fix: read signed WALOUS rasters safely

backend/app/services/walous_land_cover_service.py

@@ -247,7 +247,10 @@ class WalousLandCoverService:
                 band = source.read(1, window=window, out_shape=(height, width), masked=True, resampling=Resampling.nearest)
                 output_transform = from_bounds(*bounds, width, height)
                 outside_scope = geometry_mask([mapping(clipped_geometry)], out_shape=(height, width), transform=output_transform, invert=False)
-                raw = np.asarray(band.filled(WalousLandCoverService.NODATA), dtype="uint8")
+                # The official 2023 GeoTIFF is signed int8 while GDAL exposes
+                # its nodata sentinel as 255. Filling before widening would
+                # therefore reject the sentinel as out of range for int8.
+                raw = np.asarray(np.ma.getdata(band), dtype="uint8")
                 invalid = np.ma.getmaskarray(band) | outside_scope
                 if source.nodata is not None:
                     invalid |= np.isclose(raw.astype("float64"), float(source.nodata))
@@ -476,7 +479,7 @@ class WalousLandCoverService:
                     raise AppError(code="WALOUS_SELECTION_OUTSIDE_DATASET", message="Selection does not overlap the persisted WALOUS raster", status_code=422)
                 clipped, transform = mask(source, [mapping(geometry)], crop=True, filled=False, indexes=[1])
                 band = np.ma.asarray(clipped[0])
-                raw = np.asarray(band.filled(WalousLandCoverService.NODATA), dtype="uint8")
+                raw = np.asarray(np.ma.getdata(band), dtype="uint8")
                 selected = geometry_mask([mapping(geometry)], out_shape=raw.shape, transform=transform, invert=True)
                 valid = selected & ~np.ma.getmaskarray(band) & (raw != WalousLandCoverService.NODATA)
                 values = raw[valid]
@@ -553,7 +556,7 @@ class WalousLandCoverService:
             scale = min(1.0, max_dimension / max(source.width, source.height))
             width, height = max(1, round(source.width * scale)), max(1, round(source.height * scale))
             values = source.read(1, out_shape=(height, width), masked=True, resampling=Resampling.nearest)
-            raw = np.asarray(values.filled(WalousLandCoverService.NODATA), dtype="uint8")
+            raw = np.asarray(np.ma.getdata(values), dtype="uint8")
             rgba = np.zeros((height, width, 4), dtype="uint8")
             for value, color in WalousLandCoverService.CLASS_COLORS.items():
                 selected = raw == value