def _apply_sliding_windows(self, raster):
with rasterio.open(raster) as src:
windows = sliding_windows(self.size,
self.step_size,
height=src.height,
width=src.width)
return [(win, box(*src.window_bounds(win))) for win in windows]
def test_sliding_windows_whole_width_only():
windows = list(sliding_windows(size=2, step_size=2, width=6, height=5))
assert windows == [
# row 0
Window(col_off=0, row_off=0, width=2, height=2),
Window(col_off=2, row_off=0, width=2, height=2),
Window(col_off=4, row_off=0, width=2, height=2),
# row 1
Window(col_off=0, row_off=2, width=2, height=2),
Window(col_off=2, row_off=2, width=2, height=2),
Window(col_off=4, row_off=2, width=2, height=2),
]
def test_sliding_windows_odd_size():
windows = list(sliding_windows(size=4, step_size=2, width=7, height=9))
assert windows == [
# row 0
Window(col_off=0, row_off=0, width=4, height=4),
Window(col_off=2, row_off=0, width=4, height=4),
# row 1
Window(col_off=0, row_off=2, width=4, height=4),
Window(col_off=2, row_off=2, width=4, height=4),
# row 2
Window(col_off=0, row_off=4, width=4, height=4),
Window(col_off=2, row_off=4, width=4, height=4),
]
def predict_image(fname,
model,
size,
threshold,
step_size=None,
rescale_intensity=False,
lower_cut=2,
upper_cut=98):
if not step_size:
step_size = size
with rio.open(fname) as src:
matching_windows = []
windows = sliding_windows(
size, step_size, height=src.shape[0], width=src.shape[1])
for window in windows:
window_box = box(*src.window_bounds(window))
img = np.dstack([src.read(b, window=window) for b in range(1, 4)])
if rescale_intensity:
low, high = np.percentile(img, (lower_cut, upper_cut))
img = exposure.rescale_intensity(img, in_range=(low, high))
img = img / 255.
preds = model.predict(np.array([img]))
preds_b = preds[:, 0]
for i in np.nonzero(preds_b > threshold)[0]:
_logger.info((window, float(preds_b[i])))
reproject_window_box = reproject_shape(window_box, src.crs,
WGS84_CRS)
shape_with = ShapeWithProps(
shape=reproject_window_box, props={})
shape_with.props["prob"] = float(preds_b[i])
matching_windows.append(shape_with)
return matching_windows
def predict_image(fname,
model,
size,
threshold,
step_size=None,
rasters_contour=None,
rescale_intensity=True,
lower_cut=2,
upper_cut=98):
if not step_size:
step_size = size
if rescale_intensity:
percentiles = calculate_percentiles(fname,
lower_cut=lower_cut,
upper_cut=upper_cut)
with rio.open(fname) as src:
matching_windows = []
windows = sliding_windows(size,
step_size,
height=src.shape[0],
width=src.shape[1])
windows_and_boxes = [(w, box(*src.window_bounds(w))) for w in windows]
_logger.info('Total windows: %d', len(windows_and_boxes))
if rasters_contour:
contour_polygon, contour_crs = load_raster_contour_polygon(
rasters_contour)
contour_polygon = reproject_shape(contour_polygon, contour_crs,
src.crs)
windows_and_boxes = [(w, b) for w, b in windows_and_boxes
if contour_polygon.intersection(b)]
_logger.info(
'Total windows (after filtering with raster contour shape): %d',
len(windows_and_boxes))
total = len(windows_and_boxes) // BATCH_SIZE
for group in tqdm.tqdm(grouper(windows_and_boxes, BATCH_SIZE),
total=total):
imgs = []
window_boxes = []
for pair in group:
if pair:
window, window_box = pair
img = np.dstack(
[src.read(b, window=window) for b in range(1, 4)])
if rescale_intensity:
img = exposure.rescale_intensity(img,
in_range=percentiles)
img = resnet50.preprocess_input(img)
imgs.append(img)
window_boxes.append(window_box)
preds = model.predict(np.array(imgs))
preds_b = preds[:, 0]
for i in np.nonzero(preds_b >= threshold)[0]:
_logger.info((window, float(preds_b[i])))
window_box = window_boxes[i]
reproject_window_box = reproject_shape(window_box, src.crs,
WGS84_CRS)
s = ShapeWithProps(shape=reproject_window_box, props={})
s.props['prob'] = float(preds_b[i])
matching_windows.append(s)
return matching_windows
