You must be signed in to change notification settings - Fork 3
/
evaluate.py
466 lines (379 loc) · 20.3 KB
/
evaluate.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
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
import os
import pickle
from pathlib import Path
import cv2
import numpy as np
import rasterio
import sklearn.metrics as sk_metrics
import torch
from albumentations import Compose, Normalize
from albumentations.pytorch import ToTensorV2
from skimage import img_as_ubyte
from skimage.exposure import rescale_intensity
from tifffile import tifffile
from torch.utils.data import DataLoader
from tqdm import tqdm
from data_loader import PatchDataset
from metrics import compute_confusion_matrix, accuracy, precision_recall_fscore_support, iou_score
from prepare_landsat8_biome import get_landsat8_images, read_image, get_ground_truth
split = 'test'
# -------------------- Fixed-Point GAN ----------------------
def test_landsat8_biome(solver, config, split='test'):
l8_images = get_l8_images(config, split=split)
solver.restore_model(config.test_iters, only_g=True)
# When creating generated masks for the training dataset for FCD+, we find the threshold on the training dataset
# to avoid leaking information from the validation set into the generated labels.
# Without pixel-wise ground truth, this could be done similarly by manual inspection.
# best_threshold = solver.find_best_threshold('train' if split == 'train' else 'val')
best_threshold = 0.02
print('Evaluating model on split={} using threshold={:.4}'.format(split, best_threshold))
overall_confusion_matrix = np.zeros((2, 2))
all_metrics = []
biome_wise_cm = {x.biome: np.zeros((2, 2)) for x in l8_images}
for l8_image in tqdm(l8_images, 'Testing'):
# Read scene to memory
full_image, target, profile = read_image(l8_image, return_profile=True)
valid_pixels_mask = target > 0
target[valid_pixels_mask] -= 1
# Compute difference and prediction using trained model
difference = predict_scene(full_image, config, solver)
difference[np.logical_not(valid_pixels_mask)] = 0
mask = solver.binarize(difference, threshold=best_threshold)
# Write difference and prediction to file
write_tifs(difference, mask, valid_pixels_mask, profile, config, l8_image)
write_thumbnails(full_image, mask, target, config, l8_image, thumb_size=full_image.shape[:2]) # (1000, 1000))
# Ignore invalid pixels in metrics
mask = mask[valid_pixels_mask].flatten()
target = target[valid_pixels_mask].flatten()
# Compute metrics
cm = compute_confusion_matrix(mask, target, num_classes=2)
metrics = get_metrics_dict(cm, l8_image.name, l8_image.biome)
print('Result for {} ({}, {:.2%} cloud): Accuracy={:.2%}, F1={:.4} by predicting {:.2%} clouds'
.format(l8_image.name, l8_image.biome, (target == 1).mean(), metrics['accuracy'],
metrics['macro avg']['f1-score'], (mask == 1).mean()))
all_metrics.append(metrics)
overall_confusion_matrix += cm
biome_wise_cm[l8_image.biome] += cm
# Store metrics to file
metrics = get_metrics_dict(overall_confusion_matrix, name='overall', biome='all')
print('Overall Result: Accuracy={:.2%}, F1={:.4}'.format(metrics['accuracy'], metrics['macro avg']['f1-score']))
all_metrics.append(metrics)
pickle.dump(all_metrics, open(os.path.join(config.result_dir, 'biome_metrics.pkl'), 'wb'))
biome_wise_metrics = []
for biome, cm in biome_wise_cm.items():
biome_wise_metrics.append(get_metrics_dict(cm, name='all', biome=biome))
pickle.dump(biome_wise_metrics, open(os.path.join(config.result_dir, 'biome_metrics_biomewise.pkl'), 'wb'))
def predict_scene(full_image, config, solver):
patch_size = config.image_size
batch_size = config.batch_size
crop_size = 0
raster_height, raster_width, _ = full_image.shape
transforms = Compose([
Normalize(mean=(0.5,) * config.num_channels, std=(0.5,) * config.num_channels, max_pixel_value=2 ** 16 - 1),
ToTensorV2()
])
# Cut full image into patches
dataset = PatchDataset(full_image, patch_size, crop_size, transforms)
data_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=False,
num_workers=config.num_workers)
# Predict each patch individually and stitch together to a full image.
difference = np.zeros((raster_height, raster_width))
with torch.no_grad():
for sample in data_loader:
if (sample['image'] == -1).all(): # Don't waste time on invalid patches
continue
x = sample['image'].to(config.device)
c_trg = torch.zeros(x.shape[0], 1).to(config.device) # translate to clear
x_fake = solver.G(x, c_trg)
delta = torch.abs(x_fake - x) / 2 # Divide by 2 so delta goes from [0, 2] to [0, 1]
delta = np.mean(delta.data.cpu().numpy(), axis=1)
# Optionally crop generated image to avoid border artifacts
if crop_size > 0:
delta = delta[:, crop_size:-crop_size, crop_size:-crop_size]
rows, cols = sample['row'], sample['col']
for batch_idx in range(len(x)):
row, col = rows[batch_idx], cols[batch_idx]
height = min(raster_height - row, dataset.cropped_patch_size)
width = min(raster_width - col, dataset.cropped_patch_size)
difference[row:row + height, col:col + width] += delta[batch_idx, :height, :width]
return difference
# -------------------- Supervised ----------------------
def test_landsat8_biome_supervised(solver, config):
l8_images = get_l8_images(config, split='test')
solver.restore_model(Path(config.model_save_dir) / 'best.pt')
overall_confusion_matrix = np.zeros((2, 2))
all_metrics = []
biome_wise_cm = {x.biome: np.zeros((2, 2)) for x in l8_images}
for l8_image in tqdm(l8_images, 'Testing Supervised'):
# Read scene to memory
full_image, target, profile = read_image(l8_image, return_profile=True)
valid_pixels_mask = target > 0
target[valid_pixels_mask] -= 1
# Compute difference and prediction using trained model
mask = predict_scene_supervised(full_image, config, solver)
mask[np.logical_not(valid_pixels_mask)] = 0
write_thumbnails(full_image, mask, target, config, l8_image, thumb_size=full_image.shape[:2])
# Ignore invalid pixels in metrics
mask = mask[valid_pixels_mask].flatten()
target = target[valid_pixels_mask].flatten()
# Compute metrics
cm = compute_confusion_matrix(mask, target, num_classes=2)
biome_wise_cm[l8_image.biome] += cm
metrics = get_metrics_dict(cm, l8_image.name, l8_image.biome)
print('Result for {} ({}, {:.2%} cloud): Accuracy={:.2%}, F1={:.4} by predicting {:.2%} clouds'
.format(l8_image.name, l8_image.biome, (target == 1).mean(), metrics['accuracy'],
metrics['macro avg']['f1-score'], (mask == 1).mean()))
all_metrics.append(metrics)
overall_confusion_matrix += cm
# Compute overall metrics
metrics = get_metrics_dict(overall_confusion_matrix, name='overall', biome='all')
print('Overall Result: Accuracy={:.2%}, F1={:.4}'.format(metrics['accuracy'], metrics['macro avg']['f1-score']))
all_metrics.append(metrics)
pickle.dump(all_metrics, open(os.path.join(config.result_dir, 'biome_metrics.pkl'), 'wb'))
biome_wise_metrics = []
for biome, cm in biome_wise_cm.items():
biome_wise_metrics.append(get_metrics_dict(cm, name='all', biome=biome))
pickle.dump(biome_wise_metrics, open(os.path.join(config.result_dir, 'biome_metrics_biomewise.pkl'), 'wb'))
def predict_scene_supervised(full_image, config, solver):
patch_size = config.image_size
batch_size = config.batch_size
crop_size = 0
raster_height, raster_width, _ = full_image.shape
transforms = Compose([
Normalize(mean=(0.5,) * config.num_channels, std=(0.5,) * config.num_channels, max_pixel_value=2 ** 16 - 1),
ToTensorV2()
])
# Cut full image into patches
dataset = PatchDataset(full_image, patch_size, crop_size, transforms)
data_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=False,
num_workers=config.num_workers)
# Predict each patch individually and stitch together to a full image.
predictions = np.zeros((raster_height, raster_width))
solver.model.eval()
with torch.no_grad():
for sample in data_loader:
if (sample['image'] == -1).all(): # Don't waste time on invalid patches
continue
x = sample['image'].to(config.device)
output, _ = solver.model(x)
prediction = output.argmax(1).cpu().numpy()
# Optionally crop generated image to avoid border artifacts
if crop_size > 0:
prediction = prediction[:, crop_size:-crop_size, crop_size:-crop_size]
rows, cols = sample['row'], sample['col']
for batch_idx in range(len(x)):
row, col = rows[batch_idx], cols[batch_idx]
height = min(raster_height - row, dataset.cropped_patch_size)
width = min(raster_width - col, dataset.cropped_patch_size)
predictions[row:row + height, col:col + width] += prediction[batch_idx, :height, :width]
return predictions
# -------------------- U-CAM ----------------------
def test_landsat8_biome_cam(solver, config):
l8_images = get_l8_images(config, split='test')
solver.restore_model(Path(config.model_save_dir) / 'best.pt')
best_val_threshold = solver.find_best_threshold()
print('Best threshold', best_val_threshold)
overall_confusion_matrix = np.zeros((2, 2))
all_metrics = []
biome_wise_cm = {x.biome: np.zeros((2, 2)) for x in l8_images}
for l8_image in tqdm(l8_images, 'Testing CAM'):
# Read scene to memory
full_image, target, profile = read_image(l8_image, return_profile=True)
valid_pixels_mask = target > 0
target[valid_pixels_mask] -= 1
# Compute difference and prediction using trained model
prediction = predict_scene_cam(full_image, config, solver)
mask = (prediction > best_val_threshold).astype(np.uint8)
mask[np.logical_not(valid_pixels_mask)] = 0
write_thumbnails(full_image, mask, target, config, l8_image, thumb_size=(1000, 1000))
# Ignore invalid pixels in metrics
mask = mask[valid_pixels_mask].flatten()
target = target[valid_pixels_mask].flatten()
# Compute metrics
cm = compute_confusion_matrix(mask, target, num_classes=2)
biome_wise_cm[l8_image.biome] += cm
metrics = get_metrics_dict(cm, l8_image.name, l8_image.biome)
print('Result for {} ({}, {:.2%} cloud): Accuracy={:.2%}, F1={:.4} by predicting {:.2%} clouds'
.format(l8_image.name, l8_image.biome, (target == 1).mean(), metrics['accuracy'],
metrics['macro avg']['f1-score'], (mask == 1).mean()))
all_metrics.append(metrics)
overall_confusion_matrix += cm
# Store metrics to file
metrics = get_metrics_dict(overall_confusion_matrix, name='overall', biome='all')
print('Overall Result: Accuracy={:.2%}, F1={:.4}'.format(metrics['accuracy'], metrics['macro avg']['f1-score']))
all_metrics.append(metrics)
pickle.dump(all_metrics, open(os.path.join(config.result_dir, 'biome_metrics.pkl'), 'wb'))
biome_wise_metrics = []
for biome, cm in biome_wise_cm.items():
biome_wise_metrics.append(get_metrics_dict(cm, name='all', biome=biome))
pickle.dump(biome_wise_metrics, open(os.path.join(config.result_dir, 'biome_metrics_biomewise.pkl'), 'wb'))
def predict_scene_cam(full_image, config, solver, crop_size=0):
patch_size = config.image_size
batch_size = config.batch_size
raster_height, raster_width, _ = full_image.shape
transforms = Compose([
Normalize(mean=(0.5,) * config.num_channels, std=(0.5,) * config.num_channels, max_pixel_value=2 ** 16 - 1),
ToTensorV2()
])
# Cut full image into patches
dataset = PatchDataset(full_image, patch_size, crop_size, transforms)
data_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=False,
num_workers=config.num_workers)
# Predict each patch individually and stitch together to a full image.
predictions = np.zeros((raster_height, raster_width))
solver.model.eval()
with torch.no_grad():
for sample in data_loader:
if (sample['image'] == -1).all(): # Don't waste time on invalid patches
continue
x = sample['image'].to(config.device)
cam, _ = solver.model.cam(x)
prediction = cam.cpu().numpy().squeeze()
# Optionally crop generated image to avoid border artifacts
if crop_size > 0:
prediction = prediction[:, crop_size:-crop_size, crop_size:-crop_size]
rows, cols = sample['row'], sample['col']
for batch_idx in range(len(x)):
row, col = rows[batch_idx], cols[batch_idx]
height = min(raster_height - row, dataset.cropped_patch_size)
width = min(raster_width - col, dataset.cropped_patch_size)
predictions[row:row + height, col:col + width] += prediction[batch_idx, :height, :width]
return predictions
# -------------------- CFMask ----------------------
def test_landsat8_biome_fmask(config):
l8_images = get_l8_images(config, split='test')
overall_confusion_matrix = np.zeros((2, 2))
all_metrics = []
biome_wise_cm = {x.biome: np.zeros((2, 2)) for x in l8_images}
for l8_image in l8_images:
orig_qa_mask = tifffile.imread(str(l8_image.qa_cloud_mask))
# Determined from https://landsat.usgs.gov/sites/default/files/documents/landsat_QA_tools_userguide.pdf
# for pre-collection QA band values
qa_mask = np.ones_like(orig_qa_mask) # Start from all background
qa_mask[orig_qa_mask == 1] = 0 # invalid pixels
qa_mask[orig_qa_mask & 0b1000000000000000 == 0b1000000000000000] = 2 # medium confidence cloud
qa_mask[orig_qa_mask & 0b1100000000000000 == 0b1100000000000000] = 2 # high confidence cloud
qa_mask[orig_qa_mask & 0b0010000000000000 == 0b0010000000000000] = 2 # medium confidence cirrus
qa_mask[orig_qa_mask & 0b0011000000000000 == 0b0011000000000000] = 2 # high confidence cirrus
target = get_ground_truth(l8_image)
# Ignore invalid pixels
valid_pixels_mask = target > 0
target = target[valid_pixels_mask] - 1
qa_mask = qa_mask[valid_pixels_mask] - 1
# Compute metrics
cm = compute_confusion_matrix(qa_mask, target, num_classes=2)
biome_wise_cm[l8_image.biome] += cm
metrics = get_metrics_dict(cm, l8_image.name, l8_image.biome)
print('Result for {} ({}, {:.2%} cloud): Accuracy={:.2%}, F1={:.4} by predicting {:.2%} clouds'
.format(l8_image.name, l8_image.biome, (target == 1).mean(), metrics['accuracy'],
metrics['macro avg']['f1-score'], (qa_mask == 1).mean()))
all_metrics.append(metrics)
overall_confusion_matrix += cm
# Store metrics to file
metrics = get_metrics_dict(overall_confusion_matrix, name='overall', biome='all')
print('Overall Result: Accuracy={:.2%}, F1={:.4}'.format(metrics['accuracy'], metrics['macro avg']['f1-score']))
all_metrics.append(metrics)
fmask_result_dir = os.path.join('outputs', 'fmask')
os.makedirs(fmask_result_dir, exist_ok=True)
pickle.dump(all_metrics, open(os.path.join(fmask_result_dir, 'biome_metrics.pkl'), 'wb'))
biome_wise_metrics = []
for biome, cm in biome_wise_cm.items():
biome_wise_metrics.append(get_metrics_dict(cm, name='all', biome=biome))
pickle.dump(biome_wise_metrics, open(os.path.join(fmask_result_dir, 'biome_metrics_biomewise.pkl'), 'wb'))
# ----- Helper functions -------
# Taken from https://github.com/developmentseed/landsat-util/blob/develop/landsat/image.py
def to_uint8_based_on_cloud_coverage(bands, cloud_coverage=100.0):
output = np.zeros_like(bands, dtype=np.uint8)
for i, band in enumerate(bands):
# Color Correction
band = _color_correction(band, 0, cloud_coverage)
band = img_as_ubyte(band)
output[i] = band
return output
def _color_correction(band, low, coverage):
p_low, cloud_cut_low = _percent_cut(band, low, 100 - (coverage * 3 / 4))
temp = np.zeros(np.shape(band), dtype=np.uint16)
cloud_divide = 65000 - coverage * 100
mask = np.logical_and(band < cloud_cut_low, band > 0)
temp[mask] = rescale_intensity(band[mask], in_range=(p_low, cloud_cut_low), out_range=(256, cloud_divide))
temp[band >= cloud_cut_low] = rescale_intensity(band[band >= cloud_cut_low],
out_range=(cloud_divide, 65535))
return temp
def _percent_cut(color, low, high):
return np.percentile(color[np.logical_and(color > 0, color < 65535)], (low, high))
def write_tifs(difference, prediction, valid_mask, profile, config, l8_image):
profile.update(driver='GTiff', count=1, compress='lzw', dtype='uint8')
output_dir = os.path.join(config.result_dir, 'tifs')
os.makedirs(output_dir, exist_ok=True)
prefix = os.path.join(output_dir, f'{l8_image.biome}_{l8_image.name}')
with rasterio.open('{}_difference.tif'.format(prefix), 'w', **profile) as dst:
dst.write((difference[np.newaxis, ...] * 255).astype(np.uint8))
mask = prediction.copy()
mask[mask == 0] = 128
mask[mask == 1] = 255
mask[~valid_mask] = 0
with rasterio.open('{}_mask.tif'.format(prefix), 'w', **profile) as dst:
dst.write((mask[np.newaxis, ...]).astype(np.uint8))
def get_l8_images(config, split='test'):
l8_images = get_landsat8_images(Path(config.orig_image_dir))
with open(os.path.join(config.l8biome_image_dir, 'assignment.txt')) as f:
lines = [x.split(',') for x in f.read().splitlines()]
lines = [(x[-1]) for x in lines if x[0] == split]
return [x for x in l8_images if x.name in lines]
def write_thumbnails(image, prediction, target, config, l8_image, thumb_size=(1000, 1000)):
output_dir = os.path.join(config.result_dir, 'l8_thumbnails')
file_prefix = f'{l8_image.biome}_{l8_image.name}'
image = image.copy()[..., [3, 2, 1]]
image = np.moveaxis(image, -1, 0) # Bands first
image = to_uint8_based_on_cloud_coverage(image, cloud_coverage=(target == 1).mean() * 100.0)
image = np.moveaxis(image, 0, -1) # Bands last
thumb = cv2.resize(image, thumb_size)
# color = (63, 255, 63)
color = (26, 178, 255)
# color = (255, 0, 255)
mask = cv2.resize(prediction, thumb_size, interpolation=cv2.INTER_NEAREST)
mask = (mask[..., np.newaxis] * np.array(color)).astype(np.uint8)
weighted_sum = cv2.addWeighted(mask, 0.5, thumb, 0.5, 0.)
ind = np.any(mask > 0, axis=-1)
thumb_pred = thumb.copy()
thumb_pred[ind] = weighted_sum[ind]
mask = cv2.resize(target, thumb_size, interpolation=cv2.INTER_NEAREST)
mask = (mask[..., np.newaxis] * np.array(color)).astype(np.uint8)
weighted_sum = cv2.addWeighted(mask, 0.5, thumb, 0.5, 0.)
ind = np.any(mask > 0, axis=-1)
thumb_target = thumb.copy()
thumb_target[ind] = weighted_sum[ind]
os.makedirs(output_dir, exist_ok=True)
cv2.imwrite(os.path.join(output_dir, file_prefix + '_image.jpg'), thumb[..., [2, 1, 0]]) # cv2 uses BGR format
cv2.imwrite(os.path.join(output_dir, file_prefix + '_prediction.jpg'), thumb_pred[..., [2, 1, 0]])
cv2.imwrite(os.path.join(output_dir, file_prefix + '_target.jpg'), thumb_target[..., [2, 1, 0]])
def get_metrics_dict(confusion_matrix, name=None, biome=None):
acc = accuracy(confusion_matrix)
p, r, f1, s = precision_recall_fscore_support(confusion_matrix)
iou = iou_score(confusion_matrix, reduce_mean=False)
return {
'name': name,
'biome': biome,
'clear': {
'precision': p[0],
'recall': r[0],
'f1-score': f1[0],
'iou-score': iou[0],
'support': s[0],
},
'cloudy': {
'precision': p[1],
'recall': r[1],
'f1-score': f1[1],
'iou-score': iou[0],
'support': s[1],
},
'accuracy': acc,
'macro avg': {
'precision': np.mean(p),
'recall': np.mean(r),
'f1-score': np.mean(f1),
'iou-score': np.mean(iou),
'support': np.sum(s),
}
}