DAS-2276: retain 3 and 4 band information in browse images

CHANGELOG.md

@@ -4,10 +4,11 @@ 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/).
 
-## [unreleased] - 2024-12-10
+## [v2.1.0] - 2024-12-13
 
 ### Changed
 
+* Input GeoTIFF RGB[A] images are **no longer palettized** when converted to a PNG. The new resulting output browse images are now 3 or 4 band PNG retaining the color information of the input image.[#39](https://github.com/nasa/harmony-browse-image-generator/pull/39)
 * Changed pre-commit configuration to remove `black-jupyter` dependency [#38](https://github.com/nasa/harmony-browse-image-generator/pull/38)
 * Updates service image's python to 3.12 [#38](https://github.com/nasa/harmony-browse-image-generator/pull/38)
 * Simplifies test scripts to run with pytest and pytest plugins [#38](https://github.com/nasa/harmony-browse-image-generator/pull/38)
@@ -90,7 +91,8 @@ outlined by the NASA open-source guidelines.
 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/compare/2.0.2..HEAD
+[unreleased]:https://github.com/nasa/harmony-browse-image-generator/compare/2.1.0..HEAD
+[v2.1.0]:https://github.com/nasa/harmony-browse-image-generator/compare/2.0.2..2.1.0
 [v2.0.2]:https://github.com/nasa/harmony-browse-image-generator/compare/2.0.1..2.0.2
 [v2.0.1]:https://github.com/nasa/harmony-browse-image-generator/compare/2.0.0..2.0.1
 [v2.0.0]:https://github.com/nasa/harmony-browse-image-generator/compare/1.2.2..2.0.0

docker/service_version.txt

@@ -1 +1 @@
-2.0.2
+2.1.0

hybig/browse.py

@@ -13,7 +13,7 @@
 from harmony_service_lib.message import Source as HarmonySource
 from matplotlib.cm import ScalarMappable
 from matplotlib.colors import Normalize
-from numpy import ndarray
+from numpy import ndarray, uint8
 from osgeo_utils.auxiliary.color_palette import ColorPalette
 from PIL import Image
 from rasterio.io import DatasetReader
@@ -181,7 +181,9 @@ def create_browse_imagery(
                     f'incorrect number of bands for image: {rio_in_array.rio.count}'
                 )
 
-            raster, color_map = prepare_raster_for_writing(raster, output_driver)
+            raster, color_map = prepare_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(
@@ -217,12 +219,14 @@ def create_browse_imagery(
     return processed_files
 
 
-def convert_mulitband_to_raster(data_array: DataArray) -> ndarray:
+def convert_mulitband_to_raster(data_array: DataArray) -> ndarray[uint8]:
     """Convert multiband to a raster image.
 
-    Reads the three or four bands from the file, then normalizes them to the range
-    0 to 255. This assumes the input image is already in RGB or RGBA format and
-    just ensures that the output is 8bit.
+    Return a 4-band raster, where the alpha layer is presumed to be the missing
+    data mask.
+
+    Convert 3-band data into a 4-band raster by generating an alpha layer from
+    any missing data in the RGB bands.
 
     """
     if data_array.rio.count not in [3, 4]:
@@ -233,26 +237,49 @@ def convert_mulitband_to_raster(data_array: DataArray) -> ndarray:
 
     bands = data_array.to_numpy()
 
-    # Create an alpha layer where input NaN values are transparent.
+    if data_array.rio.count == 4:
+        return convert_to_uint8(bands, original_dtype(data_array))
+
+    # Input NaNs in any of the RGB bands are made transparent.
     nan_mask = np.isnan(bands).any(axis=0)
     nan_alpha = np.where(nan_mask, TRANSPARENT, OPAQUE)
 
-    # grab any existing alpha layer
-    bands, image_alpha = remove_alpha(bands)
+    raster = convert_to_uint8(bands, original_dtype(data_array))
 
-    norm = Normalize(vmin=np.nanmin(bands), vmax=np.nanmax(bands))
-    raster = np.nan_to_num(np.around(norm(bands) * 255.0), copy=False, nan=0.0).astype(
-        'uint8'
-    )
+    return np.concatenate((raster, nan_alpha[None, ...]), axis=0)
+
+
+def convert_to_uint8(bands: ndarray, dtype: str | None) -> ndarray[uint8]:
+    """Convert Banded data with NaNs (missing) into a uint8 data cube.
+
+    99.99% of the time this will simply pass through the data coercing it back
+    into unsigned ints and setting the missing values to 0 that will be masked
+    as transparent in the output png.
+
+    There is a some small non-zero chance that the input RGB image was 16-bit
+    and if any of the values exceed 255, we must normalize all of input data to
+    the range 0-255.
+
+    """
 
-    if image_alpha is not None:
-        # merge missing alpha with the image alpha band prefering transparency
-        # to opaqueness.
-        alpha = np.minimum(nan_alpha, image_alpha).astype(np.uint8)
+    if dtype != 'uint8' and np.nanmax(bands) > 255:
+        norm = Normalize(vmin=np.nanmin(bands), vmax=np.nanmax(bands))
+        scaled = np.around(norm(bands) * 255.0)
+        raster = scaled.filled(0).astype('uint8')
     else:
-        alpha = nan_alpha
+        raster = np.nan_to_num(bands).astype('uint8')
 
-    return np.concatenate((raster, alpha[None, ...]), axis=0)
+    return raster
+
+
+def original_dtype(data_array: DataArray) -> str | None:
+    """Return the original input data's type.
+
+    rastero_optn retains the input dtype in the encoding dictionary and is used
+    to understand what kind of casts are safe.
+
+    """
+    return data_array.encoding.get('dtype') or data_array.encoding.get('rasterio_dtype')
 
 
 def convert_singleband_to_raster(
@@ -331,15 +358,19 @@ def image_driver(mime: str) -> str:
 
 
 def prepare_raster_for_writing(
-    raster: ndarray, driver: str
+    raster: ndarray,
+    driver: str,
+    input_bands: int,
 ) -> tuple[ndarray, dict | None]:
     """Remove alpha layer if writing a jpeg."""
-    if driver == 'JPEG':
-        if raster.shape[0] == 4:
-            raster = raster[0:3, :, :]
-        return raster, None
+    if driver == 'JPEG' and raster.shape[0] == 4:
+        return raster[0:3, :, :], None
+
+    if driver == 'PNG' and input_bands == 1:
+        # we only paletize single band input data
+        return palettize_raster(raster)
 
-    return palettize_raster(raster)
+    return raster, None
 
 
 def palettize_raster(raster: ndarray) -> tuple[ndarray, dict]:
@@ -476,9 +507,13 @@ def write_georaster_as_browse(
 
     """
     n_bands = raster.shape[0]
-    dst_nodata = NODATA_IDX
+
     if color_map is not None:
+        dst_nodata = NODATA_IDX
         color_map[dst_nodata] = NODATA_RGBA
+    else:
+        # for banded data set the each band's destination nodata to zero (TRANSPARENT).
+        dst_nodata = TRANSPARENT
 
     creation_options = {
         **grid_parameters,

hybig/sizes.py

@@ -422,7 +422,7 @@ def find_closest_resolution(
 
     """
     best_info = None
-    smallest_diff = np.Infinity
+    smallest_diff = np.inf
     for res in resolutions:
         for info in resolution_info:
             resolution_diff = np.abs(res - info.pixel_size)

tests/test_service/test_adapter.py

@@ -270,10 +270,14 @@ def move_tif(*args, **kwargs):
             mock_reproject.call_args_list, expected_reproject_calls
         ):
             np.testing.assert_array_equal(
-                actual_call.kwargs['source'], expected_call.kwargs['source']
+                actual_call.kwargs['source'],
+                expected_call.kwargs['source'],
+                strict=True,
             )
             np.testing.assert_array_equal(
-                actual_call.kwargs['destination'], expected_call.kwargs['destination']
+                actual_call.kwargs['destination'],
+                expected_call.kwargs['destination'],
+                strict=True,
             )
             self.assertEqual(
                 actual_call.kwargs['src_transform'],
@@ -452,11 +456,11 @@ def move_tif(*args, **kwargs):
             'transform': expected_transform,
             'driver': 'PNG',
             'dtype': 'uint8',
-            'dst_nodata': 255,
+            'dst_nodata': 0,
             'count': 3,
         }
         raster = convert_mulitband_to_raster(rio_data_array)
-        raster, color_map = prepare_raster_for_writing(raster, 'PNG')
+        raster, color_map = prepare_raster_for_writing(raster, 'PNG', 3)
 
         dest = np.full(
             (expected_params['height'], expected_params['width']),
@@ -466,26 +470,31 @@ def move_tif(*args, **kwargs):
 
         expected_reproject_calls = [
             call(
-                source=raster[0, :, :],
+                source=raster[band, :, :],
                 destination=dest,
                 src_transform=rio_data_array.rio.transform(),
                 src_crs=rio_data_array.rio.crs,
                 dst_transform=expected_params['transform'],
                 dst_crs=expected_params['crs'],
-                dst_nodata=255,
+                dst_nodata=expected_params['dst_nodata'],
                 resampling=Resampling.nearest,
             )
+            for band in range(4)
         ]
 
-        self.assertEqual(mock_reproject.call_count, 1)
+        self.assertEqual(mock_reproject.call_count, 4)
         for actual_call, expected_call in zip(
             mock_reproject.call_args_list, expected_reproject_calls
         ):
             np.testing.assert_array_equal(
-                actual_call.kwargs['source'], expected_call.kwargs['source']
+                actual_call.kwargs['source'],
+                expected_call.kwargs['source'],
+                strict=True,
             )
             np.testing.assert_array_equal(
-                actual_call.kwargs['destination'], expected_call.kwargs['destination']
+                actual_call.kwargs['destination'],
+                expected_call.kwargs['destination'],
+                strict=True,
             )
             self.assertEqual(
                 actual_call.kwargs['src_transform'],