How to plot normalized rgb by time?

[zxdawn]

Is it possible to normalize the rgb array by time dimension? I see if there're some clouds, the nomalization will cause some dark imgs like this:

ds_rgb.isel(time=slice(0, 5)).to_array(name="rgb").wx.normalize(stretch=0.99).plot.imshow(col='time')

[zxdawn]

Well, it's easier than I thought:

def normalize(da, stretch=1):
    min_val = np.quantile(da, 1-stretch)
    max_val = np.quantile(da, stretch)
    da_norm = np.clip((da - min_val) / (max_val - min_val), 0, 1)

    return da_norm


rgb = xr.apply_ufunc(normalize,
                     ds_rgb.isel(time=slice(0, 5)).to_array(name="rgb"),
                     exclude_dims=set(('variable', 'y', 'x')),
                     input_core_dims=[['variable', 'y', 'x']],
                     output_core_dims=[['variable', 'y', 'x']],
                     vectorize=True,
                     )

I can create a PR for this.

[aazuspan]

Glad you got it figured out so quickly!

There's actually an even easier solution. If you calculate the quantiles only over the x, y, and variable dimensions (not time), you get one value per time coordinate and xarray will automatically broadcast everything correctly:

def normalize(da, stretch = 1.0):
    min_val = da.quantile(1 - stretch, dim=["x", "y", "variable"])
    max_val = da.quantile(stretch, dim=["x", "y", "variable"])

    return ((da - min_val) / (max_val - min_val)).clip(0, 1)

Do you think this should be the default behavior for the rgb method? That seems like a good idea to me. If you want to make a PR, that'd be great!

EDIT: I originally suggested calculating quantiles over just x and y, but including variable like you did with the ufunc definitely looks better since it doesn't change relationships between bands.

[zxdawn]

Oh, yes! Thanks for your tip which is cleaner and better than mine ;) I will create a PR for RGB.