Fix two scene input.

3-sentinel-data.py

@@ -597,6 +597,51 @@ def _import_distance_to_clouds():
         ) from exc
 
 
+def _normalize_s2_grid(s2_dir: Path) -> None:
+    """Remove REFL (and companion DIST_CLOUD/GCC) files that don't match the dominant shape.
+
+    Sites at MGRS tile boundaries can have S2 downloads from two tiles with different
+    spatial extents.  For example, a site at the top edge of tile 16SDA will produce
+    narrow (e.g. 30×180) REFL files from 16SDA and full-extent (180×180) files from
+    16SDB.  EFAST requires all S2 files to share the same grid, so we keep only the
+    shape with the most pixels (largest tile coverage) and discard the rest together
+    with any already-computed companions.
+    """
+    refl_files = sorted(s2_dir.glob("*_REFL.tif"))
+    if not refl_files:
+        return
+
+    shape_to_files: dict[tuple[int, int], list[Path]] = {}
+    for f in refl_files:
+        with rasterio.open(f) as src:
+            shape: tuple[int, int] = src.shape  # type: ignore[assignment]
+        shape_to_files.setdefault(shape, []).append(f)
+
+    if len(shape_to_files) <= 1:
+        return
+
+    ref_shape = max(shape_to_files.keys(), key=lambda s: s[0] * s[1])
+    shapes_summary = ", ".join(
+        f"{s[0]}×{s[1]}({len(fs)})" for s, fs in shape_to_files.items()
+    )
+
+    n_removed = 0
+    for shape, files in shape_to_files.items():
+        if shape == ref_shape:
+            continue
+        for refl_path in files:
+            stem = refl_path.stem[: -len("_REFL")]  # e.g. "S2A_16SDA_20250106_0_L2A"
+            for companion in s2_dir.glob(f"{stem}_*.tif"):
+                companion.unlink()
+            refl_path.unlink(missing_ok=True)
+            n_removed += 1
+
+    print(
+        f"[S2-NORM] Tile boundary — shapes: {shapes_summary}; "
+        f"removed {n_removed} file-sets, kept {ref_shape[0]}×{ref_shape[1]}"
+    )
+
+
 def _rescale_dist_cloud(s2_dir: Path) -> None:
     """Ensure DIST_CLOUD values are in pixel units (not normalised to [0,1])."""
     for dc_path in s2_dir.glob("*DIST_CLOUD.tif"):
@@ -796,6 +841,7 @@ def process_site(
     items = stac_search_s2(bbox, datetime(year, 1, 1), datetime(year, 12, 31))
     print(f"[{sitename}] {len(items)} S2 items found — downloading windows...")
     download_s2_window(items, bbox, s2_out, S2_BANDS, RESOLUTION_RATIO)
+    _normalize_s2_grid(s2_out)
 
     # S2 distance-to-clouds
     print(f"[{sitename}] Computing distance-to-clouds...")