File size: 10,746 Bytes
80e070e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
"""
RockGlacierNet β€” Build patches from region tiles.   [lives in patch_dataset/]

raw_region/{features,masks}  ->  patch_dataset/patches_{size}/{split}/{features,masks}/

  * pairs each feature tile with its mask (region tiles share the same name;
    legacy per-glacier feature_/mask_ pairs also work),
  * assigns each SOURCE TILE to train/val/test (split by tile -> no
    overlapping-patch leakage within a tile),
  * tiles every source into patch_size x patch_size GeoTIFF patches with overlap,
    keeping all glacier patches + a sampled fraction of background patches,
  * writes georeferenced GeoTIFF patches named feature_*.tif / mask_*.tif so the
    existing RockGlacierDataGenerator trains on them unchanged,
  * writes metadata/ (config.json, patch_metadata.csv, dataset_summary.csv,
    diagnostics_report.txt).

Run (rasterio only):
    python patch_dataset/build_patches.py                 # DEFAULT_PATCH, from raw_region/
    python patch_dataset/build_patches.py --patch-size 64
    python patch_dataset/build_patches.py --patch-size 256
All tunables live in patch_dataset/config.py.
"""

import os
import sys
import glob
import json
import math
import argparse
import numpy as np
import rasterio
import pandas as pd
from rasterio.windows import Window, transform as window_transform

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))   # import local config
import config as C


# ─────────────────────────────────────────────────────────────────────────────
# Geometry
# ─────────────────────────────────────────────────────────────────────────────

def padded_size(dim, patch, stride):
    if dim <= patch:
        return patch
    return math.ceil((dim - patch) / stride) * stride + patch


def iter_patches(feat, mask, patch, stride, mode):
    H, W = feat.shape[:2]
    pH, pW = padded_size(H, patch, stride), padded_size(W, patch, stride)
    dh, dw = pH - H, pW - W
    padded = dh > 0 or dw > 0
    if padded:
        feat = np.pad(feat, ((0, dh), (0, dw), (0, 0)),
                      mode=mode if mode == "reflect" else "constant")
        mask = np.pad(mask, ((0, dh), (0, dw), (0, 0)),
                      mode="constant", constant_values=0)
    for r in range(0, pH - patch + 1, stride):
        for c in range(0, pW - patch + 1, stride):
            yield feat[r:r+patch, c:c+patch, :], mask[r:r+patch, c:c+patch, :], r, c, padded


def assign_splits(ids, ratios, seed):
    rng = np.random.RandomState(seed)
    ids = sorted(ids); rng.shuffle(ids)
    n = len(ids); n_tr = int(n * ratios["train"]); n_va = int(n * (ratios["train"] + ratios["val"]))
    return {iid: ("train" if i < n_tr else ("val" if i < n_va else "test")) for i, iid in enumerate(ids)}


def _mask_path_for(feat_path, mask_dir):
    """Region tiles share the name; legacy uses feature_/mask_ prefixes."""
    base = os.path.basename(feat_path)
    for cand in (base, base.replace("feature_", "mask_"), base.replace("tile_", "tile_")):
        p = os.path.join(mask_dir, cand)
        if os.path.exists(p):
            return p
    return os.path.join(mask_dir, base)


def _id_of(feat_path):
    stem = os.path.splitext(os.path.basename(feat_path))[0]
    return stem.replace("feature_", "").replace("tile_", "")


# ─────────────────────────────────────────────────────────────────────────────
# Main
# ─────────────────────────────────────────────────────────────────────────────

def run(patch, feat_dir, mask_dir, limit):
    stride = C.stride_for(patch)
    outdir = C.patches_dir(patch)
    cfg = C.as_dict(patch)
    cfg["features_dir_used"], cfg["masks_dir_used"] = str(feat_dir), str(mask_dir)

    (outdir / "metadata").mkdir(parents=True, exist_ok=True)
    for sp in C.SPLIT_RATIOS:
        (outdir / sp / "features").mkdir(parents=True, exist_ok=True)
        (outdir / sp / "masks").mkdir(parents=True, exist_ok=True)
    with open(outdir / "metadata" / "config.json", "w") as f:
        json.dump(cfg, f, indent=2)

    feat_files = sorted(glob.glob(os.path.join(str(feat_dir), "*.tif")))
    if limit:
        feat_files = feat_files[:limit]
    ids = [_id_of(fp) for fp in feat_files]
    split_of = assign_splits(ids, C.SPLIT_RATIOS, C.RANDOM_SEED)
    print(f"patch_size={patch} stride={stride} | {len(feat_files)} source tiles -> {outdir.name}/")

    rng = np.random.RandomState(C.RANDOM_SEED)
    patch_records, image_records = [], []
    n_gen = n_kept = 0

    for fp in feat_files:
        iid = _id_of(fp); split = split_of[iid]
        mp = _mask_path_for(fp, str(mask_dir))
        if not os.path.exists(mp):
            print(f"  {iid}: no matching mask β€” skipped"); continue

        with rasterio.open(fp) as fs:
            feat = np.moveaxis(fs.read().astype(np.float32), 0, -1)
            base_transform, crs = fs.transform, fs.crs
        with rasterio.open(mp) as ms:
            mask = ms.read(1).astype(np.uint8)[:, :, np.newaxis]

        ch, cw = min(feat.shape[0], mask.shape[0]), min(feat.shape[1], mask.shape[1])
        feat, mask = feat[:ch, :cw, :], mask[:ch, :cw, :]
        feat = np.where((feat == C.NODATA_VALUE) | ~np.isfinite(feat), 0.0, feat)

        oH, oW = feat.shape[:2]; img_fg = int((mask > 0).sum())
        n_img_kept = n_img_gen = 0

        for fpat, mpat, r, c, padded in iter_patches(feat, mask, patch, stride, C.PADDING_MODE):
            g_pct = 100.0 * int((mpat > 0).sum()) / mpat.size
            if g_pct > C.MIN_GLACIER_PERCENTAGE:
                keep = True
            elif g_pct == 0.0:
                keep = (rng.rand() < C.BACKGROUND_KEEP_RATIO)
            else:
                keep = False
            n_gen += 1; n_img_gen += 1
            patch_id = f"{iid}_r{r:04d}_c{c:04d}"
            rec = {"patch_id": patch_id, "source_image_id": iid, "split": split,
                   "patch_row": r, "patch_col": c,
                   "glacier_pixel_count": int((mpat > 0).sum()),
                   "glacier_percentage": round(g_pct, 4),
                   "kept_or_filtered": "kept" if keep else "filtered",
                   "padding_applied": bool(padded)}
            for b, name in enumerate(C.BAND_NAMES):
                rec[f"mean_{name}"] = round(float(fpat[:, :, b].mean()), 6)
            patch_records.append(rec)
            if not keep:
                continue
            ptransform = window_transform(Window(c, r, patch, patch), base_transform)
            fout = outdir / split / "features" / f"feature_{patch_id}.tif"
            mout = outdir / split / "masks" / f"mask_{patch_id}.tif"
            with rasterio.open(fout, "w", driver="GTiff", height=patch, width=patch,
                               count=C.NUM_BANDS, dtype="float32", crs=crs, transform=ptransform) as dst:
                dst.write(np.moveaxis(fpat.astype(np.float32), -1, 0))
            with rasterio.open(mout, "w", driver="GTiff", height=patch, width=patch,
                               count=1, dtype="uint8", crs=crs, transform=ptransform, nodata=0) as dst:
                dst.write(mpat[:, :, 0].astype(np.uint8), 1)
            n_kept += 1; n_img_kept += 1

        image_records.append({"image_id": iid, "split": split,
                               "original_height": oH, "original_width": oW,
                               "glacier_pixels": img_fg,
                               "glacier_percentage": round(100.0 * img_fg / mask.size, 4),
                               "generated_patches": n_img_gen, "kept_patches": n_img_kept})

    patch_df, image_df = pd.DataFrame(patch_records), pd.DataFrame(image_records)
    patch_df.to_csv(outdir / "metadata" / "patch_metadata.csv", index=False)
    image_df.to_csv(outdir / "metadata" / "dataset_summary.csv", index=False)
    _diagnostics(patch_df, image_df, outdir, cfg)
    print(f"\nDone. generated={n_gen} kept={n_kept} -> {outdir}")


def _diagnostics(patch_df, image_df, outdir, cfg):
    kept = patch_df[patch_df["kept_or_filtered"] == "kept"] if len(patch_df) else patch_df
    L = []
    def h(t): L.extend(["", "─" * 56, f"  {t}", "─" * 56])
    L += ["=" * 56, "  ROCKGLACIERENET β€” PATCH BUILD DIAGNOSTICS", "=" * 56,
          f"  patch_size {cfg['patch_size']} | stride {cfg['stride']} | "
          f"bg_keep {cfg['background_keep_ratio']} | pad {cfg['padding_mode']}"]
    h("PATCH COUNTS")
    for sp in C.SPLIT_RATIOS:
        g = len(patch_df[patch_df["split"] == sp]) if len(patch_df) else 0
        k = len(kept[kept["split"] == sp]) if len(kept) else 0
        L.append(f"  {sp:5s}: generated={g:6d}  kept={k:6d}  filtered={g-k:6d}")
    L.append(f"  TOTAL: generated={len(patch_df):6d}  kept={len(kept):6d}")
    h("SOURCE TILES PER SPLIT")
    for sp in C.SPLIT_RATIOS:
        L.append(f"  {sp:5s}: {len(image_df[image_df['split']==sp]) if len(image_df) else 0} tiles")
    if len(kept):
        h("PIXEL CLASS BALANCE (kept patches)")
        tfg = int(kept["glacier_pixel_count"].sum()); tpx = len(kept) * cfg["patch_size"] ** 2
        tbg = tpx - tfg
        L.append(f"  Glacier   : {tfg:>12,}  ({100*tfg/tpx:.2f}%)")
        L.append(f"  Background: {tbg:>12,}  ({100*tbg/tpx:.2f}%)")
        if tfg: L.append(f"  Imbalance : {tbg/tfg:.1f} : 1")
        h("GLACIER COVERAGE OF KEPT PATCHES (%)")
        for lo, hi in [(0, 0.1), (0.1, 5), (5, 20), (20, 50), (50, 80), (80, 100.01)]:
            n = int(((kept["glacier_percentage"] >= lo) & (kept["glacier_percentage"] < hi)).sum())
            L.append(f"  {lo:5.1f}-{hi:5.1f}%: {n:6d}  {'β–ˆ' * min(n // 5, 40)}")
    report = "\n".join(L)
    print("\n" + report)
    with open(outdir / "metadata" / "diagnostics_report.txt", "w") as f:
        f.write(report)


def main():
    ap = argparse.ArgumentParser(description="Build GeoTIFF patches from region tiles.")
    ap.add_argument("--patch-size", type=int, default=C.DEFAULT_PATCH)
    ap.add_argument("--features-dir", default=str(C.RAW_REGION_FEATURES_DIR))
    ap.add_argument("--masks-dir", default=str(C.RAW_REGION_MASKS_DIR))
    ap.add_argument("--limit", type=int, default=None)
    args = ap.parse_args()
    run(args.patch_size, args.features_dir, args.masks_dir, args.limit)


if __name__ == "__main__":
    main()