fix: add more robust version of ndvi (#147)

* add more robust version of ndvi

* extend test coverage

* make test data smaller

* Update tests/test_indices.py

Co-authored-by: ValentinaHutter <85164505+ValentinaHutter@users.noreply.github.com>

* update tests

* address feedback from review

---------

Co-authored-by: ValentinaHutter <85164505+ValentinaHutter@users.noreply.github.com>

openeo_processes_dask/process_implementations/cubes/__init__.py

@@ -2,6 +2,7 @@
 from .aggregate import *
 from .apply import *
 from .general import *
+from .indices import *
 from .load import *
 from .merge import *
 from .reduce import *

openeo_processes_dask/process_implementations/cubes/indices.py

@@ -0,0 +1,59 @@
+import xarray as xr
+
+from openeo_processes_dask.process_implementations.data_model import RasterCube
+from openeo_processes_dask.process_implementations.exceptions import (
+    BandExists,
+    DimensionAmbiguous,
+    NirBandAmbiguous,
+    RedBandAmbiguous,
+)
+from openeo_processes_dask.process_implementations.math import normalized_difference
+
+__all__ = ["ndvi"]
+
+
+def ndvi(data: RasterCube, nir="nir", red="red", target_band=None):
+    if len(data.openeo.band_dims) == 0:
+        raise DimensionAmbiguous(
+            "Dimension of type `bands` is not available or is ambiguous."
+        )
+    band_dim = data.openeo.band_dims[0]
+    available_bands = data.coords[band_dim]
+
+    if nir not in available_bands or red not in available_bands:
+        try:
+            data = data.set_xindex("common_name")
+        except (ValueError, KeyError):
+            pass
+
+        if (
+            nir not in available_bands
+            and "common_name" in data.xindexes._coord_name_id.keys()
+            and nir not in data.coords["common_name"].data
+        ):
+            raise NirBandAmbiguous(
+                "The NIR band can't be resolved, please specify the specific NIR band name."
+            )
+        elif (
+            red not in available_bands
+            and "common_name" in data.xindexes._coord_name_id.keys()
+            and red not in data.coords["common_name"].data
+        ):
+            raise RedBandAmbiguous(
+                "The Red band can't be resolved, please specify the specific Red band name."
+            )
+
+    nir_band_dim = "common_name" if nir not in available_bands else band_dim
+    red_band_dim = "common_name" if red not in available_bands else band_dim
+
+    nir_band = data.sel({nir_band_dim: nir})
+    red_band = data.sel({red_band_dim: red})
+
+    nd = normalized_difference(nir_band, red_band)
+    if target_band is not None:
+        if target_band in data.coords:
+            raise BandExists("A band with the specified target name exists.")
+        nd = nd.expand_dims(band_dim).assign_coords({band_dim: [target_band]})
+        nd = xr.merge([data, nd])
+    nd.attrs = data.attrs
+    return nd

openeo_processes_dask/process_implementations/exceptions.py

@@ -64,3 +64,19 @@ class NoDataAvailable(OpenEOException):
 
 class TemporalExtentEmpty(OpenEOException):
     pass
+
+
+class DimensionAmbiguous(OpenEOException):
+    pass
+
+
+class NirBandAmbiguous(OpenEOException):
+    pass
+
+
+class RedBandAmbiguous(OpenEOException):
+    pass
+
+
+class BandExists(OpenEOException):
+    pass

openeo_processes_dask/process_implementations/math.py

@@ -59,7 +59,6 @@
     "quantiles",
     "product",
     "normalized_difference",
-    "ndvi",
 ]
 
 
@@ -326,41 +325,3 @@ def product(data, ignore_nodata=True, axis=None, keepdims=False):
 def normalized_difference(x, y):
     nd = (x - y) / (x + y)
     return nd
-
-
-def ndvi(data, nir="nir", red="red", target_band=None):
-    r = np.nan
-    n = np.nan
-    if "bands" in data.dims:
-        if red == "red":
-            if "B04" in data["bands"].values:
-                r = data.sel(bands="B04")
-        elif red == "rededge":
-            if "B05" in data["bands"].values:
-                r = data.sel(bands="B05")
-            elif "B06" in data["bands"].values:
-                r = data.sel(bands="B06")
-            elif "B07" in data["bands"].values:
-                r = data.sel(bands="B07")
-        if nir == "nir":
-            n = data.sel(bands="B08")
-        elif nir == "nir08":
-            if "B8a" in data["bands"].values:
-                n = data.sel(bands="B8a")
-            elif "B8A" in data["bands"].values:
-                n = data.sel(bands="B8A")
-            elif "B05" in data["bands"].values:
-                n = data.sel(bands="B05")
-        elif nir == "nir09":
-            if "B09" in data["bands"].values:
-                n = data.sel(bands="B09")
-        if red in data["bands"].values:
-            r = data.sel(bands=red)
-        if nir in data["bands"].values:
-            n = data.sel(bands=nir)
-    nd = normalized_difference(n, r)
-    if target_band is not None:
-        nd = nd.assign_coords(bands=target_band)
-    # TODO: Remove this once we have the .openeo accessor
-    nd.attrs = data.attrs
-    return nd

tests/conftest.py

@@ -28,14 +28,18 @@ def dask_client():
     client.shutdown()
 
 
-@pytest.fixture
-def random_raster_data(size, dtype, seed=42):
+def _random_raster_data(size, dtype, seed=42):
     rng = np.random.default_rng(seed)
     data = rng.integers(-100, 100, size=size)
     data = data.astype(dtype)
     return data
 
 
+@pytest.fixture
+def random_raster_data(size, dtype, seed=42):
+    return _random_raster_data(size, dtype, seed=seed)
+
+
 @pytest.fixture
 def bounding_box(
     west=10.45, east=10.5, south=46.1, north=46.2, crs="EPSG:4326"

tests/test_indices.py

@@ -0,0 +1,85 @@
+import numpy as np
+import pytest
+
+from openeo_processes_dask.process_implementations.cubes.indices import ndvi
+from openeo_processes_dask.process_implementations.cubes.load import load_stac
+from openeo_processes_dask.process_implementations.exceptions import (
+    BandExists,
+    DimensionAmbiguous,
+    NirBandAmbiguous,
+    RedBandAmbiguous,
+)
+from tests.conftest import _random_raster_data
+from tests.general_checks import general_output_checks
+
+
+def test_ndvi(bounding_box):
+    url = "./tests/data/stac/s2_l2a_test_item.json"
+    input_cube = load_stac(
+        url=url,
+        spatial_extent=bounding_box,
+        bands=["red", "nir"],
+    ).isel({"x": slice(0, 20), "y": slice(0, 20)})
+
+    # Test whether this works with different band names
+    input_cube = input_cube.rename({"band": "b"})
+
+    import dask.array as da
+
+    numpy_data = _random_raster_data(input_cube.data.shape, dtype=np.float64)
+
+    input_cube.data = da.from_array(numpy_data, chunks=("auto", "auto", "auto", -1))
+
+    output = ndvi(input_cube)
+
+    band_dim = input_cube.openeo.band_dims[0]
+    assert band_dim not in output.dims
+
+    expected_results = (
+        input_cube.sel({band_dim: "nir"}) - input_cube.sel({band_dim: "red"})
+    ) / (input_cube.sel({band_dim: "nir"}) + input_cube.sel({band_dim: "red"}))
+
+    general_output_checks(
+        input_cube=input_cube, output_cube=output, expected_results=expected_results
+    )
+
+    cube_with_resolvable_coords = input_cube.assign_coords(
+        {band_dim: ["blue", "yellow"]}
+    )
+    output = ndvi(cube_with_resolvable_coords)
+    general_output_checks(
+        input_cube=cube_with_resolvable_coords,
+        output_cube=output,
+        expected_results=expected_results,
+    )
+
+    with pytest.raises(DimensionAmbiguous):
+        ndvi(output)
+
+    cube_with_nir_unresolvable = cube_with_resolvable_coords
+    cube_with_nir_unresolvable.common_name.data = np.array(["blue", "red"])
+
+    with pytest.raises(NirBandAmbiguous):
+        ndvi(cube_with_nir_unresolvable)
+
+    cube_with_red_unresolvable = cube_with_resolvable_coords
+    cube_with_red_unresolvable.common_name.data = np.array(["nir", "yellow"])
+
+    with pytest.raises(RedBandAmbiguous):
+        ndvi(cube_with_red_unresolvable)
+
+    cube_with_nothing_resolvable = cube_with_resolvable_coords
+    cube_with_nothing_resolvable = cube_with_nothing_resolvable.drop_vars("common_name")
+    with pytest.raises(KeyError):
+        ndvi(cube_with_nothing_resolvable)
+
+    target_band = "yay"
+    output_with_extra_dim = ndvi(input_cube, target_band=target_band)
+    assert len(output_with_extra_dim.dims) == len(output.dims) + 1
+    assert (
+        len(output_with_extra_dim.coords[band_dim])
+        == len(input_cube.coords[band_dim]) + 1
+    )
+
+    with pytest.raises(BandExists):
+        output_with_extra_dim = ndvi(input_cube, target_band="time")