mapchete-eo 2026.2.0__py3-none-any.whl

mapchete_eo/platforms/sentinel2/product.py

@@ -0,0 +1,747 @@
+from __future__ import annotations
+
+import logging
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import numpy.ma as ma
+from mapchete.io.raster import ReferencedRaster, read_raster_window, resample_from_array
+from mapchete.geometry import reproject_geometry, buffer_antimeridian_safe
+from mapchete.path import MPath
+from mapchete.protocols import GridProtocol
+from mapchete.types import Bounds, Grid, NodataVals
+from pystac import Item
+from rasterio.enums import Resampling
+from rasterio.features import rasterize
+from shapely.geometry import shape
+
+
+from mapchete_eo.array.buffer import buffer_array
+from mapchete_eo.io.items import get_item_property
+from mapchete_eo.platforms.sentinel2.brdf.config import BRDFModels
+from mapchete_eo.platforms.sentinel2.brdf.correction import apply_correction
+from mapchete_eo.exceptions import (
+    AllMasked,
+    AssetError,
+    BRDFError,
+    CorruptedProduct,
+    EmptyFootprintException,
+    EmptyProductException,
+)
+from mapchete_eo.io.assets import get_assets, read_mask_as_raster
+from mapchete_eo.io.path import asset_mpath, get_product_cache_path
+from mapchete_eo.io.profiles import COGDeflateProfile
+from mapchete_eo.platforms.sentinel2.brdf import correction_values
+from mapchete_eo.platforms.sentinel2.bandpass_adjustment import (
+    apply_bandpass_adjustment,
+)
+from mapchete_eo.platforms.sentinel2.config import (
+    BRDFConfig,
+    BRDFModelConfig,
+    CacheConfig,
+    MaskConfig,
+)
+from mapchete_eo.platforms.sentinel2.metadata_parser.s2metadata import S2Metadata
+from mapchete_eo.platforms.sentinel2.types import (
+    CloudType,
+    L2ABand,
+    ProductQIMaskResolution,
+    Resolution,
+)
+from mapchete_eo.product import EOProduct, add_to_blacklist
+from mapchete_eo.protocols import EOProductProtocol
+from mapchete_eo.settings import mapchete_eo_settings
+
+logger = logging.getLogger(__name__)
+
+
+class Cache:
+    item: Item
+    config: CacheConfig
+    path: MPath
+
+    def __init__(self, item: Item, config: CacheConfig):
+        self.item = item
+        self.config = config
+        # TODO: maybe move this function here
+        self.path = get_product_cache_path(
+            self.item,
+            MPath.from_inp(self.config.path),
+            self.config.product_path_generation_method,
+        )
+        self.path.makedirs()
+        self._brdf_grid_cache: dict = dict()
+        if self.config.brdf:
+            self._brdf_bands = [
+                asset_name_to_l2a_band(self.item, band)
+                for band in self.config.brdf.bands
+            ]
+        else:
+            self._brdf_bands = []
+        try:
+            self._existing_files = self.path.ls()
+        except FileNotFoundError:
+            self._existing_files = None
+
+    def __repr__(self):
+        return f"<Cache: product={self.item.id}, path={self.path}>"
+
+    def cache_assets(self):
+        # cache assets
+        if self.config.assets:
+            # TODO determine already existing assets
+            self.item = get_assets(
+                self.item,
+                self.config.assets,
+                self.path,
+                resolution=self.config.assets_resolution.value,
+                ignore_if_exists=True,
+                item_href_in_dst_dir=False,
+            )
+            return self.item
+
+    def cache_brdf_grids(self, metadata: S2Metadata):
+        if self.config.brdf:
+            resolution = self.config.brdf.resolution
+            model = self.config.brdf.model
+
+            logger.debug(
+                f"prepare BRDF model '{model}' for product bands {self._brdf_bands} in {resolution} resolution"
+            )
+            for band in self._brdf_bands:
+                out_path = self.path / f"brdf_{model}_{band.name}_{resolution}.tif"
+                # TODO: do check with _existing_files again to reduce S3 requests
+                if not out_path.exists():
+                    try:
+                        grid = correction_values(
+                            metadata,
+                            band,
+                            model=model,
+                            resolution=resolution,
+                            per_detector=self.config.brdf.per_detector_correction,
+                        )
+                    except BRDFError as exc:
+                        error_msg = (
+                            f"product {self.item.get_self_href()} is corrupted: {exc}"
+                        )
+                        logger.error(error_msg)
+                        add_to_blacklist(self.item.get_self_href())
+                        raise CorruptedProduct(error_msg)
+
+                    logger.debug("cache BRDF correction grid to %s", out_path)
+                    grid.to_file(out_path, **COGDeflateProfile(grid.meta))
+                self._brdf_grid_cache[band] = out_path
+
+    def get_brdf_grid(self, band: L2ABand):
+        try:
+            return self._brdf_grid_cache[band]
+        except KeyError:
+            if band in self._brdf_bands:
+                raise KeyError(f"BRDF grid for band {band} not yet cached")
+            else:
+                raise KeyError(f"BRDF grid for band {band} not configured")
+
+
+class S2Product(EOProduct, EOProductProtocol):
+    """
+    Sentinel-2 specific EOProduct implementation.
+    """
+
+    _item_dict: Optional[dict] = None
+    cache: Optional[Cache] = None
+    _scl_cache: Dict[GridProtocol, np.ndarray]
+    _item_property_cache: Dict[str, Any]
+
+    def __init__(
+        self,
+        item: Item,
+        metadata: Optional[S2Metadata] = None,
+        cache_config: Optional[CacheConfig] = None,
+        metadata_mapper: Optional[Callable[[Item], S2Metadata]] = None,
+        item_modifier_funcs: Optional[List[Callable[[Item], Item]]] = None,
+        lazy_load_item: bool = False,
+        item_property_cache: Optional[Dict[str, Any]] = None,
+    ):
+        if lazy_load_item:
+            self._item_dict = None
+        else:
+            self._item_dict = item.to_dict()
+        self.item_href = item.self_href
+        self.id = item.id
+
+        self._metadata = metadata
+        self._metadata_mapper = metadata_mapper
+        self._item_modifier_funcs = item_modifier_funcs
+        self._scl_cache = dict()
+        self._item_property_cache = item_property_cache or dict()
+        self.cache = Cache(item, cache_config) if cache_config else None
+
+        self.__geo_interface__ = item.geometry
+        self.bounds = Bounds.from_inp(shape(self))
+        self.crs = mapchete_eo_settings.default_catalog_crs
+
+    @classmethod
+    def from_stac_item(
+        self,
+        item: Item,
+        cache_config: Optional[CacheConfig] = None,
+        cache_all: bool = False,
+        **kwargs,
+    ) -> S2Product:
+        s2product = S2Product(item, cache_config=cache_config, **kwargs)
+
+        if cache_all:
+            # cache assets if configured
+            s2product.cache_assets()
+
+            # cache BRDF grids if configured
+            s2product.cache_brdf_grids()
+
+        return s2product
+
+    @property
+    def item(self) -> Item:
+        if not self._item:
+            if self._item_dict:
+                self._item = Item.from_dict(self._item_dict)
+            else:
+                item = Item.from_file(self.item_href)
+                for modifier in self._item_modifier_funcs or []:
+                    item = modifier(item)
+                self._item = item
+        return self._item
+
+    @property
+    def metadata(self) -> S2Metadata:
+        if not self._metadata:
+            if self._metadata_mapper:
+                self._metadata = self._metadata_mapper(self.item)
+            else:
+                self._metadata = S2Metadata.from_stac_item(self.item)
+        return self._metadata
+
+    def __repr__(self):
+        return f"<S2Product product_id={self.id}>"
+
+    def clear_cached_data(self):
+        if self._metadata is not None:
+            self._metadata.clear_cached_data()
+            self._metadata = None
+        if self._item is not None:
+            self._item = None
+        self._item_property_cache = dict()
+        self._scl_cache = dict()
+
+    def read_np_array(
+        self,
+        assets: Optional[List[str]] = None,
+        eo_bands: Optional[List[str]] = None,
+        grid: Union[GridProtocol, Resolution] = Resolution["10m"],
+        resampling: Resampling = Resampling.nearest,
+        nodatavals: NodataVals = None,
+        raise_empty: bool = True,
+        apply_offset: bool = True,
+        apply_scale: bool = False,
+        apply_sentinel2_bandpass_adjustment: bool = False,
+        mask_config: MaskConfig = MaskConfig(),
+        brdf_config: Optional[BRDFConfig] = None,
+        fill_value: int = 0,
+        read_mask: Optional[np.ndarray] = None,
+        **kwargs,
+    ) -> ma.MaskedArray:
+        """
+        Read Sentinel-2 assets into a MaskedArray with masks and BRDF.
+        """
+        assets = assets or []
+        eo_bands = eo_bands or []
+        apply_offset = apply_offset and not self.metadata.boa_offset_applied
+        if eo_bands:
+            count = len(eo_bands)
+            raise NotImplementedError("please use asset names for now")
+        else:
+            count = len(assets)
+        if isinstance(grid, Resolution):
+            grid = self.metadata.grid(grid)
+        mask = self.get_mask(
+            grid, mask_config, target_mask=None if read_mask is None else ~read_mask
+        ).data
+        if nodatavals is None:
+            nodatavals = fill_value
+        elif fill_value is None and nodatavals is not None:
+            fill_value = nodatavals
+        if mask.all():
+            if raise_empty:
+                raise EmptyProductException(
+                    f"{self}: configured mask over {grid} covers everything"
+                )
+            else:
+                return self.empty_array(count, grid=grid, fill_value=fill_value)
+
+        arr = super().read_np_array(
+            assets=assets,
+            eo_bands=eo_bands,
+            grid=grid,
+            resampling=resampling,
+            raise_empty=False,
+            apply_offset=apply_offset,
+            apply_scale=apply_scale,
+        )
+
+        # bring mask to same shape as data array
+        expanded_mask = np.repeat(np.expand_dims(mask, axis=0), arr.shape[0], axis=0)
+        arr.set_fill_value(fill_value)
+        arr[expanded_mask] = fill_value
+        arr[expanded_mask] = ma.masked
+
+        if arr.mask.all():
+            if raise_empty:
+                raise EmptyProductException(
+                    f"{self}: is empty over {grid} after reading bands and applying all masks"
+                )
+            else:
+                return self.empty_array(count, grid=grid, fill_value=fill_value)
+
+        # apply Sentinel-2 bandpass adjustment
+        if apply_sentinel2_bandpass_adjustment:
+            arr = self._apply_sentinel2_bandpass_adjustment(
+                uncorrected=arr, assets=assets
+            )
+
+        # apply BRDF config if required
+        if brdf_config:
+            arr = self._apply_brdf(
+                uncorrected=arr,
+                assets=assets,
+                brdf_config=brdf_config,
+                grid=grid,
+                resampling=resampling,
+                mask_config=mask_config,
+            )
+
+        return ma.MaskedArray(arr, fill_value=fill_value)
+
+    def cache_assets(self) -> None:
+        if self.cache is not None:
+            self.cache.cache_assets()
+
+    def cache_brdf_grids(self) -> None:
+        if self.cache is not None:
+            self.cache.cache_brdf_grids(self.metadata)
+
+    def read_brdf_grid(
+        self,
+        band: L2ABand,
+        resampling: Resampling = Resampling.bilinear,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        brdf_config: BRDFModelConfig = BRDFConfig(),
+    ) -> np.ndarray:
+        grid = (
+            self.metadata.grid(grid)
+            if isinstance(grid, Resolution)
+            else Grid.from_obj(grid)
+        )
+        try:
+            # read cached file if configured
+            if self.cache:
+                return read_raster_window(
+                    self.cache.get_brdf_grid(band),
+                    grid=grid,
+                    resampling=resampling,
+                )
+            # calculate on the fly
+            return resample_from_array(
+                correction_values(
+                    self.metadata,
+                    band,
+                    model=brdf_config.model,
+                    resolution=brdf_config.resolution,
+                    footprints_cached_read=brdf_config.footprints_cached_read,
+                    per_detector=brdf_config.per_detector_correction,
+                ),
+                out_grid=grid,
+                resampling=resampling,
+                keep_2d=True,
+            )
+        except (AssetError, BRDFError) as exc:
+            error_msg = f"product {self.item.get_self_href()} is corrupted: {exc}"
+            logger.error(error_msg)
+            add_to_blacklist(self.item.get_self_href())
+            raise CorruptedProduct(error_msg)
+
+    def read_l1c_cloud_mask(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        cloud_type: CloudType = CloudType.all,
+        cached_read: bool = False,
+    ) -> ReferencedRaster:
+        """Return classification cloud mask."""
+        logger.debug("read classification cloud mask for %s", str(self))
+        return self.metadata.l1c_cloud_mask(
+            cloud_type, dst_grid=grid, cached_read=cached_read
+        )
+
+    def read_snow_ice_mask(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        cached_read: bool = False,
+    ) -> ReferencedRaster:
+        """Return classification snow and ice mask."""
+        logger.debug("read classification snow and ice mask for %s", str(self))
+        return self.metadata.snow_ice_mask(dst_grid=grid, cached_read=cached_read)
+
+    def read_cloud_probability(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        resampling: Resampling = Resampling.bilinear,
+        from_resolution: ProductQIMaskResolution = ProductQIMaskResolution["20m"],
+        cached_read: bool = False,
+    ) -> ReferencedRaster:
+        """Return cloud probability mask."""
+        if "cloud" in self.item.assets:
+            logger.debug("read cloud probability mask for %s from asset", str(self))
+            return read_mask_as_raster(
+                path=asset_mpath(item=self.item, asset="cloud"),
+                dst_grid=(
+                    self.metadata.grid(grid)
+                    if isinstance(grid, Resolution)
+                    else Grid.from_obj(grid)
+                ),
+                resampling=resampling,
+                rasterize_value_func=lambda feature: True,
+                masked=False,
+                cached_read=cached_read,
+            )
+        logger.debug(
+            "read cloud probability mask for %s from metadata archive", str(self)
+        )
+        return self.metadata.cloud_probability(
+            dst_grid=grid,
+            resampling=resampling,
+            from_resolution=from_resolution,
+            cached_read=cached_read,
+        )
+
+    def read_snow_probability(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        resampling: Resampling = Resampling.bilinear,
+        from_resolution: ProductQIMaskResolution = ProductQIMaskResolution["20m"],
+        cached_read: bool = False,
+    ) -> ReferencedRaster:
+        """Return classification snow and ice mask."""
+        if "snow" in self.item.assets:
+            logger.debug("read snow probability mask for %s from asset", str(self))
+            return read_mask_as_raster(
+                path=asset_mpath(item=self.item, asset="cloud"),
+                dst_grid=(
+                    self.metadata.grid(grid)
+                    if isinstance(grid, Resolution)
+                    else Grid.from_obj(grid)
+                ),
+                resampling=resampling,
+                rasterize_value_func=lambda feature: True,
+                masked=False,
+                cached_read=cached_read,
+            )
+        logger.debug(
+            "read snow probability mask for %s from metadata archive", str(self)
+        )
+        return self.metadata.snow_probability(
+            dst_grid=grid,
+            resampling=resampling,
+            from_resolution=from_resolution,
+            cached_read=cached_read,
+        )
+
+    def read_scl(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["20m"],
+        cached_read: bool = False,
+    ) -> ReferencedRaster:
+        """
+        Read Scene Classification Layer mask.
+        """
+        grid = (
+            self.metadata.grid(grid)
+            if isinstance(grid, Resolution)
+            else Grid.from_obj(grid)
+        )
+        grid_hash = hash((grid.transform, grid.shape))
+        if grid_hash not in self._scl_cache:
+            logger.debug("read SCL mask for %s", str(self))
+            self._scl_cache[grid_hash] = read_mask_as_raster(
+                asset_mpath(self.item, "scl"),
+                dst_grid=grid,
+                resampling=Resampling.nearest,
+                masked=True,
+                cached_read=cached_read,
+            )
+        return self._scl_cache[grid_hash]
+
+    def footprint_nodata_mask(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["10m"],
+        buffer_m: float = 0,
+    ) -> ReferencedRaster:
+        """Return rasterized footprint mask."""
+        grid = (
+            self.metadata.grid(grid)
+            if isinstance(grid, Resolution)
+            else Grid.from_obj(grid)
+        )
+        if buffer_m:
+            footprint = buffer_antimeridian_safe(shape(self), buffer_m=buffer_m)
+            if footprint.is_empty:
+                raise EmptyFootprintException(
+                    f"buffer value of {buffer_m} results in an empty geometry for footprint {shape(self).wkt}"
+                )
+        else:
+            footprint = shape(self)
+
+        return ReferencedRaster(
+            rasterize(
+                [
+                    reproject_geometry(
+                        footprint,
+                        self.crs,
+                        grid.crs,
+                        # CRS Bounds are sometimes smaller than (Mapchete) Grid Bounds,
+                        # if clipping allowed it will mask out features at CRS Bounds border,
+                        # therefore clip_to_crs_bounds: False; see mapchete.geometry.reproject reproject_geometry
+                        clip_to_crs_bounds=False,
+                    )
+                ],
+                out_shape=grid.shape,
+                transform=grid.transform,
+                all_touched=True,
+                fill=1,
+                default_value=0,
+            ).astype(bool),
+            transform=grid.transform,
+            bounds=grid.bounds,
+            crs=grid.crs,
+        )
+
+    def get_mask(
+        self,
+        grid: Union[GridProtocol, Resolution] = Resolution["10m"],
+        mask_config: MaskConfig = MaskConfig(),
+        target_mask: Optional[np.ndarray] = None,
+    ) -> ReferencedRaster:
+        """
+        Merge all configured masks into one.
+        """
+        grid = (
+            self.metadata.grid(grid)
+            if isinstance(grid, Resolution)
+            else Grid.from_obj(grid)
+        )
+        if target_mask is None:
+            target_mask = np.zeros(shape=grid.shape, dtype=bool)
+        else:
+            if target_mask.shape != grid.shape:
+                raise ValueError("a target mask must have the same shape as the grid")
+            logger.debug("got custom target mask to start with: %s", target_mask.shape)
+
+        def _check_full(arr):
+            # ATTENTION: target_mask and out have to be combined *after* mask was buffered!
+            # use 'logical or' not '+' !!!
+            if (arr | target_mask).all():
+                raise AllMasked()
+
+        out = np.zeros(shape=grid.shape, dtype=bool)
+        logger.debug("generate mask for product %s ...", str(self))
+        try:
+            _check_full(out)
+            if mask_config.footprint:
+                logger.debug("generate footprint nodata mask ...")
+                try:
+                    out |= self.footprint_nodata_mask(
+                        grid, buffer_m=mask_config.footprint_buffer_m
+                    ).data
+                    _check_full(out)
+                except EmptyFootprintException:
+                    raise AllMasked()
+            if mask_config.l1c_cloud_type:
+                logger.debug("generate L1C mask ...")
+                out |= self.read_l1c_cloud_mask(
+                    grid,
+                    mask_config.l1c_cloud_type,
+                    cached_read=mask_config.l1c_cloud_mask_cached_read,
+                ).data
+                _check_full(out)
+            if mask_config.cloud_probability_threshold != 100:
+                logger.debug(
+                    "generate cloud probability (%s) mask ...",
+                    mask_config.cloud_probability_threshold,
+                )
+                cld_prb = self.read_cloud_probability(
+                    grid,
+                    from_resolution=mask_config.cloud_probability_resolution,
+                    cached_read=mask_config.cloud_probability_cached_read,
+                ).data
+                out |= np.where(
+                    cld_prb >= mask_config.cloud_probability_threshold, True, False
+                )
+                _check_full(out)
+            if mask_config.scl_classes:
+                logger.debug(
+                    "generate SCL mask using %s ...",
+                    ", ".join(
+                        [scl_class.name for scl_class in mask_config.scl_classes]
+                    ),
+                )
+                # convert SCL classes to pixel values
+                scl_values = [scl.value for scl in mask_config.scl_classes]
+                # read SCL mask
+                scl_arr = self.read_scl(
+                    grid, cached_read=mask_config.scl_cached_read
+                ).data
+                # mask out specific pixel values
+                out |= np.isin(scl_arr, scl_values)
+                _check_full(out)
+            if mask_config.snow_ice:
+                logger.debug("generate snow & ice mask ...")
+                out |= self.read_snow_ice_mask(
+                    grid, cached_read=mask_config.snow_ice_mask_cached_read
+                ).data
+                _check_full(out)
+            if mask_config.snow_probability_threshold != 100:
+                logger.debug(
+                    "generate snow probability (%s) mask ...",
+                    mask_config.snow_probability_threshold,
+                )
+                snw_prb = self.read_snow_probability(
+                    grid,
+                    from_resolution=mask_config.snow_probability_resolution,
+                    cached_read=mask_config.snow_probability_cached_read,
+                ).data
+                out |= np.where(
+                    snw_prb >= mask_config.snow_probability_threshold, True, False
+                )
+                _check_full(out)
+            if mask_config.buffer:
+                logger.debug(
+                    "apply buffer (%s) to combined mask ...", mask_config.buffer
+                )
+                out = buffer_array(array=out, buffer=mask_config.buffer)
+                _check_full(out)
+        except AllMasked:
+            logger.debug(
+                "mask for product %s already full, skip reading other masks", self.id
+            )
+        except FileNotFoundError as exc:  # pragma: no cover
+            raise CorruptedProduct from exc
+
+        # ATTENTION: target_mask and out have to be combined *after* mask was buffered!
+        # use 'logical or' not '+' !!!
+        return ReferencedRaster(
+            out | target_mask,
+            transform=grid.transform,
+            crs=grid.crs,
+            bounds=grid.bounds,
+        )
+
+    def get_property(self, name: str) -> Any:
+        if name not in self._item_property_cache:
+            self._item_property_cache[name] = get_item_property(self.item, name)
+        return self._item_property_cache[name]
+
+    def _apply_sentinel2_bandpass_adjustment(
+        self, uncorrected: ma.MaskedArray, assets: List[str], computing_dtype=np.float32
+    ) -> ma.MaskedArray:
+        out_arr: ma.MaskedArray = ma.masked_array(
+            data=np.zeros(uncorrected.shape, uncorrected.dtype),
+            mask=uncorrected.mask.copy(),
+            fill_value=uncorrected.fill_value,
+        )
+        for band_idx, asset in enumerate(assets):
+            out_arr[band_idx] = apply_bandpass_adjustment(
+                uncorrected[band_idx],
+                item=self.item,
+                l2a_band=asset_name_to_l2a_band(self.item, asset),
+                computing_dtype=computing_dtype,
+                out_dtype=uncorrected.dtype,
+            )
+        return out_arr
+
+    def _apply_brdf(
+        self,
+        uncorrected: ma.MaskedArray,
+        assets: List[str],
+        brdf_config: BRDFConfig,
+        grid: Union[GridProtocol, Resolution, None] = Resolution["10m"],
+        resampling: Resampling = Resampling.nearest,
+        mask_config: MaskConfig = MaskConfig(),
+    ) -> ma.MaskedArray:
+        out_arr: ma.MaskedArray = ma.masked_array(
+            data=np.zeros(uncorrected.shape, uncorrected.dtype),
+            mask=uncorrected.mask.copy(),
+            fill_value=uncorrected.fill_value,
+        )
+
+        # apply default correction defined in root
+        if brdf_config.model == BRDFModels.none:
+            logger.debug("no default BRDF model specified")
+            out_arr[:] = uncorrected
+        else:
+            logger.debug("applying %s to bands", brdf_config.model)
+            for band_idx, asset in enumerate(assets):
+                out_arr[band_idx] = apply_correction(
+                    band=uncorrected[band_idx],
+                    correction=self.read_brdf_grid(
+                        asset_name_to_l2a_band(self.item, asset),
+                        resampling=resampling,
+                        grid=grid,
+                        brdf_config=brdf_config,
+                    ),
+                    correction_weight=brdf_config.correction_weight,
+                    log10_bands_scale=brdf_config.log10_bands_scale,
+                )
+
+        # if SCL-specific correction is configured, apply and overwrite values in array
+        if brdf_config.scl_specific_configurations:
+            logger.debug("SCL class specific BRDF correction required")
+            scl_arr = self.read_scl(grid, mask_config.scl_cached_read).data
+
+            for scl_config in brdf_config.scl_specific_configurations:
+                scl_mask = np.isin(
+                    scl_arr, [scl_class.value for scl_class in scl_config.scl_classes]
+                )
+
+                for band_idx, asset in enumerate(assets):
+                    if scl_config.model == BRDFModels.none:
+                        # use uncorrected values from original array
+                        out_arr[band_idx][scl_mask] = uncorrected[band_idx][scl_mask]
+
+                    elif scl_mask.any():
+                        logger.debug(
+                            "applying BRDF model %s to SCL classes %s",
+                            scl_config.model.value,
+                            ", ".join(
+                                [scl_class.name for scl_class in scl_config.scl_classes]
+                            ),
+                        )
+                        # apply correction band by band
+                        out_arr[band_idx][scl_mask] = apply_correction(
+                            uncorrected[band_idx],
+                            self.read_brdf_grid(
+                                asset_name_to_l2a_band(self.item, asset),
+                                resampling=resampling,
+                                grid=grid,
+                                brdf_config=scl_config,
+                            ),
+                            correction_weight=scl_config.correction_weight,
+                            log10_bands_scale=scl_config.log10_bands_scale,
+                        )[scl_mask]
+
+                    # leave it be for all other cases
+
+        return out_arr
+
+
+def asset_name_to_l2a_band(item: Item, asset_name: str) -> L2ABand:
+    asset = item.assets[asset_name]
+    asset_path = MPath(asset.href)
+    band_name = asset_path.name.split(".")[0]
+    return L2ABand[band_name]