Merge pull request #50 from nasa/GITC-7208

GITC-7208: Fixing Colormap Rasterization issues

Author: jackiryan

CHANGELOG.md

@@ -4,6 +4,12 @@ HyBIG follows semantic versioning. All notable changes to this project will be
 documented in this file. The format is based on [Keep a
 Changelog](http://keepachangelog.com/en/1.0.0/).
 
+## [v2.4.0] - 2025-04-28
+
+### Changed
+
+* Fix rasterization issues with palettized granules. Source images now retain their palette in a scaled form, rather than reinterpreting the palette. [[#50](https://github.com/nasa/harmony-browse-image-generator/pull/50)]
+* Minor bugfixes and type formatting improvements.
 
 ## [v2.3.0] - 2025-02-26
 
@@ -112,6 +118,7 @@ For more information on internal releases prior to NASA open-source approval,
 see legacy-CHANGELOG.md.
 
 [unreleased]: https://github.com/nasa/harmony-browse-image-generator/
+[v2.4.0]: https://github.com/nasa/harmony-browse-image-generator/releases/tag/2.4.0
 [v2.3.0]: https://github.com/nasa/harmony-browse-image-generator/releases/tag/2.3.0
 [v2.2.0]: https://github.com/nasa/harmony-browse-image-generator/releases/tag/2.2.0
 [v2.1.0]: https://github.com/nasa/harmony-browse-image-generator/releases/tag/2.1.0

docker/service.Dockerfile

@@ -15,19 +15,30 @@
 # and updates the entrypoint of the new service container
 #
 ###############################################################################
-FROM python:3.12
+FROM python:3.12-slim
 
 WORKDIR "/home"
 
-RUN apt-get update
-RUN apt-get install -y libgdal-dev
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    libgdal-dev \
+    gdal-bin \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN export GDAL_VERSION=$(gdal-config --version) && \
+    echo "GDAL version: $GDAL_VERSION"
+
+ENV CPLUS_INCLUDE_PATH=/usr/include/gdal
+ENV C_INCLUDE_PATH=/usr/include/gdal
+
+RUN pip install GDAL==$(gdal-config --version)
 
 # Install Pip dependencies
-COPY pip_requirements*.txt /home/
+COPY pip_requirements.txt /home/
 
 RUN pip install --no-input --no-cache-dir \
-    -r pip_requirements.txt \
-    -r pip_requirements_skip_snyk.txt
+    -r pip_requirements.txt
 
 # Copy service code.
 COPY ./harmony_service harmony_service

docker/service_version.txt

@@ -1 +1 @@
-2.3.0
+2.4.0

hybig/browse.py

@@ -11,13 +11,11 @@
 from affine import dumpsw
 from harmony_service_lib.message import Message as HarmonyMessage
 from harmony_service_lib.message import Source as HarmonySource
-from matplotlib.cm import ScalarMappable
 from matplotlib.colors import Normalize
 from numpy import ndarray, uint8
 from osgeo_utils.auxiliary.color_palette import ColorPalette
 from PIL import Image
 from rasterio.io import DatasetReader
-from rasterio.plot import reshape_as_image, reshape_as_raster
 from rasterio.warp import Resampling, reproject
 from rioxarray import open_rasterio
 from xarray import DataArray
@@ -28,10 +26,12 @@
     NODATA_RGBA,
     OPAQUE,
     TRANSPARENT,
+    ColorMap,
     all_black_color_map,
+    colormap_from_colors,
     get_color_palette,
+    greyscale_colormap,
     palette_from_remote_colortable,
-    remove_alpha,
 )
 from hybig.exceptions import HyBIGError
 from hybig.sizes import (
@@ -171,18 +171,19 @@ def create_browse_imagery(
                 color_palette = get_color_palette(
                     in_dataset, source, item_color_palette
                 )
-                raster = convert_singleband_to_raster(rio_in_array, color_palette)
+                raster, color_map = convert_singleband_to_raster(
+                    rio_in_array, color_palette
+                )
             elif rio_in_array.rio.count in (3, 4):
                 raster = convert_mulitband_to_raster(rio_in_array)
+                color_map = None
+                if output_driver == 'JPEG':
+                    raster = raster[0:3, :, :]
             else:
                 raise HyBIGError(
                     f'incorrect number of bands for image: {rio_in_array.rio.count}'
                 )
 
-            raster, color_map = standardize_raster_for_writing(
-                raster, output_driver, rio_in_array.rio.count
-            )
-
             grid_parameters = get_target_grid_parameters(message, rio_in_array)
             grid_parameter_list, tile_locators = create_tiled_output_parameters(
                 grid_parameters
@@ -283,69 +284,65 @@ def original_dtype(data_array: DataArray) -> str | None:
 def convert_singleband_to_raster(
     data_array: DataArray,
     color_palette: ColorPalette | None = None,
-) -> ndarray:
-    """Convert input dataset to a 4 band raster image.
+) -> tuple[ndarray, ColorMap]:
+    """Convert input dataset to a 1-band palettized image with colormap.
 
-    Use a palette if provided otherwise return a greyscale image.
+    Uses a palette if provided otherwise returns a greyscale image.
     """
     if color_palette is None:
-        return convert_gray_1band_to_raster(data_array)
-    return convert_paletted_1band_to_raster(data_array, color_palette)
+        return scale_grey_1band(data_array)
+    return scale_paletted_1band(data_array, color_palette)
 
 
-def convert_gray_1band_to_raster(data_array: DataArray) -> ndarray:
-    """Convert a 1-band raster without a color association."""
+def scale_grey_1band(data_array: DataArray) -> tuple[ndarray, ColorMap]:
+    """Normalize input array and return scaled data with greyscale ColorMap."""
     band = data_array[0, :, :]
-    cmap = matplotlib.colormaps['Greys_r']
-    cmap.set_bad(NODATA_RGBA)
     norm = Normalize(vmin=np.nanmin(band), vmax=np.nanmax(band))
-    scalar_map = ScalarMappable(cmap=cmap, norm=norm)
 
-    rgba_image = np.zeros((*band.shape, 4), dtype='uint8')
-    for row_no in range(band.shape[0]):
-        rgba_image_slice = scalar_map.to_rgba(band[row_no, :], bytes=True)
-        rgba_image[row_no, :, :] = rgba_image_slice
+    # Scale input data from 0 to 254
+    normalized_data = norm(band) * 254.0
 
-    return reshape_as_raster(rgba_image)
+    # Set any missing to missing
+    normalized_data[np.isnan(normalized_data)] = NODATA_IDX
 
+    grey_colormap = greyscale_colormap()
+    raster_data = np.expand_dims(np.round(normalized_data).data, 0)
+    return np.array(raster_data, dtype='uint8'), grey_colormap
 
-def convert_paletted_1band_to_raster(
+
+def scale_paletted_1band(
     data_array: DataArray, palette: ColorPalette
-) -> ndarray:
-    """Convert a 1 band image with palette into a rgba raster image."""
+) -> tuple[ndarray, ColorMap]:
+    """Scale a 1-band image with palette into modified image and associated color_map.
+
+    Use the palette's levels and values, transform the input data_array into
+    the correct levels indexed from 0-255 return the scaled array along side of
+    a colormap corresponding to the new levels.
+    """
     band = data_array[0, :, :]
     levels = list(palette.pal.keys())
     colors = [
         palette.color_to_color_entry(value, with_alpha=True)
         for value in palette.pal.values()
     ]
-    scaled_colors = [
-        (r / 255.0, g / 255.0, b / 255.0, a / 255.0) for r, g, b, a in colors
-    ]
-
-    cmap, norm = matplotlib.colors.from_levels_and_colors(
-        levels, scaled_colors, extend='max'
-    )
+    norm = matplotlib.colors.BoundaryNorm(levels, len(levels) - 1)
 
     # handle palette no data value
+    nodata_color = (0, 0, 0, 0)
     if palette.ndv is not None:
-        nodata_colors = palette.color_to_color_entry(palette.ndv, with_alpha=True)
-        cmap.set_bad(
-            (
-                nodata_colors[0] / 255.0,
-                nodata_colors[1] / 255.0,
-                nodata_colors[2] / 255.0,
-                nodata_colors[3] / 255.0,
-            )
-        )
+        nodata_color = palette.color_to_color_entry(palette.ndv, with_alpha=True)
 
-    scalar_map = matplotlib.cm.ScalarMappable(norm=norm, cmap=cmap)
-    rgba_image = np.zeros((*band.shape, 4), dtype='uint8')
-    for row_no in range(band.shape[0]):
-        rgba_image[row_no, :, :] = scalar_map.to_rgba(
-            np.ma.masked_invalid(band[row_no, :]), bytes=True
-        )
-    return reshape_as_raster(rgba_image)
+    colors = [*colors, nodata_color]
+
+    scaled_band = norm(band)
+
+    # Set underflow and nan values to nodata
+    scaled_band[scaled_band == -1] = len(colors) - 1
+    scaled_band[np.isnan(band)] = len(colors) - 1
+
+    color_map = colormap_from_colors(colors)
+    raster_data = np.expand_dims(scaled_band.data, 0)
+    return np.array(raster_data, dtype='uint8'), color_map
 
 
 def image_driver(mime: str) -> str:
@@ -355,81 +352,6 @@ def image_driver(mime: str) -> str:
     return 'PNG'
 
 
-def standardize_raster_for_writing(
-    raster: ndarray,
-    driver: str,
-    band_count: int,
-) -> tuple[ndarray, dict | None]:
-    """Standardize raster data for writing to browse image.
-
-    Args:
-        raster: Input raster data array
-        driver: Output image format ('JPEG' or 'PNG')
-        band_count: Number of bands in original input data
-
-    The function handles two special cases:
-    - JPEG output with 4-band data -> Drop alpha channel and return 3-band RGB
-    - PNG output with single-band data -> Convert to paletted format
-
-    Returns:
-        tuple: (prepared_raster, color_map) where:
-            - prepared_raster is the processed ndarray
-            - color_map is either None or a dict mapping palette indices to RGBA values
-
-
-    """
-    if driver == 'JPEG' and raster.shape[0] == 4:
-        return raster[0:3, :, :], None
-
-    if driver == 'PNG' and band_count == 1:
-        # Only palettize single band input data that has been converted to an
-        # RGBA raster.
-        return palettize_raster(raster)
-
-    return raster, None
-
-
-def palettize_raster(raster: ndarray) -> tuple[ndarray, dict]:
-    """Convert an RGB or RGBA image into a 1band image and palette.
-
-    Converts a 3 or 4 band np raster into a PIL image.
-    Quantizes the image into a 1band raster with palette
-
-    Transparency is handled by first removing the Alpha layer and creating
-    quantized raster from just the RGB layers. Next the Alpha layer values are
-    treated as either transparent or opaque and any transparent values are
-    written to the final raster as 254 and add the mapped RGBA value to the
-    color palette.
-    """
-    # reserves index 255 for transparent and off grid fill values
-    # 0 to 254
-    max_colors = 255
-    rgb_raster, alpha = remove_alpha(raster)
-
-    multiband_image = Image.fromarray(reshape_as_image(rgb_raster))
-    quantized_image = multiband_image.quantize(colors=max_colors)
-
-    color_map = get_color_map_from_image(quantized_image)
-
-    quantized_array, color_map = add_alpha(alpha, np.array(quantized_image), color_map)
-
-    one_band_raster = np.expand_dims(quantized_array, 0)
-    return one_band_raster, color_map
-
-
-def add_alpha(
-    alpha: ndarray | None, quantized_array: ndarray, color_map: dict
-) -> tuple[ndarray, dict]:
-    """If the input data had alpha values, manually set the quantized_image
-    index to the transparent index in those places.
-    """
-    if alpha is not None and np.any(alpha != OPAQUE):
-        # Set any alpha to the transparent index value
-        quantized_array = np.where(alpha != OPAQUE, NODATA_IDX, quantized_array)
-        color_map[NODATA_IDX] = NODATA_RGBA
-    return quantized_array, color_map
-
-
 def get_color_map_from_image(image: Image) -> dict:
     """Get a writable color map
 

hybig/color_utility.py

@@ -8,6 +8,7 @@
 import numpy as np
 import requests
 from harmony_service_lib.message import Source as HarmonySource
+from numpy import uint8
 from osgeo_utils.auxiliary.color_palette import ColorPalette
 from pystac import Item
 from rasterio.io import DatasetReader
@@ -17,16 +18,18 @@
     HyBIGNoColorInformation,
 )
 
+ColorMap = dict[uint8, tuple[uint8, uint8, uint8, uint8]]
+
 # Constants for output PNG images
 # Applied to transparent pixels where alpha < 255
-TRANSPARENT = np.uint8(0)
-OPAQUE = np.uint8(255)
+TRANSPARENT = uint8(0)
+OPAQUE = uint8(255)
 # Applied to off grid areas during reprojection
 NODATA_RGBA = (0, 0, 0, 0)
 NODATA_IDX = 255
 
 
-def remove_alpha(raster: np.ndarray) -> tuple[np.ndarray, np.ndarray, None]:
+def remove_alpha(raster: np.ndarray) -> tuple[np.ndarray, np.ndarray | None]:
     """Remove alpha layer when it exists."""
     if raster.shape[0] == 4:
         return raster[0:3, :, :], raster[3, :, :]
@@ -87,7 +90,16 @@ def get_color_palette(
         return get_remote_palette_from_source(source)
     except HyBIGNoColorInformation:
         try:
-            return convert_colormap_to_palette(dataset.colormap(1))
+            ds_cmap = dataset.colormap(1)
+            # very defensive since this function is not documented in rasterio
+            ndv_tuple: tuple[float, ...] = dataset.get_nodatavals()
+            if ndv_tuple is not None and len(ndv_tuple) > 0:
+                # this service only supports one ndv, so just use the first one
+                # (usually the only one)
+                ds_cmap['nv'] = ds_cmap[ndv_tuple[0]]
+                # then remove the value associated with the ndv key
+                ds_cmap.pop(ndv_tuple[0])
+            return convert_colormap_to_palette(ds_cmap)
         except ValueError:
             return None
 
@@ -120,11 +132,28 @@ def get_remote_palette_from_source(source: HarmonySource) -> dict:
         raise HyBIGNoColorInformation('No color in source') from exc
 
 
-def all_black_color_map():
+def all_black_color_map() -> ColorMap:
     """Return a full length rgba color map with all black values."""
     return {idx: (0, 0, 0, 255) for idx in range(256)}
 
 
+def colormap_from_colors(
+    colors: list[tuple[uint8, uint8, uint8, uint8]],
+) -> ColorMap:
+    color_map = {}
+    for idx, rgba in enumerate(colors):
+        color_map[idx] = rgba
+    return color_map
+
+
+def greyscale_colormap() -> ColorMap:
+    color_map = {}
+    for idx in range(255):
+        color_map[idx] = (idx, idx, idx, 255)
+    color_map[NODATA_IDX] = NODATA_RGBA
+    return color_map
+
+
 def convert_colormap_to_palette(colormap: dict) -> ColorPalette:
     """Convert a GeoTIFF palette to GDAL ColorPalette.