Refactor python API

[perrygeo]

We should build a pure function (data-in, data-out, no IO) that we can unit test easily. This might take some refactoring to accomplish.

I'm thinking of a function with a signature like:

def pansharpen(vis, vis_transform, pan, pan_transform, method="Brovey"):
    """Pansharpen a lower-resolution visual band

    Parameters
    =========
    vis: ndarray, 3D with shape == (3, vh, vw)
        Visual band array with RGB bands
    vis_transform: Affine
        affine transform defining the georeferencing of the vis array
    pan: ndarray, 2D with shape == (ph, pw)
        Panchromatic band array
    pan_transform: Affine
        affine transform defining the georeferencing of the pan array
    method: string
        Algorithm for pansharpening; default Brovey

    Returns:
    ======
    pansharp: ndarray, 3D with shape == (3, ph, pw)
        pansharpened visual band
        affine transform is identical to `pan_transform`
    """

[virginiayung]

These are the current steps I am taking to refactor the current worker script in the refactor_api branch and PR #13

• Renaming current pansharpen function to calculate_landsat_pansharpen. This function handles:
• reading, updating input meta data
• writing the result of pansharpening (derived by _pansharpen_worker) into each window respectively via riomucho.
• _pansharpen_worker function is the riomucho worker for pansharpening. It is responsible for:
  • Reading panchromatic band data, determines the datatype for the panchromatic band from meta data
  • Determining RGB window sizes (includes padding)
  • Determining affines for pan, RGB windows
  • Reading RGB data
  • Passing necessary arguments to the new pansharpen function to perform pansharpening
  • rescaling pansharpened array to destination datatype
• new pansharpen is the pansharpening python API (data-in, data-out, no IO) that performs pansharpening operations:
  • Takes ndarrays (RGB bands, pan band), and arguments (Affine transforms) etc
  • Creates a mask of pixels where any channel is 0 (nodata)
  • Apply the mask to each of the RGB bands
  • Perform pansharpening using the specified algorithm. Default is Brovey.
  • Returns ndarray, 3D with shape == (3, panband_height, panband_width)

The biggest advantage for refactoring the current pansharpening script into a pure function, separating IO from array operations is so that we can easily write unit tests that specifically test for array operations:

def test_create_apply_mask(arr):
  .
  .
def test_upsample(masked_arr, pan.shape, pan.shape, rgb_affine, r_crs, pan_affine, dst_crs)
  .
  .
def test_Brovey(arr1, arr2, weight, dtype):
  .
  .
def test_rescale(arr):
  .
  .

cc @dnomadb @perrygeo @celoyd @mapbox/satellite

[virginiayung]

🚢 in PR #13