def is_between_candidate(
anc_a_points: tuple, anc_b_points: tuple, target_points: tuple,
occ_thresh: float, min_forbidden_occ_ratio: float,
target_anchor_intersect_ratio_thresh: float) -> (bool, bool):
"""
Check whether a target object lies in the convex hull of the two anchors.
@param anc_a_points: The vertices of the first anchor's 2d top face.
@param anc_b_points: The vertices of the second anchor's 2d top face.
@param target_points: The vertices of the target's 2d top face.
@param occ_thresh: By considering the target intersection ratio with the convexhull of the two anchor,
which is calculated by dividing the target intersection area to the target's area, if the ratio is
bigger than the occ_thresh, then we consider this target is between the two anchors.
@param min_forbidden_occ_ratio: used to create a range of intersection area ratios wherever any target
object occupies the convexhull with a ratio within this range, we consider this case is ambiguous and we
ignore generating between references with such combination of possible targets and those two anchors
@param target_anchor_intersect_ratio_thresh: The max allowed target-to-anchor intersection ratio, if the target
is intersecting with any of the anchors with a ratio above this thresh, we should ignore generating between
references for such combinations
@return: (bool, bool) --> (target_lies_in_convex_hull_statisfying_constraints, bad_target_anchor_combination)
"""
bad_comb = False
forbidden_occ_range = [min_forbidden_occ_ratio, occ_thresh - 0.001]
intersect_ratio_thresh = target_anchor_intersect_ratio_thresh
# Get the convex hull of all points of the two anchors
convex_hull = MultiPoint(anc_a_points + anc_b_points).convex_hull
# Get anchor a, b polygons
polygon_a = MultiPoint(anc_a_points).convex_hull
polygon_b = MultiPoint(anc_b_points).convex_hull
polygon_t = MultiPoint(target_points).convex_hull
# Candidate should fall completely/with a certain ratio in the convex_hull polygon
occ_ratio = convex_hull.intersection(polygon_t).area / polygon_t.area
if occ_ratio < occ_thresh: # The object is not in the convex-hull enough to be considered between
if forbidden_occ_range[0] < occ_ratio < forbidden_occ_range[1]:
# but also should not be causing any ambiguities for other candidate targets
bad_comb = True
return False, bad_comb
# Candidate target should never be intersecting any of the anchors
if polygon_t.intersection(
polygon_a).area / polygon_t.area > intersect_ratio_thresh:
bad_comb = True
return False, bad_comb
if polygon_t.intersection(
polygon_b).area / polygon_t.area > intersect_ratio_thresh:
bad_comb = True
return False, bad_comb
return True, bad_comb
def _calculate_shading(weight: float,
shadows: list,
site_points: MultiPoint,
heat_map: np.ndarray,
gridcell_width: float,
gridcell_height: float,
normalize_by_area=False) -> None:
"""
Update the heat_map with shading losses in POA irradiance
:param weight: loss to apply to shaded cells
:param shadows: list of shadow (Multi)Polygons for each blade angle
:param site_points: points of solar panels
:param heat_map: array with shading losses
:param gridcell_width: width of cells in the heat map
:param gridcell_height: height of cells in the heat map
:param normalize_by_area: if True, normalize weight per cell by how much area is shaded
"""
if not shadows:
return
module_width_half = gridcell_width / 2
module_height_half = gridcell_height / 2
for shadow in shadows:
if normalize_by_area:
intersecting_points = site_points.intersection(
shadow.buffer(
np.linalg.norm([gridcell_height, gridcell_width])))
else:
intersecting_points = site_points.intersection(shadow)
if intersecting_points:
if isinstance(intersecting_points, Point):
intersecting_points = (intersecting_points, )
# break up into separate instructions for minor speed up by vectorization
xs = np.array([pt.x for pt in intersecting_points])
ys = np.array([pt.y for pt in intersecting_points])
x_ind = (xs - site_points.bounds[0]) / gridcell_width
y_ind = (ys - site_points.bounds[1]) / gridcell_height
x_ind = np.round(x_ind).astype(int)
y_ind = np.round(y_ind).astype(int)
for n in range(len(intersecting_points)):
x = x_ind[n]
y = y_ind[n]
pt = intersecting_points[n]
if normalize_by_area:
cell = box(pt.x - module_width_half,
pt.y - module_height_half,
pt.x + module_width_half,
pt.y + module_height_half)
intersection = cell.intersection(shadow)
area_weight = intersection.area / cell.area
heat_map[y, x] += weight * area_weight
else:
heat_map[y, x] += weight
def intersect(poly_1,poly_2): poly_1 = np.array(poly_1) poly_1 = MultiPoint(poly_1).convex_hull poly_2 = np.array(poly_2) poly_2 = MultiPoint(poly_2).convex_hull intersect_area = poly_2.intersection(poly_1).area return intersect_area
def project_camera( corners, cam_model ):
""" -------------------------------------------------------------------------------------------------------------
Project a 3D bounding box into camera (image) space (the origin of the camera space is top left corner).
This function is taken from post_process_coords() in
nuscenes-devkit/python-sdk/nuscenes/scripts/export_2d_annotations_as_json.py
corners: [np.array] the 8 corners of the 3D box
cam_model: ???
return: [tuple of int] ??? INT OR FLOAT
coordinates of the 2D box, None if the object is outside the camera frame
------------------------------------------------------------------------------------------------------------- """
front = np.argwhere( corners[ 2, :] > 0 ) # check which corners are in front of the camera
corners = corners[ :, front.flatten() ] # and take those only
corners_p = view_points( corners, cam_model, True ) # project the 3D corners in camera space
corners_p = corners_p.T[ :, : 2 ] # take only X and Y in camera space
poly = MultiPoint( corners_p.tolist() ).convex_hull
img = box( 0, 0, img_size[ 0 ], img_size[ 1 ] )
if not poly.intersects( img ):
return None # return None if the projection is out of the camera frame
inters = poly.intersection( img )
coords = np.array( [ c for c in inters.exterior.coords ] )
min_x = min( coords[ :, 0 ] )
min_y = min( coords[ :, 1 ] )
max_x = max( coords[ :, 0 ] )
max_y = max( coords[ :, 1 ] )
return min_x, min_y, max_x, max_y
def _calculate_shading(weight: float, shadows: list,
site_points: MultiPoint,
heat_map: np.ndarray) -> None:
"""
Update the heat_map with shading losses in POA irradiance
:param weight: loss to apply to shaded cells
:param shadows: list of shadow (Multi)Polygons for each blade angle
:param site_points: points of solar panels
:param heat_map: array with shading losses
"""
if not shadows:
return
for shadow in shadows:
intersecting_points = site_points.intersection(shadow)
if intersecting_points:
if isinstance(intersecting_points, Point):
intersecting_points = (intersecting_points, )
# break up into separate instructions for minor speed up by vectorization
xs = np.array([pt.x for pt in intersecting_points])
ys = np.array([pt.y for pt in intersecting_points])
x_ind = (xs - site_points.bounds[0]) / module_width
y_ind = (ys - site_points.bounds[1]) / module_height
x_ind = np.round(x_ind).astype(int)
y_ind = np.round(y_ind).astype(int)
for x, y in zip(x_ind, y_ind):
heat_map[y, x] += weight
def sample_geoseries(geoseries: gpd.GeoSeries,
count: int,
overestimate: float = 2) -> gpd.GeoSeries:
"""Creates random geographic point samples inside a polygon/multipolygon
Args:
geoseries: geometry dataset (e.g. `gdf['geometry']`) with polygons/multipolygons
count: number of samples to generate
overestimate: scaler to generate extra samples
to toss points outside of the polygon/inside it's bounds
Returns:
points: Point geometry geoseries
"""
if type(geoseries) is gpd.GeoDataFrame:
geoseries = geoseries.geometry
polygon = geoseries.unary_union
xmin, ymin, xmax, ymax = polygon.bounds
ratio = polygon.area / polygon.envelope.area
samples = np.random.uniform((xmin, ymin), (xmax, ymax),
(int(count / ratio * overestimate), 2))
multipoint = MultiPoint(samples)
multipoint = multipoint.intersection(polygon)
sample_array = np.zeros((len(multipoint.geoms), 2))
for idx, point in enumerate(multipoint.geoms):
sample_array[idx] = (point.x, point.y)
xy = sample_array[np.random.choice(len(sample_array), count)]
points = xy_to_geoseries(xy[:, 0], xy[:, 1], crs=geoseries.crs)
return points
def post_process_coords(
corner_coords: List, imsize: Tuple[int, int] = (1600, 900)
) -> Union[Tuple[float, float, float, float], None]:
"""
Get the intersection of the convex hull of the reprojected bbox corners and the image canvas, return None if no
intersection.
:param corner_coords: Corner coordinates of reprojected bounding box.
:param imsize: Size of the image canvas.
:return: Intersection of the convex hull of the 2D box corners and the image canvas.
"""
polygon_from_2d_box = MultiPoint(corner_coords).convex_hull
img_canvas = box(0, 0, imsize[0], imsize[1])
if polygon_from_2d_box.intersects(img_canvas):
img_intersection = polygon_from_2d_box.intersection(img_canvas)
intersection_coords = np.array(
[coord for coord in img_intersection.exterior.coords])
min_x = min(intersection_coords[:, 0])
min_y = min(intersection_coords[:, 1])
max_x = max(intersection_coords[:, 0])
max_y = max(intersection_coords[:, 1])
return min_x, min_y, max_x, max_y
else:
return None
def CalculateSafeArea(points, f):
k = len(points) - f
area = MultiPoint(points).convex_hull
for subset in combinations(points, k):
#print subset
subarea = MultiPoint(subset).convex_hull
area = area.intersection(subarea)
return area
def create_map(n=5e-6):
ocean = gpd.read_file(
'C://Users//potis//Desktop//b//ne_10m_ocean_scale_rank.shp')
ocean = ocean.to_crs('ESRI:54030')
result = []
for i in range(len(ocean)):
xmin, ymin, xmax, ymax = ocean.bounds.loc[i].T.values
x = np.arange(np.floor(xmin * n) / n, np.ceil(xmax * n) / n, 1 / n)
y = np.arange(np.floor(ymin * n) / n, np.ceil(ymax * n) / n, 1 / n)
grid = np.transpose([np.tile(x, len(y)), np.repeat(y, len(x))])
points = MultiPoint(grid)
if ocean.geometry[i].is_valid:
result.append(points.intersection(ocean.geometry.loc[i]))
else:
result.append(points.intersection(ocean.geometry.loc[i].buffer(0)))
results = [j for i in result for j in i]
return gpd.GeoDataFrame(results, columns=['geometry'], crs=ocean.crs)
def _anno_to_2d_bbox(self, anno, pc_file, cam_front, lidar_top,
ego_pose_cam, ego_pose_lidar, cam_intrinsic):
# Make pixel indexes 0-based
dists = []
nusc_box = self.nusc.get_box(anno['token'])
# Move them to the ego-pose frame.
nusc_box.translate(-np.array(ego_pose_cam['translation']))
nusc_box.rotate(Quaternion(ego_pose_cam['rotation']).inverse)
# Move them to the calibrated sensor frame.
nusc_box.translate(-np.array(cam_front['translation']))
nusc_box.rotate(Quaternion(cam_front['rotation']).inverse)
dists.append(np.linalg.norm(nusc_box.center))
# Filter out the corners that are not in front of the calibrated sensor.
#Corners is a 3x8 matrix, first four corners are the ones facing forward, last 4 are ons facing backward
#(0,1) top, forward
#(2,3) bottom, forward
#(4,5) top, backward
#(6,7) bottom, backward
corners_3d = nusc_box.corners()
#Getting first 4 values of Z
dists.append(np.mean(corners_3d[2, :4]))
# z is height of object for ego pose or lidar
# y is height of object for camera frame
#TODO: Discover why this is taking the Z axis
in_front = np.argwhere(corners_3d[2, :] > 0).flatten()
corners_3d = corners_3d[:, in_front]
#print(corners_3d)
#above = np.argwhere(corners_3d[2, :] > 0).flatten()
#corners_3d = corners_3d[:, above]
# Project 3d box to 2d.
corner_coords = view_points(corners_3d, cam_intrinsic,
True).T[:, :2].tolist()
#print(corner_coords)
# Keep only corners that fall within the image.
polygon_from_2d_box = MultiPoint(corner_coords).convex_hull
img_canvas = box(0, 0, self._imwidth - 1, self._imheight - 1)
if polygon_from_2d_box.intersects(img_canvas):
img_intersection = polygon_from_2d_box.intersection(img_canvas)
intersection_coords = np.array(
[coord for coord in img_intersection.exterior.coords])
min_x = min(intersection_coords[:, 0])
min_y = min(intersection_coords[:, 1])
max_x = max(intersection_coords[:, 0])
max_y = max(intersection_coords[:, 1])
#print('contained pts {}'.format(contained_points))
return [min_x, min_y, max_x, max_y], dists
else:
return None, dists
def _sample_points_(cnt,
n_points=None,
spacing=None,
mode='uniform_random',
random_seed=None):
'''sample point within an oblong contour with shapely
[deprecated]
'''
if (n_points is None) and (spacing is None):
raise ValueError('either `spacing` or `n_points` must be specified')
if isinstance(cnt, Polygon):
pg = cnt
else:
pg = Polygon(cnt)
center, sz, angle_ = cv2.minAreaRect(np.asarray(pg.boundary, dtype=int))
w, h = sz
if n_points is None:
n_points = h * w / spacing**2
pg_rot = rotate(pg, -angle_)
x0, y0, x1, y1 = pg_rot.bounds
# calculate area ratio and increment number of sampled points accordingly
area_ratio = pg_rot.area / Polygon([(x0, y0), (x0, y1), (x1, y1),
(x1, y0)]).area
n_points = int(np.ceil(n_points / area_ratio))
if mode == 'grid':
points_rot = sample_grid(w,
h,
n_points=n_points,
spacing=spacing,
angle=angle_)
points_rot = points_rot + np.r_[x0, y0]
elif mode == 'rotated_grid':
points_rot = sample_grid(w, h, n_points=n_points, spacing=spacing)
points_rot = points_rot + np.r_[x0, y0]
elif mode == 'uniform_random':
np.random.seed(random_seed)
points_rot = np.random.rand(
n_points,
2,
) * np.r_[w, h] + np.r_[x0, y0]
else:
raise ValueError('unknown mode:%s' % mode)
points_rot = MultiPoint(asMultiPoint(points_rot))
points_rot = points_rot.intersection(pg_rot)
points = rotate(points_rot, angle_)
return points
def generate_sample_points(sample_area, spacing):
sample_bound = sample_area.bounds
x = np.arange(sample_bound[0] - spacing, sample_bound[2] + spacing,
spacing)
y = np.arange(sample_bound[1] - spacing, sample_bound[3] + spacing,
spacing)
pts = np.dstack(np.meshgrid(x, y)).reshape(-1, 2)
pts = np.hstack((pts, np.zeros((pts.shape[0], 1))))
pts = MultiPoint(pts)
pts = pts.intersection(sample_area)
pts = np.array(mapping(pts)['coordinates'])
return pts
def _identical_systemic_reward(self, world):
''' reward all agents the same for perimeter forming around reward function
Notes:
- computes surface integral over surface defined by convex hull of agents
'''
assert self.identical_rewards == True
assert world.identical_rewards == True
# reward function is designed for a single landmark case
assert len(world.landmarks) == 1
assert len(self.scenario_landmarks) == 1
# get convex hull created by agents
landmark_circle = Point(world.landmarks[0].state.p_pos).buffer(
world.landmarks[0].size)
agent_positions = [ag.state.p_pos for ag in world.agents]
agent_hull = MultiPoint(agent_positions).convex_hull
if agent_hull.geom_type == 'LineString':
# if hull is degenerate (e.g. all points collinear), then set integral to zero
return [0.0] * self.num_agents
assert agent_hull.geom_type == 'Polygon', "Unexpected geometry type for convex hull: {}".format(
agent_hull.geom_type)
assert agent_hull.is_valid, "Invalid Polygon for convex hull"
assert landmark_circle.is_valid, "Invalid Polygon for convex hull"
# find intersection of convex hull and landmark
landmark_coverage = agent_hull.intersection(landmark_circle)
# compute reward
A = agent_hull.area
B = landmark_circle.area
AB = landmark_coverage.area
w_p = _PRECISION_WEIGHT
assert w_p >= 0.0
w_c = _COVERAGE_WEIGHT
assert w_c >= 0.0
reward_signal = w_p * AB / A - w_c * (B - AB) / B
assert (reward_signal >= -w_c and reward_signal <= w_p)
# normalize by number of agents
reward_signal /= float(self.num_agents)
# assign to all agents
reward_n = [reward_signal] * self.num_agents
return reward_n
def grid_points(shape):
bds = shape.bounds
i = 24
lon_min = (bds[0] * i) // 1 / i
lon_max = (bds[2] * i + 1) // 1 / i
lat_min = (bds[1] * i) // 1 / i
lat_max = (bds[3] * i + 1) // 1 / i
grid = []
for lon in np.arange(lon_min, lon_max, 1 / i):
for lat in np.arange(lat_min, lat_max, 1 / i):
grid.append(Point(lon, lat))
grid = MultiPoint(grid)
point_list = list(grid.intersection(shape))
return point_list
def intersect_polygon(self): # Returns coordinates
"""
Creates grid with dimensions contained in self.gridsize;
Creates an intersection with the polygon;
Returned value: set of coordinates detailing only the grid
points which were overlaid on the polygon
"""
# Calculate bounds of the grid around the polygon
x_min, y_min, x_max, y_max = self.polygon.bounds
x_min = round_to_multiple(x_min, self.x_spacing)
x_max = round_to_multiple(x_max, self.x_spacing)
y_min = round_to_multiple(y_min, self.y_spacing)
y_max = round_to_multiple(y_max, self.y_spacing)
# Initialise large grid of points spaced using x & y spacing
# This is to "quantise" our grid
x_grid = np.linspace(x_min, x_max,
int((x_max - x_min) / self.x_spacing) + 2)
y_grid = np.linspace(y_min, y_max,
int((y_max - y_min) / self.y_spacing) + 1)
bounds_grid = np.transpose(
[np.tile(x_grid, len(y_grid)),
np.repeat(y_grid, len(x_grid))])
points = MultiPoint(bounds_grid)
# Shapely intersect grid with polygon and store lattice coordinates in "lattice"
# (Shapely isn't well documented enough and there might have been a better way
# to do this)
result = points.intersection(self.polygon)
lattice = np.array(result.__geo_interface__["coordinates"])
# Normalise values to integer values, e.g. 0.34 -> 1
lattice[:, 0] /= self.x_spacing
lattice[:, 1] /= self.y_spacing
lattice = np.around(lattice).astype(int)
# Move grid to begin at [0,0,(0)]
x_min = lattice[:, 0].min()
y_min = lattice[:, 1].min()
lattice[:, 0] -= x_min
lattice[:, 1] -= y_min
return lattice
def compute_intersection_area(
A_east_min_vec,
A_east_max_vec,
A_north_min_vec,
A_north_max_vec,
B_east_min,
B_east_max,
B_north_min,
B_north_max,
):
B_coords = tuple(
zip(
[B_north_min, B_north_min, B_north_max, B_north_max],
[B_east_min, B_east_max, B_east_min, B_east_max],
))
B_polygon_obj = MultiPoint(B_coords).convex_hull
num_areas = A_east_min_vec.shape[0]
intersection_areas = np.zeros(num_areas)
for area_idx in np.arange(num_areas):
A_coords = tuple(
zip(
[
A_north_min_vec[area_idx],
A_north_min_vec[area_idx],
A_north_max_vec[area_idx],
A_north_max_vec[area_idx],
],
[
A_east_min_vec[area_idx],
A_east_max_vec[area_idx],
A_east_min_vec[area_idx],
A_east_max_vec[area_idx],
],
))
A_polygon_obj = MultiPoint(A_coords).convex_hull
intersection_areas[area_idx] = B_polygon_obj.intersection(
A_polygon_obj).area
return intersection_areas
def getZones(tpolygon,d):
ozones=[]
zones=[]
for unique in udensity:
mps=MultiPoint(xy[density==unique]).buffer(d)
_d = DF.getD_l(unique,minGrowth,d)
_d = np.maximum(minDensity,_d)
ozone=mps.intersection(tpolygon)
zone=ozone.buffer(-_d*0.2).buffer(_d*0.2).removeHoles(cArea(_d*0.2)).simplify(_d*0.01)
if not zone.is_empty:
ozones.append(zone.getExterior().union(mps.buffer(-_d*0.2)))
# ozones.append(mps)
zones.append(zone)
ozones=cascaded_union(ozones)
zones=cascaded_union(zones)
# zones.plot("o-")
return zones,ozones
def visible_points(self, points, return_coords=True):
"""
Determine if points are visible - lying inside polygon
:param points: inptu points
:param return_coords: True to return coordinates of points, False to return indicies pf visible points in
input array
:return: see above
"""
if not isinstance(points, MultiPoint):
points = np.array(points)
multipoint = MultiPoint(
points
) # Convert all points at once to shapely format for performance
else:
multipoint = points
visible = multipoint.intersection(self.geometry)
if return_coords:
return np.array(visible).reshape((-1, 2))
else:
return visible
def check_next_to():
if self.type != 'closest':
return False
# if not close to each other, return false
if self.target.distance_from_other_object(self.anchor,
optimized=True) > 1.2:
return False
# Get the anchor and the target convex hull
anchor_face = self.anchor.get_bbox().z_faces()[0] # Top face
anchor_points = tuple(map(tuple,
anchor_face[:, :2])) # x, y coordinates
target_z_face = self.target.get_bbox().z_faces()[0]
target_points = tuple(map(tuple, target_z_face[:, :2]))
anchor_target_polygon = MultiPoint(anchor_points +
target_points).convex_hull
# Loop over the scan object and check no one intersects this convex hull
for o in self.scan().three_d_objects:
if o.instance_label in ['wall', 'floor']:
continue
if o.object_id in [
self.anchor.object_id, self.target.object_id
]:
continue
o_z_face = o.get_bbox().z_faces()[0]
o_points = tuple(map(tuple, o_z_face[:, :2]))
o_polygon = MultiPoint(o_points).convex_hull
# if it is found in the area between the two objects, return false
if o_polygon.intersection(
anchor_target_polygon).area / o_polygon.area > 0.5:
return False
return True
def assign_label(self, tile_starts, tile_ends, regions, region_labels):
''' calculates overlap of tile with xml regions and creates dictionary based on unique labels '''
tile_box = [tile_starts[0],
tile_starts[1]], [tile_starts[0], tile_ends[1]
], [tile_ends[0], tile_starts[1]
], [tile_ends[0], tile_ends[1]]
tile_box = list(tile_box)
tile_box = MultiPoint(tile_box).convex_hull
tile_label = {}
# create a dictionary of label/value pairs: returns percent of tile containing unique
for label in set(region_labels):
# grab regions that correspond to this label
label_list = [i for i, e in enumerate(region_labels) if e == label]
labels = tuple(regions[i] for i in label_list)
# loop over every region associated with a given label, sum the overlap
box_label = False # initialize
ov = 0 # initialize
for reg in labels:
poly = Polygon(reg)
if poly.is_valid == False:
poly = poly.buffer(0)
poly_label = tile_box.intersects(poly)
if poly_label == True:
box_label = True
ov_reg = tile_box.intersection(poly)
ov += ov_reg.area / tile_box.area
if box_label == True:
tile_label[label] = ov
# p.s. if you are curious, you can plot the polygons by the following
# plt.plot(*poly.exterior.xy) and plt.plot(*tile_box.exterior.xy)
return tile_label
def get_cuboid2d_visibility(cuboid2d, img_width, img_height):
cuboid_poly = MultiPoint(cuboid2d).convex_hull
img_poly = MultiPoint([(0, 0), (0, img_height), (img_width, img_height),
(img_width, 0)]).convex_hull
return cuboid_poly.intersection(img_poly).area / cuboid_poly.area
class PolygonAnnotatedCoordinates(GeoDataFrameWrapper):
"""
Class for retrieving ground truth cluster labels from a set of coordinate points and polygons.
From the provided 2-dim. coordinates only points within the ground truth region will be considered.
"""
def __init__(self,
coordinates: TCoordinates,
groundTruthPolygons: Union[str, Sequence[Polygon],
GeoDataFrame],
noiseLabel: Optional[int] = -1):
"""
:param coordinates: coordinates of points. These points should be spread over an area larger or equal to
the ground truth area
:param groundTruthPolygons: sequence of polygons, GeoDataFrame or path to a shapefile containing such a sequence.
The polygons represent the ground truth for clustering.
*Important*: the first polygon in the sequence is assumed to be the region within
which ground truth was provided and has to cover all remaining polygons. This also means that all non-noise
clusters in that region should be covered by a polygon
:param noiseLabel: label to associate with noise or None
"""
# The constructor might seem bloated but it really mostly does input validation for the polygons
coordinates = extractCoordinatesArray(coordinates)
if isinstance(groundTruthPolygons, str):
polygons: Sequence[Polygon] = gp.read_file(
groundTruthPolygons).geometry.values
elif isinstance(groundTruthPolygons, GeoDataFrame):
polygons: Sequence[Polygon] = groundTruthPolygons.geometry.values
else:
polygons = groundTruthPolygons
self.regionPolygon = polygons[0]
self.noiseLabel = noiseLabel
self.clusterPolygons = MultiPolygon(polygons[1:])
self.noisePolygon = self.regionPolygon.difference(self.clusterPolygons)
self.regionMultipoint = MultiPoint(coordinates).intersection(
self.regionPolygon)
if self.regionMultipoint.is_empty:
raise Exception(
f"The ground truth region contains no datapoints. "
f"This can happen if you have provided unsuitable coordinates")
self.noiseMultipoint = self.regionMultipoint.intersection(
self.noisePolygon)
if self.noiseLabel is None and not self.noisePolygon.is_empty:
raise Exception(
f"No noise_label was provided but there is noise: {len(self.noiseMultipoint)} datapoints"
f"in annotated area do not belong to any cluster polygon")
self.clustersMultipoints = []
intermediatePolygon = Polygon()
for i, clusterPolygon in enumerate(self.clusterPolygons, start=1):
if not intermediatePolygon.intersection(clusterPolygon).is_empty:
raise Exception(
f"The polygons should be non-intersecting: polygon {i} intersects with previous polygons"
)
intermediatePolygon = intermediatePolygon.union(clusterPolygon)
clusterMultipoint = self.regionMultipoint.intersection(
clusterPolygon)
if clusterMultipoint.is_empty:
raise Exception(
f"The annotated cluster for polygon {i} is empty - check your data!"
)
self.clustersMultipoints.append(clusterMultipoint)
def toGeoDF(self, crs='epsg:3857', include_noise=True):
"""
:return: GeoDataFrame with clusters as MultiPoint instance indexed by the clusters' identifiers
"""
clusters = self.clustersMultipoints
firstLabel = 0
if self.noiseLabel is not None and include_noise:
clusters = [self.noiseMultipoint] + clusters
firstLabel = self.noiseLabel
gdf = gp.GeoDataFrame(
{
"geometry":
clusters,
"identifier":
list(range(firstLabel, firstLabel + len(clusters), 1))
},
crs=crs)
gdf.set_index("identifier", drop=True, inplace=True)
return gdf
def plot(self, includeNoise=True, **kwargs):
"""
Plots the ground truth clusters
:param includeNoise:
:param kwargs:
:return:
"""
gdf = self.toGeoDF(include_noise=includeNoise)
gdf["color"] = np.random.random(len(gdf))
if includeNoise and self.noiseLabel is not None:
gdf.loc[self.noiseLabel, "color"] = 0
gdf.plot(column="color", **kwargs)
def getCoordinatesLabels(self):
"""
Extract cluster coordinates and labels as numpy arrays from the provided ground truth region and
cluster polygons
:return: tuple of arrays of the type (coordinates, labels)
"""
coords, labels = [], []
for row in self.toGeoDF(include_noise=True).itertuples():
clusterMultipoint, label = row.geometry, row.Index
coords += [[p.x, p.y] for p in clusterMultipoint]
labels += [label] * len(clusterMultipoint)
return np.array(coords), np.array(labels)
def readshp(trnfile, inDataset, pos):
# projection info from input image
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
gt = list(geotransform)
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile, 0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
# coordinate transformation from training to image projection
ct = osr.CoordinateTransformation(trnsr, imsr)
# image bands
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# number of classes
K = 1
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
if int(classid) > K:
K = int(classid)
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K += 1
# loop through the polygons
Gs = [] # train observations (data matrix)
ls = [] # class labels (lists)
classnames = []
classids = set()
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = str(feature.GetField('CLASS_ID'))
classname = feature.GetField('CLASS_NAME')
if classid not in classids:
classnames.append(classname)
classids = classids | set(classid)
# label for this ROI
y = int(classid)
l = [0 for i in range(K)]
l[y] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:, 0] = bdry[:, 0] - gt[0]
bdry[:, 1] = bdry[:, 1] - gt[3]
GT = np.mat([[gt[1], gt[2]], [gt[4], gt[5]]])
bdry = bdry * np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx, miny, maxx, maxy = map(int, list(polygon1.bounds))
pts = []
for i in range(minx, maxx + 1):
for j in range(miny, maxy + 1):
pts.append((i, j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),
dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy - miny + 1, maxx - minx + 1, len(rasterBands)))
k = 0
for band in rasterBands:
cube[:, :, k] = band.ReadAsArray(minx, miny, maxx - minx + 1,
maxy - miny + 1)
k += 1
# get the training vectors
for (x, y) in intersection:
Gs.append(cube[y - miny, x - minx, :])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
Gs = np.array(Gs)
ls = np.array(ls)
return (Gs, ls, K, classnames)
def main():
gdal.AllRegister()
path = auxil.select_directory('Input directory')
if path:
os.chdir(path)
# input image
infile = auxil.select_infile(title='Image file')
if infile:
inDataset = gdal.Open(infile,GA_ReadOnly)
cols = inDataset.RasterXSize
rows = inDataset.RasterYSize
bands = inDataset.RasterCount
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
gt = list(geotransform)
else:
print 'No geotransform available'
return
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
else:
return
pos = auxil.select_pos(bands)
if not pos:
return
N = len(pos)
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# training algorithm
trainalg = auxil.select_integer(1,msg='1:Maxlike,2:Backprop,3:Congrad,4:SVM')
if not trainalg:
return
# training data (shapefile)
trnfile = auxil.select_infile(filt='.shp',title='Train shapefile')
if trnfile:
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile,0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
else:
return
tstfile = auxil.select_outfile(filt='.tst', title='Test results file')
if not tstfile:
print 'No test output'
# outfile
outfile, outfmt = auxil.select_outfilefmt(title='Classification file')
if not outfile:
return
if trainalg in (2,3,4):
# class probabilities file, hidden neurons
probfile, probfmt = auxil.select_outfilefmt(title='Probabilities file')
else:
probfile = None
if trainalg in (2,3):
L = auxil.select_integer(8,'Number of hidden neurons')
if not L:
return
# coordinate transformation from training to image projection
ct= osr.CoordinateTransformation(trnsr,imsr)
# number of classes
K = 1
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
if int(classid)>K:
K = int(classid)
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K += 1
print '========================='
print 'supervised classification'
print '========================='
print time.asctime()
print 'image: '+infile
print 'training: '+trnfile
if trainalg == 1:
print 'Maximum Likelihood'
elif trainalg == 2:
print 'Neural Net (Backprop)'
elif trainalg ==3:
print 'Neural Net (Congrad)'
else:
print 'Support Vector Machine'
# loop through the polygons
Gs = [] # train observations
ls = [] # class labels
classnames = '{unclassified'
classids = set()
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = str(feature.GetField('CLASS_ID'))
classname = feature.GetField('CLASS_NAME')
if classid not in classids:
classnames += ', '+ classname
classids = classids | set(classid)
l = [0 for i in range(K)]
l[int(classid)] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:,0] = bdry[:,0]-gt[0]
bdry[:,1] = bdry[:,1]-gt[3]
GT = np.mat([[gt[1],gt[2]],[gt[4],gt[5]]])
bdry = bdry*np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx,miny,maxx,maxy = map(int,list(polygon1.bounds))
pts = []
for i in range(minx,maxx+1):
for j in range(miny,maxy+1):
pts.append((i,j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy-miny+1,maxx-minx+1,len(rasterBands)))
k=0
for band in rasterBands:
cube[:,:,k] = band.ReadAsArray(minx,miny,maxx-minx+1,maxy-miny+1)
k += 1
# get the training vectors
for (x,y) in intersection:
Gs.append(cube[y-miny,x-minx,:])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
classnames += '}'
m = len(ls)
print str(m) + ' training pixel vectors were read in'
Gs = np.array(Gs)
ls = np.array(ls)
# stretch the pixel vectors to [-1,1] for ffn
maxx = np.max(Gs,0)
minx = np.min(Gs,0)
for j in range(N):
Gs[:,j] = 2*(Gs[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
# random permutation of training data
idx = np.random.permutation(m)
Gs = Gs[idx,:]
ls = ls[idx,:]
# setup output datasets
driver = gdal.GetDriverByName(outfmt)
outDataset = driver.Create(outfile,cols,rows,1,GDT_Byte)
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
outDataset.SetGeoTransform(tuple(gt))
if projection is not None:
outDataset.SetProjection(projection)
outBand = outDataset.GetRasterBand(1)
if probfile:
driver = gdal.GetDriverByName(probfmt)
probDataset = driver.Create(probfile,cols,rows,K,GDT_Byte)
if geotransform is not None:
probDataset.SetGeoTransform(tuple(gt))
if projection is not None:
probDataset.SetProjection(projection)
probBands = []
for k in range(K):
probBands.append(probDataset.GetRasterBand(k+1))
if tstfile:
# train on 2/3 training examples
Gstrn = Gs[0:2*m//3,:]
lstrn = ls[0:2*m//3,:]
Gstst = Gs[2*m//3:,:]
lstst = ls[2*m//3:,:]
else:
Gstrn = Gs
lstrn = ls
if trainalg == 1:
classifier = sc.Maxlike(Gstrn,lstrn)
elif trainalg == 2:
classifier = sc.Ffnbp(Gstrn,lstrn,L)
elif trainalg == 3:
classifier = sc.Ffncg(Gstrn,lstrn,L)
elif trainalg == 4:
classifier = sc.Svm(Gstrn,lstrn)
print 'training on %i pixel vectors...' % np.shape(Gstrn)[0]
start = time.time()
result = classifier.train()
print 'elapsed time %s' %str(time.time()-start)
if result:
if trainalg in [2,3]:
cost = np.log10(result)
ymax = np.max(cost)
ymin = np.min(cost)
xmax = len(cost)
plt.plot(range(xmax),cost,'k')
plt.axis([0,xmax,ymin-1,ymax])
plt.title('Log(Cross entropy)')
plt.xlabel('Epoch')
# classify the image
print 'classifying...'
start = time.time()
tile = np.zeros((cols,N))
for row in range(rows):
for j in range(N):
tile[:,j] = rasterBands[j].ReadAsArray(0,row,cols,1)
tile[:,j] = 2*(tile[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
cls, Ms = classifier.classify(tile)
outBand.WriteArray(np.reshape(cls,(1,cols)),0,row)
if probfile:
Ms = np.byte(Ms*255)
for k in range(K):
probBands[k].WriteArray(np.reshape(Ms[k,:],(1,cols)),0,row)
outBand.FlushCache()
print 'elapsed time %s' %str(time.time()-start)
outDataset = None
inDataset = None
if probfile:
for probBand in probBands:
probBand.FlushCache()
probDataset = None
print 'class probabilities written to: %s'%probfile
K = lstrn.shape[1]+1
if (outfmt == 'ENVI') and (K<19):
# try to make an ENVI classification header file
hdr = header.Header()
headerfile = outfile+'.hdr'
f = open(headerfile)
line = f.readline()
envihdr = ''
while line:
envihdr += line
line = f.readline()
f.close()
hdr.read(envihdr)
hdr['file type'] ='ENVI Classification'
hdr['classes'] = str(K)
classlookup = '{0'
for i in range(1,3*K):
classlookup += ', '+str(str(ctable[i]))
classlookup +='}'
hdr['class lookup'] = classlookup
hdr['class names'] = classnames
f = open(headerfile,'w')
f.write(str(hdr))
f.close()
print 'thematic map written to: %s'%outfile
if trainalg in [2,3]:
print 'please close the cross entropy plot to continue'
plt.show()
if tstfile:
with open(tstfile,'w') as f:
print >>f, 'FFN test results for %s'%infile
print >>f, time.asctime()
print >>f, 'Classification image: %s'%outfile
print >>f, 'Class probabilities image: %s'%probfile
print >>f, lstst.shape[0],lstst.shape[1]
classes, _ = classifier.classify(Gstst)
labels = np.argmax(lstst,axis=1)+1
for i in range(len(classes)):
print >>f, classes[i], labels[i]
f.close()
print 'test results written to: %s'%tstfile
print 'done'
else:
print 'an error occured'
return
def draw(self, vsk: vsketch.Vsketch) -> None:
print(os.getcwd())
vsk.size("a6", landscape=False, center=False)
vsk.scale(1)
vsk.penWidth(self.pen_width)
glyph_poly = load_glyph(self.font, self.glyph, self.face_index)
# normalize glyph size
bounds = glyph_poly.bounds
scale_factor = min(
(vsk.width - 2 * self.glyph_margin) / (bounds[2] - bounds[0]),
(vsk.height - 2 * self.glyph_margin) / (bounds[3] - bounds[1]),
)
glyph_poly = scale(glyph_poly, scale_factor, scale_factor)
bounds = glyph_poly.bounds
glyph_poly = translate(
glyph_poly,
vsk.width / 2 - bounds[0] - (bounds[2] - bounds[0]) / 2,
vsk.height / 2 - bounds[1] - (bounds[3] - bounds[1]) / 2 +
self.glyph_voffset,
)
if self.draw_glyph:
vsk.strokeWeight(self.glyph_weight)
if self.fill_glyph:
vsk.fill(1)
vsk.geometry(glyph_poly)
if self.fill_glyph and self.glyph_chroma:
angle = self.glyph_chroma_angle / 180.0 * math.pi
glyph_poly_chroma1 = translate(
glyph_poly,
-self.glyph_chroma_offset * math.cos(angle),
-self.glyph_chroma_offset * math.sin(angle),
).difference(glyph_poly)
glyph_poly_chroma2 = translate(
glyph_poly,
self.glyph_chroma_offset * math.cos(angle),
self.glyph_chroma_offset * math.sin(angle),
).difference(glyph_poly)
vsk.strokeWeight(1)
vsk.stroke(2)
vsk.fill(2)
vsk.geometry(glyph_poly_chroma1)
vsk.stroke(3)
vsk.fill(3)
vsk.geometry(glyph_poly_chroma2)
glyph_poly = unary_union(
[glyph_poly, glyph_poly_chroma1, glyph_poly_chroma2])
vsk.strokeWeight(1)
vsk.stroke(1)
vsk.noFill()
glyph_shadow = None
if self.glyph_shadow:
angle = self.glyph_chroma_angle / 180.0 * math.pi
glyph_shadow = translate(
glyph_poly,
self.glyph_chroma_offset * math.cos(angle),
self.glyph_chroma_offset * math.sin(angle),
).difference(glyph_poly)
vsk.fill(3)
vsk.stroke(3)
vsk.geometry(glyph_shadow)
vsk.noFill()
vsk.stroke(1)
glyph_poly = glyph_poly.union(glyph_shadow)
if self.glyph_weight == 1:
glyph_poly_ext = glyph_poly.buffer(
self.glyph_space,
join_style=JOIN_STYLE.mitre,
)
glyph_poly_int = glyph_poly.buffer(
-self.glyph_space_inside,
join_style=JOIN_STYLE.mitre,
)
else:
buf_len = (self.glyph_weight - 1) / 2 * self.pen_width
glyph_poly_ext = glyph_poly.buffer(
buf_len * 2 + self.glyph_space,
join_style=JOIN_STYLE.mitre,
)
glyph_poly_int = glyph_poly.buffer(
-buf_len - self.glyph_space_inside,
join_style=JOIN_STYLE.mitre,
)
if glyph_shadow is not None:
glyph_poly_int = glyph_poly_int.difference(glyph_shadow)
# horizontal stripes
if self.draw_h_stripes:
count = round(
(vsk.height - 2 * self.margin) / self.h_stripes_pitch)
corrected_pitch = (vsk.height - 2 * self.margin) / count
hstripes = MultiLineString([[
(self.margin, self.margin + i * corrected_pitch),
(vsk.width - self.margin, self.margin + i * corrected_pitch),
] for i in range(count + 1)])
vsk.geometry(hstripes.difference(glyph_poly_ext))
if self.h_stripes_inside:
inside_stripes = translate(hstripes, 0, corrected_pitch /
2).intersection(glyph_poly_int)
vsk.geometry(inside_stripes)
if self.h_stripes_inside_chroma:
chroma_offset = math.sqrt(2) * self.pen_width
vsk.stroke(2)
vsk.geometry(
translate(inside_stripes, -chroma_offset,
-chroma_offset))
vsk.stroke(3)
vsk.geometry(
translate(inside_stripes, chroma_offset,
chroma_offset))
vsk.stroke(1)
# concentric
if self.draw_concentric:
circle_count = int(
math.ceil(
math.hypot(vsk.width, vsk.height) / 2 /
self.concentric_pitch))
circles = unary_union([
Point(vsk.width / 2, vsk.height / 2).buffer(
(i + 1) * self.concentric_pitch,
resolution=int(1 * (i + 1) * self.concentric_pitch),
).exterior for i in range(circle_count)
])
vsk.geometry(
circles.difference(glyph_poly_ext).intersection(
box(
self.margin,
self.margin,
vsk.width - self.margin,
vsk.height - self.margin,
)))
# dots
vsk.fill(1)
if self.draw_dots or self.draw_cut_circles:
v_pitch = self.pitch * math.tan(math.pi / 3) / 2
h_count = int((vsk.width - 2 * self.margin) // self.pitch)
v_count = int((vsk.height - 2 * self.margin) // v_pitch)
h_offset = (vsk.width - h_count * self.pitch) / 2
v_offset = (vsk.height - v_count * v_pitch) / 2
dot_array = []
for j in range(v_count + 1):
odd_line = j % 2 == 1
for i in range(h_count + (0 if odd_line else 1)):
dot = Point(
h_offset + i * self.pitch +
(self.pitch / 2 if odd_line else 0),
v_offset + j * v_pitch,
).buffer(self.thickness / 2)
if self.draw_dots:
if not dot.buffer(
self.thickness / 2).intersects(glyph_poly_ext):
dot_array.append(dot)
else:
dot_array.append(dot)
dots = unary_union(dot_array)
if self.draw_dots:
vsk.geometry(dots)
if self.draw_cut_circles:
if self.cut_circles_inside:
op_func = lambda geom: geom.intersection(glyph_poly_int)
else:
op_func = lambda geom: geom.difference(glyph_poly_ext)
vsk.geometry(op_func(dots))
if self.cut_circle_chroma:
angle = math.pi / 6
dist = self.pitch * 0.1
vsk.fill(2)
vsk.stroke(2)
vsk.geometry(
op_func(
translate(dots, -dist * math.cos(angle), -dist *
math.sin(angle)).difference(dots)))
vsk.fill(3)
vsk.stroke(3)
vsk.geometry(
op_func(
translate(dots, dist * math.cos(angle),
dist *
math.sin(angle)).difference(dots)))
vsk.fill(1)
vsk.stroke(1)
vsk.stroke(4) # apply line sort, see finalize()
if self.draw_dot_matrix:
h_count = int(
(vsk.width - 2 * self.margin) // self.dot_matrix_pitch) + 1
v_count = int(
(vsk.height - 2 * self.margin) // self.dot_matrix_pitch) + 1
h_pitch = (vsk.width - 2 * self.margin) / (h_count - 1)
v_pitch = (vsk.height - 2 * self.margin) / (v_count - 1)
mp = MultiPoint([
(self.margin + i * h_pitch, self.margin + j * v_pitch)
for i, j in itertools.product(range(h_count), range(v_count))
if vsk.random(1) < self.dot_matrix_density
])
if self.draw_dot_matrix_inside:
mp = mp.intersection(glyph_poly_int)
else:
mp = mp.difference(glyph_poly_ext)
vsk.geometry(mp)
vsk.vpype("color -l4 black")
vsk.vpype("color -l1 black color -l2 cyan color -l3 magenta")
def main():
gdal.AllRegister()
path = auxil.select_directory('Input directory')
if path:
os.chdir(path)
# input image
infile = auxil.select_infile(title='Image file')
if infile:
inDataset = gdal.Open(infile, GA_ReadOnly)
cols = inDataset.RasterXSize
rows = inDataset.RasterYSize
bands = inDataset.RasterCount
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
gt = list(geotransform)
else:
print 'No geotransform available'
return
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
else:
return
pos = auxil.select_pos(bands)
if not pos:
return
N = len(pos)
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# training algorithm
trainalg = auxil.select_integer(1,
msg='1:Maxlike,2:Backprop,3:Congrad,4:SVM')
if not trainalg:
return
# training data (shapefile)
trnfile = auxil.select_infile(filt='.shp', title='Train shapefile')
if trnfile:
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile, 0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
else:
return
tstfile = auxil.select_outfile(filt='.tst', title='Test results file')
if not tstfile:
print 'No test output'
# outfile
outfile, outfmt = auxil.select_outfilefmt(title='Classification file')
if not outfile:
return
if trainalg in (2, 3, 4):
# class probabilities file, hidden neurons
probfile, probfmt = auxil.select_outfilefmt(title='Probabilities file')
else:
probfile = None
if trainalg in (2, 3):
L = auxil.select_integer(8, 'Number of hidden neurons')
if not L:
return
# coordinate transformation from training to image projection
ct = osr.CoordinateTransformation(trnsr, imsr)
# number of classes
K = 1
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
if int(classid) > K:
K = int(classid)
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K += 1
print '========================='
print 'supervised classification'
print '========================='
print time.asctime()
print 'image: ' + infile
print 'training: ' + trnfile
if trainalg == 1:
print 'Maximum Likelihood'
elif trainalg == 2:
print 'Neural Net (Backprop)'
elif trainalg == 3:
print 'Neural Net (Congrad)'
else:
print 'Support Vector Machine'
# loop through the polygons
Gs = [] # train observations
ls = [] # class labels
classnames = '{unclassified'
classids = set()
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = str(feature.GetField('CLASS_ID'))
classname = feature.GetField('CLASS_NAME')
if classid not in classids:
classnames += ', ' + classname
classids = classids | set(classid)
l = [0 for i in range(K)]
l[int(classid)] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:, 0] = bdry[:, 0] - gt[0]
bdry[:, 1] = bdry[:, 1] - gt[3]
GT = np.mat([[gt[1], gt[2]], [gt[4], gt[5]]])
bdry = bdry * np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx, miny, maxx, maxy = map(int, list(polygon1.bounds))
pts = []
for i in range(minx, maxx + 1):
for j in range(miny, maxy + 1):
pts.append((i, j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),
dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy - miny + 1, maxx - minx + 1, len(rasterBands)))
k = 0
for band in rasterBands:
cube[:, :, k] = band.ReadAsArray(minx, miny, maxx - minx + 1,
maxy - miny + 1)
k += 1
# get the training vectors
for (x, y) in intersection:
Gs.append(cube[y - miny, x - minx, :])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
classnames += '}'
m = len(ls)
print str(m) + ' training pixel vectors were read in'
Gs = np.array(Gs)
ls = np.array(ls)
# stretch the pixel vectors to [-1,1] for ffn
maxx = np.max(Gs, 0)
minx = np.min(Gs, 0)
for j in range(N):
Gs[:, j] = 2 * (Gs[:, j] - minx[j]) / (maxx[j] - minx[j]) - 1.0
# random permutation of training data
idx = np.random.permutation(m)
Gs = Gs[idx, :]
ls = ls[idx, :]
# setup output datasets
driver = gdal.GetDriverByName(outfmt)
outDataset = driver.Create(outfile, cols, rows, 1, GDT_Byte)
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
outDataset.SetGeoTransform(tuple(gt))
if projection is not None:
outDataset.SetProjection(projection)
outBand = outDataset.GetRasterBand(1)
if probfile:
driver = gdal.GetDriverByName(probfmt)
probDataset = driver.Create(probfile, cols, rows, K, GDT_Byte)
if geotransform is not None:
probDataset.SetGeoTransform(tuple(gt))
if projection is not None:
probDataset.SetProjection(projection)
probBands = []
for k in range(K):
probBands.append(probDataset.GetRasterBand(k + 1))
if tstfile:
# train on 2/3 training examples
Gstrn = Gs[0:2 * m // 3, :]
lstrn = ls[0:2 * m // 3, :]
Gstst = Gs[2 * m // 3:, :]
lstst = ls[2 * m // 3:, :]
else:
Gstrn = Gs
lstrn = ls
if trainalg == 1:
classifier = sc.Maxlike(Gstrn, lstrn)
elif trainalg == 2:
classifier = sc.Ffnbp(Gstrn, lstrn, L)
elif trainalg == 3:
classifier = sc.Ffncg(Gstrn, lstrn, L)
elif trainalg == 4:
classifier = sc.Svm(Gstrn, lstrn)
print 'training on %i pixel vectors...' % np.shape(Gstrn)[0]
start = time.time()
result = classifier.train()
print 'elapsed time %s' % str(time.time() - start)
if result:
if trainalg in [2, 3]:
cost = np.log10(result)
ymax = np.max(cost)
ymin = np.min(cost)
xmax = len(cost)
plt.plot(range(xmax), cost, 'k')
plt.axis([0, xmax, ymin - 1, ymax])
plt.title('Log(Cross entropy)')
plt.xlabel('Epoch')
# classify the image
print 'classifying...'
start = time.time()
tile = np.zeros((cols, N))
for row in range(rows):
for j in range(N):
tile[:, j] = rasterBands[j].ReadAsArray(0, row, cols, 1)
tile[:, j] = 2 * (tile[:, j] - minx[j]) / (maxx[j] -
minx[j]) - 1.0
cls, Ms = classifier.classify(tile)
outBand.WriteArray(np.reshape(cls, (1, cols)), 0, row)
if probfile:
Ms = np.byte(Ms * 255)
for k in range(K):
probBands[k].WriteArray(np.reshape(Ms[k, :], (1, cols)), 0,
row)
outBand.FlushCache()
print 'elapsed time %s' % str(time.time() - start)
outDataset = None
inDataset = None
if probfile:
for probBand in probBands:
probBand.FlushCache()
probDataset = None
print 'class probabilities written to: %s' % probfile
K = lstrn.shape[1] + 1
if (outfmt == 'ENVI') and (K < 19):
# try to make an ENVI classification header file
hdr = header.Header()
headerfile = outfile + '.hdr'
f = open(headerfile)
line = f.readline()
envihdr = ''
while line:
envihdr += line
line = f.readline()
f.close()
hdr.read(envihdr)
hdr['file type'] = 'ENVI Classification'
hdr['classes'] = str(K)
classlookup = '{0'
for i in range(1, 3 * K):
classlookup += ', ' + str(str(ctable[i]))
classlookup += '}'
hdr['class lookup'] = classlookup
hdr['class names'] = classnames
f = open(headerfile, 'w')
f.write(str(hdr))
f.close()
print 'thematic map written to: %s' % outfile
if trainalg in [2, 3]:
print 'please close the cross entropy plot to continue'
plt.show()
if tstfile:
with open(tstfile, 'w') as f:
print >> f, 'FFN test results for %s' % infile
print >> f, time.asctime()
print >> f, 'Classification image: %s' % outfile
print >> f, 'Class probabilities image: %s' % probfile
print >> f, lstst.shape[0], lstst.shape[1]
classes, _ = classifier.classify(Gstst)
labels = np.argmax(lstst, axis=1) + 1
for i in range(len(classes)):
print >> f, classes[i], labels[i]
f.close()
print 'test results written to: %s' % tstfile
print 'done'
else:
print 'an error occured'
return
def to_stations(da: Var,
*,
stations: pd.DataFrame,
datastore: DataStore = None,
chunks: dict = None,
**kwargs) -> xr.DataArray:
if isinstance(da, camps.Variable):
da = da(datastore=datastore, chunks=chunks, **kwargs)
elif isinstance(da, xr.DataArray):
if chunks:
da = da.camps.chunk(chunks)
# Determine x and y
# Use Projected crs as common crs for grid and station
try:
x = da.camps.projx.name
y = da.camps.projy.name
except KeyError:
# projected coordinates do not exist
if da.camps.grid_mapping:
# try to make them from grid_mapping and lat/lons
da = da.metpy.assign_crs(da.camps.grid_mapping)
da = da.metpy.assign_y_x()
da = da.drop('metpy_crs')
x = da.camps.projx.name
y = da.camps.projy.name
else:
# exception for mercator data without grid_mapping
lat = da.camps.latitude
if lat.ndim == 1:
y = lat.name
lon = da.camps.longitude
if lon.ndim == 1:
x = lon.name
# ensure longitude expressed as degrees east from prime meridian with -179 and 180 bounds
lon_attrs = lon.attrs
da[lon.name] = xr.where(lon > 180, lon - 360, lon)
da[lon.name].attrs = lon_attrs
# Add the lat lon grid mapping since it didn't have one
# Latitude and longitude on the WGS 1984 datum
lat_lon_wgs84 = {
'grid_mapping_name': "latitude_longitude",
'longitude_of_prime_meridian': 0.0,
'semi_major_axis': 6378137.0,
'inverse_flattening': 298.257223563
}
gm = xr.DataArray()
gm.attrs.update(lat_lon_wgs84)
da = da.assign_coords({'lat_lon_wgs84': gm})
da.attrs['grid_mapping'] = 'lat_lon_wgs84'
pyproj_crs = CFProjection(da.camps.grid_mapping).to_pyproj()
stations['x'], stations['y'] = Proj(pyproj_crs)(stations.lon.values,
stations.lat.values)
# rechunk so that multiple chunks don't span x and y dims
if da.chunks is not None:
da = da.chunk({x: -1, y: -1})
da = da.unify_chunks()
# load x,y data
da[x].load()
da[y].load()
# make horizonontal space 1-D by stacking x and y
stacked = da.stack(xy=(x, y))
gridxy = np.column_stack((stacked[x].data, stacked[y].data))
stationxy = np.column_stack((stations.x, stations.y))
tree = cKDTree(gridxy) # fast nearest neighbor search algorith
dist_ix = tree.query(
stationxy) # find distance to nearest and index of nearest
# make a grid polygon
from shapely.geometry import Polygon, MultiPoint, Point
unit_y = xr.DataArray(np.ones(da[y].shape), dims=da[y].dims)
unit_x = xr.DataArray(np.ones(da[x].shape), dims=da[x].dims)
edge_x = edge(da[x] * unit_y) # broadcast constant x along the y dim
edge_y = edge(unit_x * da[y]) # broadcast constant y along the x dim
xy = zip(edge_x, edge_y)
grid_polygon = Polygon(xy)
# make station points
xy = zip(stations.x.values, stations.y.values)
station_points = MultiPoint(list(xy))
# determine which stations lie outside grid domain (polygon)
stations['point_str'] = pd.Series([str(p) for p in station_points])
stations['ix'] = stations.index
points_outside_grid = station_points - station_points.intersection(
grid_polygon)
if not points_outside_grid.is_empty:
if isinstance(points_outside_grid, Point):
points_outside_grid = [points_outside_grid]
ix_stations_outside_grid = stations.set_index('point_str').loc[[
str(p) for p in points_outside_grid
]].ix
else:
ix_stations_outside_grid = list() # let be empty list
print(len(ix_stations_outside_grid))
def nearest_worker(da: xr.DataArray, *, x, y, ix, ix_nan) -> xr.DataArray:
da = da.stack(station=(x,
y)) # squash the horizontal space dims into one
da = da.isel(station=ix)
da = da.drop_vars('station') # remove station coord
da.loc[{
'station': ix_nan
}] = np.nan # use integer index location to set stations outside grid to missing
return da
# make template for expressing the change in shape in map_blocks
template = da.copy()
# reshape action
template = template.stack(
station=(x, y)) # combine the lat/lon dims into one dim called station
template = template.isel(
station=[0] *
len(stations)) # select only the first; this removes the dim station
template = template.drop(
'station'
) # drop the multiindex lat/lon coord associated with 'station' from the 0th grid point
mb_kwargs = dict(x=x, y=y, ix=dist_ix[1], ix_nan=ix_stations_outside_grid)
da = xr.map_blocks(nearest_worker, da, kwargs=mb_kwargs, template=template)
# remove any metadata that may be leftover from the grid
da = da.drop_vars([x, y], errors='ignore')
# configure station metadata
# prep station coord with numeric index called 'station'
stations = stations.reset_index()
stations.index.set_names('station', inplace=True)
# assign the new coords
da = da.assign_coords({'platform_id': stations.call})
da.platform_id.attrs['standard_name'] = 'platform_id'
# assign the new coords with numeric index
da = da.assign_coords({'lat': stations.lat})
da.lat.attrs['standard_name'] = 'latitude'
da.lat.attrs['units'] = 'degrees_north'
da = da.assign_coords({'lon': stations.lon})
da.lon.attrs['standard_name'] = 'longitude'
da.lon.attrs['units'] = 'degrees_east'
# drop the numeric index;
da = da.reset_index('station', drop=True)
return da
def main():
gdal.AllRegister()
path = auxil.select_directory('Choose input directory')
if path:
os.chdir(path)
# input image
infile = auxil.select_infile(title='Choose image file')
if infile:
inDataset = gdal.Open(infile,GA_ReadOnly)
cols = inDataset.RasterXSize
rows = inDataset.RasterYSize
bands = inDataset.RasterCount
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
gt = list(geotransform)
else:
print 'No geotransform available'
return
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
else:
return
pos = auxil.select_pos(bands)
if not pos:
return
N = len(pos)
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# training data (shapefile)
trnfile = auxil.select_infile(filt='.shp',title='Choose train shapefile')
if trnfile:
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile,0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
else:
return
# hidden neurons
L = auxil.select_integer(8,'number of hidden neurons')
if not L:
return
# outfile
outfile, fmt = auxil.select_outfilefmt()
if not outfile:
return
# coordinate transformation from training to image projection
ct= osr.CoordinateTransformation(trnsr,imsr)
# number of classes
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K = int(classid)+1
print '========================='
print ' ffncg'
print '========================='
print time.asctime()
print 'image: '+infile
print 'training: '+trnfile
# loop through the polygons
Gs = [] # train observations
ls = [] # class labels
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = feature.GetField('CLASS_ID')
l = [0 for i in range(K)]
l[int(classid)] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:,0] = bdry[:,0]-gt[0]
bdry[:,1] = bdry[:,1]-gt[3]
GT = np.mat([[gt[1],gt[2]],[gt[4],gt[5]]])
bdry = bdry*np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx,miny,maxx,maxy = map(int,list(polygon1.bounds))
pts = []
for i in range(minx,maxx+1):
for j in range(miny,maxy+1):
pts.append((i,j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy-miny+1,maxx-minx+1,len(rasterBands)))
k=0
for band in rasterBands:
cube[:,:,k] = band.ReadAsArray(minx,miny,maxx-minx+1,maxy-miny+1)
k += 1
# get the training vectors
for (x,y) in intersection:
Gs.append(cube[y-miny,x-minx,:])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
m = len(ls)
print str(m) + ' training pixel vectors were read in'
Gs = np.array(Gs)
ls = np.array(ls)
# stretch the pixel vectors to [-1,1]
maxx = np.max(Gs,0)
minx = np.min(Gs,0)
for j in range(N):
Gs[:,j] = 2*(Gs[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
# random permutation of training data
idx = np.random.permutation(m)
Gs = Gs[idx,:]
ls = ls[idx,:]
# setup output dataset
driver = gdal.GetDriverByName(fmt)
outDataset = driver.Create(outfile,cols,rows,1,GDT_Byte)
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
outDataset.SetGeoTransform(tuple(gt))
if projection is not None:
outDataset.SetProjection(projection)
outBand = outDataset.GetRasterBand(1)
# train on 9/10 training examples
Gstrn = Gs[0:9*m//10,:]
lstrn = ls[0:9*m//10,:]
affn = Ffncg(Gstrn,lstrn,L)
print 'training on %i pixel vectors...' % np.shape(Gstrn)[0]
start = time.time()
cost = affn.train(epochs=epochs)
print 'elapsed time %s' %str(time.time()-start)
if cost is not None:
# cost = np.log10(cost)
ymax = np.max(cost)
ymin = np.min(cost)
xmax = len(cost)
plt.plot(range(xmax),cost,'k')
plt.axis([0,xmax,ymin-1,ymax])
plt.title('Cross entropy')
plt.xlabel('Epoch')
# classify the image
print 'classifying...'
tile = np.zeros((cols,N))
for row in range(rows):
for j in range(N):
tile[:,j] = rasterBands[j].ReadAsArray(0,row,cols,1)
tile[:,j] = 2*(tile[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
cls, _ = affn.classify(tile)
outBand.WriteArray(np.reshape(cls,(1,cols)),0,row)
outBand.FlushCache()
outDataset = None
inDataset = None
print 'thematic map written to: ' + outfile
print 'please close the cross entropy plot to continue'
plt.show()
else:
print 'an error occured'
return
print 'submitting cross-validation to multyvac'
start = time.time()
jid = mv.submit(traintst,Gs,ls,L,_layer='ms_image_analysis')
print 'submission time: %s' %str(time.time()-start)
start = time.time()
job = mv.get(jid)
result = job.get_result(job)
print 'execution time: %s' %str(time.time()-start)
print 'misclassification rate: %f' %np.mean(result)
print 'standard deviation: %f' %np.std(result)
print '--------done---------------------'
def main():
usage = '''
Usage:
---------------------------------------------------------
python %s [-p bandPositions] [- a algorithm] [-L number of hidden neurons]
[-P generate class probabilities image] filename trainShapefile
bandPositions is a list, e.g., -p [1,2,4]
algorithm 1=MaxLike
2=NNet(backprop)
3=NNet(congrad)
4=SVM
If the input file is named
path/filenbasename.ext then
The output classification file is named
path/filebasename_class.ext
the class probabilities output file is named
path/filebasename_classprobs.ext
and the test results file is named
path/filebasename_<classifier>.tst
--------------------------------------------------------''' %sys.argv[0]
options, args = getopt.getopt(sys.argv[1:],'hnPp:a:L:')
pos = None
probs = False
L = 8
trainalg = 1
graphics = True
for option, value in options:
if option == '-h':
print usage
return
elif option == '-p':
pos = eval(value)
elif option == '-n':
graphics = False
elif option == '-a':
trainalg = eval(value)
elif option == '-L':
L = eval(value)
elif option == '-P':
probs = True
if len(args) != 2:
print 'Incorrect number of arguments'
print usage
sys.exit(1)
if trainalg == 1:
algorithm = 'MaxLike'
elif trainalg == 2:
algorithm = 'NNet(Backprop)'
elif trainalg == 3:
algorithm = 'NNet(Congrad)'
else:
algorithm = 'SVM'
infile = args[0]
trnfile = args[1]
gdal.AllRegister()
if infile:
inDataset = gdal.Open(infile,GA_ReadOnly)
cols = inDataset.RasterXSize
rows = inDataset.RasterYSize
bands = inDataset.RasterCount
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
gt = list(geotransform)
else:
print 'No geotransform available'
return
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
else:
return
if pos is None:
pos = range(1,bands+1)
N = len(pos)
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# output files
path = os.path.dirname(infile)
basename = os.path.basename(infile)
root, ext = os.path.splitext(basename)
outfile = '%s/%s_class%s'%(path,root,ext)
tstfile = '%s/%s_%s.tst'%(path,root,algorithm)
if (trainalg in (2,3,4)) and probs:
# class probabilities file
probfile = '%s/%s_classprobs%s'%(path,root,ext)
else:
probfile = None
# training data
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile,0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
# coordinate transformation from training to image projection
ct = osr.CoordinateTransformation(trnsr,imsr)
# number of classes
K = 1
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
if int(classid)>K:
K = int(classid)
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K += 1
# es kann losgehen
print '========================='
print 'supervised classification'
print '========================='
print time.asctime()
print 'image: '+infile
print 'training: '+trnfile
print 'algorithm: '+algorithm
# loop through the polygons
Gs = [] # train observations
ls = [] # class labels
classnames = '{unclassified'
classids = set()
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = str(feature.GetField('CLASS_ID'))
classname = feature.GetField('CLASS_NAME')
if classid not in classids:
classnames += ', '+ classname
classids = classids | set(classid)
# label for this ROI
l = [0 for i in range(K)]
l[int(classid)] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:,0] = bdry[:,0]-gt[0]
bdry[:,1] = bdry[:,1]-gt[3]
GT = np.mat([[gt[1],gt[2]],[gt[4],gt[5]]])
bdry = bdry*np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx,miny,maxx,maxy = map(int,list(polygon1.bounds))
pts = []
for i in range(minx,maxx+1):
for j in range(miny,maxy+1):
pts.append((i,j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy-miny+1,maxx-minx+1,len(rasterBands)))
k=0
for band in rasterBands:
cube[:,:,k] = band.ReadAsArray(minx,miny,maxx-minx+1,maxy-miny+1)
k += 1
# get the training vectors
for (x,y) in intersection:
Gs.append(cube[y-miny,x-minx,:])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
classnames += '}'
m = len(ls)
print str(m) + ' training pixel vectors were read in'
Gs = np.array(Gs)
ls = np.array(ls)
# stretch the pixel vectors to [-1,1] for ffn
maxx = np.max(Gs,0)
minx = np.min(Gs,0)
for j in range(N):
Gs[:,j] = 2*(Gs[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
# random permutation of training data
idx = np.random.permutation(m)
Gs = Gs[idx,:]
ls = ls[idx,:]
# train on 2/3 training examples
Gstrn = Gs[0:2*m//3,:]
lstrn = ls[0:2*m//3,:]
Gstst = Gs[2*m//3:,:]
lstst = ls[2*m//3:,:]
# setup output datasets
driver = inDataset.GetDriver()
outDataset = driver.Create(outfile,cols,rows,1,GDT_Byte)
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
outDataset.SetGeoTransform(tuple(gt))
if projection is not None:
outDataset.SetProjection(projection)
outBand = outDataset.GetRasterBand(1)
if probfile:
probDataset = driver.Create(probfile,cols,rows,K,GDT_Byte)
if geotransform is not None:
probDataset.SetGeoTransform(tuple(gt))
if projection is not None:
probDataset.SetProjection(projection)
probBands = []
for k in range(K):
probBands.append(probDataset.GetRasterBand(k+1))
# initialize classifier
if trainalg == 1:
classifier = sc.Maxlike(Gstrn,lstrn)
elif trainalg == 2:
classifier = sc.Ffnbp(Gstrn,lstrn,L)
elif trainalg == 3:
classifier = sc.Ffncg(Gstrn,lstrn,L)
elif trainalg == 4:
classifier = sc.Svm(Gstrn,lstrn)
# train it
print 'training on %i pixel vectors...' % np.shape(Gstrn)[0]
start = time.time()
result = classifier.train()
print 'elapsed time %s' %str(time.time()-start)
if result:
if (trainalg in [2,3]) and graphics:
cost = np.log10(result)
ymax = np.max(cost)
ymin = np.min(cost)
xmax = len(cost)
plt.plot(range(xmax),cost,'k')
plt.axis([0,xmax,ymin-1,ymax])
plt.title('Log(Cross entropy)')
plt.xlabel('Epoch')
# classify the image
print 'classifying...'
start = time.time()
tile = np.zeros((cols,N),dtype=np.float32)
for row in range(rows):
for j in range(N):
tile[:,j] = rasterBands[j].ReadAsArray(0,row,cols,1)
tile[:,j] = 2*(tile[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
cls, Ms = classifier.classify(tile)
outBand.WriteArray(np.reshape(cls,(1,cols)),0,row)
if probfile:
Ms = np.byte(Ms*255)
for k in range(K):
probBands[k].WriteArray(np.reshape(Ms[k,:],(1,cols)),0,row)
outBand.FlushCache()
print 'elapsed time %s' %str(time.time()-start)
outDataset = None
inDataset = None
if probfile:
for probBand in probBands:
probBand.FlushCache()
probDataset = None
print 'class probabilities written to: %s'%probfile
K = lstrn.shape[1]+1
print 'thematic map written to: %s'%outfile
if trainalg in [2,3]:
plt.show()
if tstfile:
with open(tstfile,'w') as f:
print >>f, algorithm +'test results for %s'%infile
print >>f, time.asctime()
print >>f, 'Classification image: %s'%outfile
print >>f, 'Class probabilities image: %s'%probfile
print >>f, lstst.shape[0],lstst.shape[1]
classes, _ = classifier.classify(Gstst)
labels = np.argmax(lstst,axis=1)+1
for i in range(len(classes)):
print >>f, classes[i], labels[i]
f.close()
print 'test results written to: %s'%tstfile
print 'done'
else:
print 'an error occured'
return
def main():
usage = '''
Usage:
---------------------------------------------------------
python %s [-p bandPositions] [- a algorithm] [-L number of hidden neurons]
[-P generate class probabilities image] filename trainShapefile
bandPositions is a list, e.g., -p [1,2,4]
algorithm 1=MaxLike
2=NNet(backprop)
3=NNet(congrad)
4=SVM
If the input file is named
path/filenbasename.ext then
The output classification file is named
path/filebasename_class.ext
the class probabilities output file is named
path/filebasename_classprobs.ext
and the test results file is named
path/filebasename_<classifier>.tst
--------------------------------------------------------''' %sys.argv[0]
options, args = getopt.getopt(sys.argv[1:],'hnPp:a:L:')
pos = None
probs = False
L = 8
graphics = True
trainalg = 1
for option, value in options:
if option == '-h':
print usage
return
elif option == '-p':
pos = eval(value)
elif option == '-n':
graphics = False
elif option == '-a':
trainalg = eval(value)
elif option == '-L':
L = eval(value)
elif option == '-P':
probs = True
if len(args) != 2:
print 'Incorrect number of arguments'
print usage
sys.exit(1)
if trainalg == 1:
algorithm = 'MaxLike'
elif trainalg == 2:
algorithm = 'NNet(Backprop)'
elif trainalg == 3:
algorithm = 'NNet(Congrad)'
elif trainalg == 4:
algorithm = 'SVM'
infile = args[0]
trnfile = args[1]
gdal.AllRegister()
if infile:
inDataset = gdal.Open(infile,GA_ReadOnly)
cols = inDataset.RasterXSize
rows = inDataset.RasterYSize
bands = inDataset.RasterCount
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
gt = list(geotransform)
else:
print 'No geotransform available'
return
imsr = osr.SpatialReference()
imsr.ImportFromWkt(projection)
else:
return
if pos is None:
pos = range(1,bands+1)
N = len(pos)
rasterBands = []
for b in pos:
rasterBands.append(inDataset.GetRasterBand(b))
# output files
path = os.path.dirname(infile)
basename = os.path.basename(infile)
root, ext = os.path.splitext(basename)
outfile = '%s/%s_class%s'%(path,root,ext)
tstfile = '%s/%s_%s.tst'%(path,root,algorithm)
if (trainalg in (2,3,4)) and probs:
# class probabilities file
probfile = '%s/%s_classprobs%s'%(path,root,ext)
else:
probfile = None
# training data
trnDriver = ogr.GetDriverByName('ESRI Shapefile')
trnDatasource = trnDriver.Open(trnfile,0)
trnLayer = trnDatasource.GetLayer()
trnsr = trnLayer.GetSpatialRef()
# coordinate transformation from training to image projection
ct = osr.CoordinateTransformation(trnsr,imsr)
# number of classes
K = 1
feature = trnLayer.GetNextFeature()
while feature:
classid = feature.GetField('CLASS_ID')
if int(classid)>K:
K = int(classid)
feature = trnLayer.GetNextFeature()
trnLayer.ResetReading()
K += 1
# es kann losgehen
print '========================='
print 'supervised classification'
print '========================='
print time.asctime()
print 'image: '+infile
print 'training: '+trnfile
print 'algorithm: '+algorithm
# loop through the polygons
Gs = [] # train observations
ls = [] # class labels
classnames = '{unclassified'
classids = set()
print 'reading training data...'
for i in range(trnLayer.GetFeatureCount()):
feature = trnLayer.GetFeature(i)
classid = str(feature.GetField('CLASS_ID'))
classname = feature.GetField('CLASS_NAME')
if classid not in classids:
classnames += ', '+ classname
classids = classids | set(classid)
# label for this ROI
l = [0 for i in range(K)]
l[int(classid)] = 1.0
polygon = feature.GetGeometryRef()
# transform to same projection as image
polygon.Transform(ct)
# convert to a Shapely object
poly = shapely.wkt.loads(polygon.ExportToWkt())
# transform the boundary to pixel coords in numpy
bdry = np.array(poly.boundary)
bdry[:,0] = bdry[:,0]-gt[0]
bdry[:,1] = bdry[:,1]-gt[3]
GT = np.mat([[gt[1],gt[2]],[gt[4],gt[5]]])
bdry = bdry*np.linalg.inv(GT)
# polygon in pixel coords
polygon1 = asPolygon(bdry)
# raster over the bounding rectangle
minx,miny,maxx,maxy = map(int,list(polygon1.bounds))
pts = []
for i in range(minx,maxx+1):
for j in range(miny,maxy+1):
pts.append((i,j))
multipt = MultiPoint(pts)
# intersection as list
intersection = np.array(multipt.intersection(polygon1),dtype=np.int).tolist()
# cut out the bounded image cube
cube = np.zeros((maxy-miny+1,maxx-minx+1,len(rasterBands)))
k=0
for band in rasterBands:
cube[:,:,k] = band.ReadAsArray(minx,miny,maxx-minx+1,maxy-miny+1)
k += 1
# get the training vectors
for (x,y) in intersection:
Gs.append(cube[y-miny,x-minx,:])
ls.append(l)
polygon = None
polygon1 = None
feature.Destroy()
trnDatasource.Destroy()
classnames += '}'
m = len(ls)
print str(m) + ' training pixel vectors were read in'
Gs = np.array(Gs)
ls = np.array(ls)
# stretch the pixel vectors to [-1,1] (for ffn)
maxx = np.max(Gs,0)
minx = np.min(Gs,0)
for j in range(N):
Gs[:,j] = 2*(Gs[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
# random permutation of training data
idx = np.random.permutation(m)
Gs = Gs[idx,:]
ls = ls[idx,:]
# setup output datasets
driver = inDataset.GetDriver()
outDataset = driver.Create(outfile,cols,rows,1,GDT_Byte)
projection = inDataset.GetProjection()
geotransform = inDataset.GetGeoTransform()
if geotransform is not None:
outDataset.SetGeoTransform(tuple(gt))
if projection is not None:
outDataset.SetProjection(projection)
outBand = outDataset.GetRasterBand(1)
if probfile:
probDataset = driver.Create(probfile,cols,rows,K,GDT_Byte)
if geotransform is not None:
probDataset.SetGeoTransform(tuple(gt))
if projection is not None:
probDataset.SetProjection(projection)
probBands = []
for k in range(K):
probBands.append(probDataset.GetRasterBand(k+1))
# initialize classifier
if trainalg == 1:
classifier = sc.Maxlike(Gs,ls)
elif trainalg == 2:
classifier = sc.Ffnbp(Gs,ls,L)
elif trainalg == 3:
classifier = sc.Ffncg(Gs,ls,L)
elif trainalg == 4:
classifier = sc.Svm(Gs,ls)
# train it
print 'training on %i pixel vectors...' % np.shape(Gs)[0]
start = time.time()
result = classifier.train()
print 'elapsed time %s' %str(time.time()-start)
if result:
if (trainalg in [2,3]) and graphics:
cost = np.log10(result)
ymax = np.max(cost)
ymin = np.min(cost)
xmax = len(cost)
plt.plot(range(xmax),cost,'k')
plt.axis([0,xmax,ymin-1,ymax])
plt.title('Log(Cross entropy)')
plt.xlabel('Epoch')
plt.show()
# classify the image
print 'classifying...'
start = time.time()
tile = np.zeros((cols,N),dtype=np.float32)
for row in range(rows):
for j in range(N):
tile[:,j] = rasterBands[j].ReadAsArray(0,row,cols,1)
tile[:,j] = 2*(tile[:,j]-minx[j])/(maxx[j]-minx[j]) - 1.0
cls, Ms = classifier.classify(tile)
outBand.WriteArray(np.reshape(cls,(1,cols)),0,row)
if probfile:
Ms = np.byte(Ms*255)
for k in range(K):
probBands[k].WriteArray(np.reshape(Ms[k,:],(1,cols)),0,row)
outBand.FlushCache()
print 'elapsed time %s' %str(time.time()-start)
outDataset = None
inDataset = None
if probfile:
for probBand in probBands:
probBand.FlushCache()
probDataset = None
print 'class probabilities written to: %s'%probfile
K = ls.shape[1]+1
print 'thematic map written to: %s'%outfile
else:
print 'an error occured'
return
# cross-validation
start = time.time()
rc = Client()
print 'submitting cross-validation to %i IPython engines'%len(rc)
m = np.shape(Gs)[0]
traintest = []
for i in range(10):
sl = slice(i*m//10,(i+1)*m//10)
traintest.append( (np.delete(Gs,sl,0),np.delete(ls,sl,0), \
Gs[sl,:],ls[sl,:],L,trainalg) )
v = rc[:]
v.execute('import auxil.supervisedclass as sc')
result = v.map(crossvalidate,traintest).get()
print 'parallel execution time: %s' %str(time.time()-start)
print 'misclassification rate: %f' %np.mean(result)
print 'standard deviation: %f' %np.std(result)
points.append(l)
a = f2.readline().split(sep = '\t')
points1 = list()
a[0] = int(a[0])
a[1] = int(a[1])
a[2] = int(a[2])
a[3] = int(a[3])
points1.append([a[0], a[1]])
points1.append([a[2],a[1]])
points1.append([a[2], a[3]])
points1.append([a[0], a[3]])
polygon = MultiPoint(points).convex_hull
polygon1 = MultiPoint(points1).convex_hull
intersect = polygon.intersection(polygon1).area
if polygon1.area>0:
probability = intersect/polygon1.area
else:
probability = 0
f1.write(str(probability)+'\n')
else:
src_pts = np.float32([x.pt for x in matched1]).reshape(-1, 1, 2)
dst_pts = np.float32([x.pt for x in matched2]).reshape(-1, 1, 2)
M, mask = cv.findHomography(src_pts, dst_pts, cv.RANSAC, 6.0)
matchesMask = mask.ravel().tolist()
pts = np.float32([[167, 290], [167, 1016], [719, 1016], [719, 290]]).reshape(-1, 1, 2)
img1 = cv.polylines(img1, [np.int32(pts)], True, 255, 3, cv.LINE_AA)
dst = cv.perspectiveTransform(pts, M)