class CDSST:
def __init__(self):
path = 'frames/'
day_images = SkyCameraFile.glob(path)
day_images.sort()
times = np.array(
[SkyCameraFile.parseTime(x).datetime for x in day_images])
self.day_images = day_images
self.times = times
self.calibration = Calibration(catalog=SkyCatalog(True))
self.calibration.load()
self.helper = CloudDetectionHelper()
try:
with open('cd_sst.cache', 'rb') as f:
self.cache = pickle.load(f)
except:
self.cache = pd.DataFrame(index=pd.Index([], dtype=np.datetime64),
columns=[
'pos_x', 'pos_y', 'radius',
'stripe_min', 'stripe_max'
])
def save_cache(self):
with open('cd_sst.cache', 'wb') as f:
pickle.dump(self.cache, f)
def detect(self, filename):
image = cv2.imread(filename)
time = SkyCameraFile.parseTime(filename).datetime
mask = np.uint8(self.helper.get_mask(image).copy())
self.calibration.selectImage(filename)
pos = self.calibration.project()
pos = (pos[0], pos[1])
if time in self.cache.index:
sun_x, sun_y, radius, min_x, max_x = self.cache.loc[time]
sun = None
sun_pos = None
if not np.isnan(sun_x):
sun_pos = (sun_x, sun_y)
sun = CDSunRemoval.circle_mask(sun_pos, radius, mask.shape)
sun_line = None
if not np.isnan(min_x):
sun_line = np.zeros(mask.shape, np.uint8)
sun_line[:, min_x:max_x + 1] = True
else:
sun_line, min_x, max_x = CDSunRemoval.find_sun_line(image, pos)
sun, sun_pos, radius = CDSunRemoval.find_sun(pos, mask)
sun_x = sun_y = None
if sun_pos is not None:
sun_x = sun_pos[0]
sun_y = sun_pos[1]
self.cache.loc[time] = [sun_x, sun_y, radius, min_x, max_x]
if sun_pos is None:
sun_pos = pos
mask = self.helper.fullmask.copy()
mask[self.helper.mask == 0] = 0
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE,
np.ones((11, 11), np.uint8))
_, contours, _ = cv2.findContours(mask, cv2.RETR_LIST,
cv2.CHAIN_APPROX_SIMPLE)
cloudiness = np.ones(mask.shape, np.float32) * .5
did_sun = False
for contour in contours:
area = cv2.contourArea(contour)
is_sun = False
if area > 100: # TODO: try different numbers
single_contour = np.zeros(mask.shape, np.uint8)
cv2.drawContours(single_contour, [contour], 0, 1, cv2.FILLED)
if sun is not None and not did_sun:
sun_area = np.sum(sun[self.helper.mask == 1])
if area > 0.9 * sun_area:
joint_area = np.sum(np.logical_and(
single_contour, sun))
if sun_area / joint_area > 0.9:
is_sun = True
if is_sun:
if sun_area * 1.2 < area:
difference = np.uint8(
np.logical_and(np.logical_not(sun),
single_contour))
_, contours2, _ = cv2.findContours(
difference, cv2.RETR_LIST,
cv2.CHAIN_APPROX_SIMPLE)
# filter smaller ones here! currently done with the if at the beginning
contours += contours2
did_sun = True
if not is_sun:
cloudiness[single_contour > 0] = 1.0
b, g, r = cv2.split(np.int32(image))
mask = self.helper.mask
rmb = r - b
rmb[mask == 0] = 0
cloudiness[rmb < -10] = 0
cloudiness[rmb > 50] = 1
delta = np.timedelta64(39, 's')
delta2 = np.timedelta64(0, 's')
time_diff = time - self.times
before = np.logical_and(time_diff > delta2, time_diff < delta)
if np.sum(before) == 0:
raise ValueError('No previous image found.')
current_index = np.where(before)[0][0]
prev_img = cv2.imread(self.day_images[current_index])
gray_prev = cv2.cvtColor(prev_img, cv2.COLOR_BGR2GRAY)
gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
flow = cv2.calcOpticalFlowFarneback(gray_prev, gray_image, None, 0.5,
3, 15, 3, 5, 1.2, 0)
flow2 = -cv2.calcOpticalFlowFarneback(gray_image, gray_prev, None, 0.5,
3, 15, 3, 5, 1.2, 0)
mag1, _ = cv2.cartToPolar(flow[..., 0], flow[..., 1])
mag2, _ = cv2.cartToPolar(flow2[..., 0], flow2[..., 1])
movement = np.logical_and(mag1 > 1, mag2 > 1)
no_movement = np.logical_not(movement)
brightness = np.mean(image, 2)
cloudiness[np.logical_and(movement == 1, cloudiness != 0)] = 1
cloudiness[self.helper.mask == 0] = 1
if sun is not None:
cloudiness[sun == 1] = 1
if sun_line is not None:
sun_line_dilated = cv2.morphologyEx(sun_line, cv2.MORPH_DILATE,
np.ones((1, 3)))
cloudiness[sun_line_dilated == 1] = 1
y, x = np.mgrid[0:brightness.shape[0], 0:brightness.shape[1]]
x1 = []
y1 = []
out = []
for i in range(brightness.shape[0]):
for j in range(brightness.shape[1]):
if cloudiness[i, j] != 1:
x1.append(x[i, j])
y1.append(y[i, j])
out.append(brightness[i, j])
x = np.array(x1) - sun_pos[0]
y = np.array(y1) - sun_pos[1]
out = np.array(out)
dist = np.sqrt(x * x + y * y)
A = np.array([dist, np.ones(x.shape), x, y]).transpose()
A_inv = np.linalg.pinv(A)
param = np.dot(A_inv, out)
y, x = np.mgrid[0:brightness.shape[0], 0:brightness.shape[1]]
x = x - pos[0]
y = y - pos[1]
dist = np.sqrt(x * x + y * y)
A = np.array([dist, np.ones(x.shape), x, y]).transpose()
gradient = np.dot(A, param).transpose()
rect_size = 15
rect_border = (rect_size - 1) // 2
brightness_norm = brightness - gradient
stddev = np.zeros(brightness.shape)
for y in range(image.shape[0]):
for x in range(image.shape[1]):
if cloudiness[y, x] == 1:
continue
lx = x - rect_border
rx = x + rect_border + 1
uy = y - rect_border
dy = y + rect_border + 1
if lx < 0: lx = 0
if uy < 0: uy = 0
if rx > image.shape[1]: rx = image.shape[1]
if dy > image.shape[0]: dy = image.shape[0]
mask_part = cloudiness[uy:dy, lx:rx]
stddev[y, x] = np.std(brightness_norm[uy:dy,
lx:rx][mask_part != 1])
def_clear = np.sum(cloudiness == 0)
cloudiness[cloudiness == 0.5] = (stddev > 3)[cloudiness == 0.5]
if sun is None or (sun_line is None and radius < 100):
if def_clear < 0.1 * np.sum(self.helper.mask == 1):
cloudiness[np.logical_and(cloudiness == 0, rmb > -8)] = 1
cloudiness = self.helper.close_result(cloudiness)
cloudiness[self.helper.mask == 0] = 0.5
if sun is not None:
cloudiness[sun == 1] = 0.5
if sun_line is not None:
cloudiness[sun_line_dilated == 1] = 0.5
return cloudiness
def get_cloud_cover(self, cloudiness):
return np.sum(cloudiness == 1) / np.sum(cloudiness != 0.5)
class StarCheckerHelper:
def __init__(self, calibration_file):
self.calibration = Calibration()
self.calibration.load(calibration_file)
def prepare(self, path, star_finder):
with warnings.catch_warnings():
warnings.simplefilter('ignore', AstropyWarning)
self.calibration.selectImage(path)
self.names, self.vmag, alt, az = self.calibration.catalog.filter(Configuration.min_alt * u.deg, Configuration.max_mag)
altaz = np.array([alt.radian, az.radian]).transpose()
self.pos = np.column_stack(self.calibration.project(altaz))
self.finder = star_finder
self.image = cv2.imread(path)
self.finder.setImage(self.image)
self.altaz = np.array([alt.degree, az.degree]).transpose()
def count_stars(self):
min_az = 0
max_az = 360
min_alt = Configuration.min_alt
max_alt = 90
alt_step = Configuration.alt_step
az_step = Configuration.az_step
alt_bins = int((max_alt - min_alt) / alt_step)
az_bins = int((max_az - min_az) / az_step)
counts = np.zeros([alt_bins, az_bins, 4])
for alt_bin in range(alt_bins):
alt = min_alt + alt_step * alt_bin
for az_bin in range(az_bins):
az = min_az + az_step * az_bin
counts[alt_bin, az_bin, 2] = alt
counts[alt_bin, az_bin, 3] = az
for i in range(self.pos.shape[0]):
aa = self.altaz[i]
if aa[0] > alt and aa[0] <= alt + alt_step and aa[1] > az and aa[1] <= az + az_step:
counts[alt_bin, az_bin, 0] += 1
if self.finder.isStar(self.pos[i][0], self.pos[i][1]):
counts[alt_bin, az_bin, 1] += 1
return counts
def get_image(self):
result = self.image.copy()
good_color = (0, 255, 0)
bad_color = (0, 0, 255)
for i in range(self.pos.shape[0]):
if self.finder.isStar(self.pos[i][0], self.pos[i][1]):
color = good_color
else:
color = bad_color
cv2.circle(result, (int(self.pos[i][0]), int(self.pos[i][1])), 3, color)
return result
class CDSST:
def __init__(self):
path = 'frames/'
day_images = SkyCameraFile.glob(path)
day_images.sort()
times = np.array([SkyCameraFile.parseTime(x).datetime for x in day_images])
self.day_images = day_images
self.times = times
self.calibration = Calibration(catalog=SkyCatalog(True))
self.calibration.load()
self.helper = CloudDetectionHelper()
try:
with open('cd_sst.cache', 'rb') as f:
self.cache = pickle.load(f)
except:
self.cache = pd.DataFrame(index=pd.Index([], dtype=np.datetime64), columns=['pos_x', 'pos_y', 'radius', 'stripe_min', 'stripe_max'])
def save_cache(self):
with open('cd_sst.cache', 'wb') as f:
pickle.dump(self.cache, f)
def detect(self, filename):
image = cv2.imread(filename)
time = SkyCameraFile.parseTime(filename).datetime
mask = np.uint8(self.helper.get_mask(image).copy())
self.calibration.selectImage(filename)
pos = self.calibration.project()
pos = (pos[0], pos[1])
if time in self.cache.index:
sun_x, sun_y, radius, min_x, max_x = self.cache.loc[time]
sun = None
sun_pos = None
if not np.isnan(sun_x):
sun_pos = (sun_x, sun_y)
sun = CDSunRemoval.circle_mask(sun_pos, radius, mask.shape)
sun_line = None
if not np.isnan(min_x):
sun_line = np.zeros(mask.shape, np.uint8)
sun_line[:, min_x:max_x + 1] = True
else:
sun_line, min_x, max_x = CDSunRemoval.find_sun_line(image, pos)
sun, sun_pos, radius = CDSunRemoval.find_sun(pos, mask)
sun_x = sun_y = None
if sun_pos is not None:
sun_x = sun_pos[0]
sun_y = sun_pos[1]
self.cache.loc[time] = [sun_x, sun_y, radius, min_x, max_x]
if sun_pos is None:
sun_pos = pos
mask = self.helper.fullmask.copy()
mask[self.helper.mask == 0] = 0
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, np.ones((11, 11), np.uint8))
_, contours, _ = cv2.findContours(mask, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
cloudiness = np.ones(mask.shape, np.float32) * .5
did_sun = False
for contour in contours:
area = cv2.contourArea(contour)
is_sun = False
if area > 100: # TODO: try different numbers
single_contour = np.zeros(mask.shape, np.uint8)
cv2.drawContours(single_contour, [contour], 0, 1, cv2.FILLED)
if sun is not None and not did_sun:
sun_area = np.sum(sun[self.helper.mask == 1])
if area > 0.9 * sun_area:
joint_area = np.sum(np.logical_and(single_contour, sun))
if sun_area / joint_area > 0.9:
is_sun = True
if is_sun:
if sun_area * 1.2 < area:
difference = np.uint8(np.logical_and(np.logical_not(sun), single_contour))
_, contours2, _ = cv2.findContours(difference, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
# filter smaller ones here! currently done with the if at the beginning
contours += contours2
did_sun = True
if not is_sun:
cloudiness[single_contour > 0] = 1.0
b, g, r = cv2.split(np.int32(image))
mask = self.helper.mask
rmb = r - b
rmb[mask == 0] = 0
cloudiness[rmb < -10] = 0
cloudiness[rmb > 50] = 1
delta = np.timedelta64(39, 's')
delta2 = np.timedelta64(0, 's')
time_diff = time - self.times
before = np.logical_and(time_diff > delta2, time_diff < delta)
if np.sum(before) == 0:
raise ValueError('No previous image found.')
current_index = np.where(before)[0][0]
prev_img = cv2.imread(self.day_images[current_index])
gray_prev = cv2.cvtColor(prev_img, cv2.COLOR_BGR2GRAY)
gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
flow = cv2.calcOpticalFlowFarneback(gray_prev, gray_image, None, 0.5, 3, 15, 3, 5, 1.2, 0)
flow2 = -cv2.calcOpticalFlowFarneback(gray_image, gray_prev, None, 0.5, 3, 15, 3, 5, 1.2, 0)
mag1, _ = cv2.cartToPolar(flow[...,0], flow[...,1])
mag2, _ = cv2.cartToPolar(flow2[...,0], flow2[...,1])
movement = np.logical_and(mag1 > 1, mag2 > 1)
no_movement = np.logical_not(movement)
brightness = np.mean(image, 2)
cloudiness[np.logical_and(movement == 1, cloudiness != 0)] = 1
cloudiness[self.helper.mask == 0] = 1
if sun is not None:
cloudiness[sun == 1] = 1
if sun_line is not None:
sun_line_dilated = cv2.morphologyEx(sun_line, cv2.MORPH_DILATE, np.ones((1, 3)))
cloudiness[sun_line_dilated == 1] = 1
y, x = np.mgrid[0:brightness.shape[0], 0:brightness.shape[1]]
x1 = []
y1 = []
out = []
for i in range(brightness.shape[0]):
for j in range(brightness.shape[1]):
if cloudiness[i, j] != 1:
x1.append(x[i, j])
y1.append(y[i, j])
out.append(brightness[i, j])
x = np.array(x1) - sun_pos[0]
y = np.array(y1) - sun_pos[1]
out = np.array(out)
dist = np.sqrt(x * x + y * y)
A = np.array([dist, np.ones(x.shape), x, y]).transpose()
A_inv = np.linalg.pinv(A)
param = np.dot(A_inv, out)
y, x = np.mgrid[0:brightness.shape[0], 0:brightness.shape[1]]
x = x - pos[0]
y = y - pos[1]
dist = np.sqrt(x * x + y * y)
A = np.array([dist, np.ones(x.shape), x, y]).transpose()
gradient = np.dot(A, param).transpose()
rect_size = 15
rect_border = (rect_size - 1) // 2
brightness_norm = brightness - gradient
stddev = np.zeros(brightness.shape)
for y in range(image.shape[0]):
for x in range(image.shape[1]):
if cloudiness[y, x] == 1:
continue
lx = x - rect_border
rx = x + rect_border + 1
uy = y - rect_border
dy = y + rect_border + 1
if lx < 0: lx = 0
if uy < 0: uy = 0
if rx > image.shape[1]: rx = image.shape[1]
if dy > image.shape[0]: dy = image.shape[0]
mask_part = cloudiness[uy:dy, lx:rx]
stddev[y, x] = np.std(brightness_norm[uy:dy, lx:rx][mask_part != 1])
def_clear = np.sum(cloudiness == 0)
cloudiness[cloudiness == 0.5] = (stddev > 3)[cloudiness == 0.5]
if sun is None or (sun_line is None and radius < 100):
if def_clear < 0.1 * np.sum(self.helper.mask == 1):
cloudiness[np.logical_and(cloudiness == 0, rmb > -8)] = 1
cloudiness = self.helper.close_result(cloudiness)
cloudiness[self.helper.mask == 0] = 0.5
if sun is not None:
cloudiness[sun == 1] = 0.5
if sun_line is not None:
cloudiness[sun_line_dilated == 1] = 0.5
return cloudiness
def get_cloud_cover(self, cloudiness):
return np.sum(cloudiness == 1) / np.sum(cloudiness != 0.5)
