63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153 | def build_training_chips(
tiles_dir: Path,
out_dir: Path,
chip_size: int = 512,
stride: int | None = None,
image_bands: tuple[int, ...] = (1, 2, 3),
skip_tiles: tuple[str, ...] = (),
) -> int:
"""Build proposal-decoder training chips from labeled tiles."""
stride = chip_size if stride is None else int(stride)
skip_set = set(skip_tiles)
for subdir in ("images", "masks", "chm", "prompts"):
(out_dir / subdir).mkdir(parents=True, exist_ok=True)
written = 0
for tile in _iter_tiles(tiles_dir, skip_set):
img_path = tile / "img.tif"
crowns_path = tile / "crowns.gpkg"
chm_path = tile / "chm.tif"
if not img_path.exists() or not crowns_path.exists() or not chm_path.exists():
print(f"Skipping {tile.name}: missing img.tif, crowns.gpkg, or chm.tif")
continue
with rasterio.open(img_path) as src:
height, width = src.height, src.width
transform = src.transform
crs = src.crs
image = src.read(image_bands).astype(np.float32)
with rasterio.open(chm_path) as chm_src:
chm = chm_src.read(1).astype(np.float32)
crowns = gpd.read_file(crowns_path)
if crowns.empty:
print(f"Skipping {tile.name}: empty crowns")
continue
if crowns.crs is None:
crowns = crowns.set_crs(crs)
elif crowns.crs != crs:
crowns = crowns.to_crs(crs)
tile_count = 0
chip_id = 0
for y0 in range(0, height, stride):
for x0 in range(0, width, stride):
y1 = min(y0 + chip_size, height)
x1 = min(x0 + chip_size, width)
if y1 - y0 < 1 or x1 - x0 < 1:
continue
window = rasterio.windows.Window(x0, y0, x1 - x0, y1 - y0)
win_transform = rasterio.windows.transform(window, transform)
chip_bounds = box(*rasterio.transform.array_bounds(y1 - y0, x1 - x0, win_transform))
clipped = _clip_annotations(crowns, chip_bounds)
if clipped.empty:
continue
image_chip = image[:, y0:y1, x0:x1]
chm_chip = chm[y0:y1, x0:x1]
pad_h = chip_size - image_chip.shape[1]
pad_w = chip_size - image_chip.shape[2]
if pad_h or pad_w:
image_chip = np.pad(image_chip, ((0, 0), (0, pad_h), (0, pad_w)), mode="constant")
chm_chip = np.pad(chm_chip, ((0, pad_h), (0, pad_w)), mode="constant")
base_image_chip = image_chip.astype(np.float32)
base_chm_chip = chm_chip.astype(np.float32)
for geom in clipped.geometry:
mask_chip = _rasterize_one_geometry(geom, (y1 - y0, x1 - x0), win_transform)
if pad_h or pad_w:
mask_chip = np.pad(mask_chip, ((0, pad_h), (0, pad_w)), mode="constant")
prompt = _prompt_from_mask(mask_chip)
if prompt is None:
continue
stem = f"{tile.name}_{chip_id}"
np.save(out_dir / "images" / f"{stem}.npy", base_image_chip)
np.save(out_dir / "masks" / f"{stem}.npy", mask_chip.astype(np.uint8))
np.save(out_dir / "chm" / f"{stem}.npy", base_chm_chip)
_save_prompt_json(out_dir / "prompts" / f"{stem}.json", prompt)
chip_id += 1
tile_count += 1
written += 1
print(f"{tile.name}: wrote {tile_count} training chips")
print(f"Built {written} training chips in {out_dir}")
return written
|