| """ |
| RockGlacierNet — compute per-band normalization stats for a built patch set. |
| |
| The training data_generator z-score normalizes each band. Its built-in |
| GLOBAL_MEAN/STD were measured on the OLD glacier-centric crops and badly misfit the |
| whole-Sikkim region patches (e.g. Elev std 282 vs actual ~1150, NDVI off by +3.7 sigma). |
| This writes correct stats next to the patches; data_generator loads them automatically. |
| |
| python patch_dataset/compute_norm_stats.py --patch-size 128 |
| |
| Output: patch_dataset/patches_{size}/metadata/norm_stats.json |
| """ |
|
|
| import os |
| import sys |
| import json |
| import glob |
| import argparse |
| import numpy as np |
| import rasterio |
|
|
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
| import config as C |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--patch-size", type=int, default=C.DEFAULT_PATCH) |
| args = ap.parse_args() |
|
|
| outdir = C.patches_dir(args.patch_size) |
| feats = sorted(glob.glob(str(outdir / "train" / "features" / "*.tif"))) |
| if not feats: |
| sys.exit(f"No train features in {outdir}/train/features — build patches first.") |
|
|
| n = C.NUM_BANDS |
| |
| count = 0 |
| s1 = np.zeros(n, dtype=np.float64) |
| s2 = np.zeros(n, dtype=np.float64) |
| for fp in feats: |
| with rasterio.open(fp) as s: |
| a = s.read().astype(np.float64) |
| a = np.where(np.isfinite(a) & (a != C.NODATA_VALUE), a, 0.0) |
| flat = a.reshape(n, -1) |
| s1 += flat.sum(axis=1) |
| s2 += (flat * flat).sum(axis=1) |
| count += flat.shape[1] |
|
|
| mean = s1 / count |
| std = np.sqrt(np.maximum(s2 / count - mean * mean, 1e-12)) |
|
|
| stats = {"band_names": C.BAND_NAMES, |
| "mean": [round(float(x), 6) for x in mean], |
| "std": [round(float(x), 6) for x in std], |
| "num_train_patches": len(feats), |
| "num_pixels": int(count)} |
| out = outdir / "metadata" / "norm_stats.json" |
| with open(out, "w") as f: |
| json.dump(stats, f, indent=2) |
|
|
| print(f"Computed over {len(feats)} train patches ({count:,} px/band)") |
| print(f"{'band':7s} {'mean':>12s} {'std':>12s}") |
| for nm, m, sd in zip(C.BAND_NAMES, stats["mean"], stats["std"]): |
| print(f"{nm:7s} {m:12.4f} {sd:12.4f}") |
| print(f"\nSaved -> {out}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
