Creating an in memory rasterio Dataset from numpy array

You need to recreate a DatasetReader manually and you can use a MemoryFile to avoid writing to disk.

You can re-use the metadata from the input raster in the DatasetReader, but you'll need to modify the height and width properties and the transform. From documentation:

After these resolution changing operations, the dataset’s resolution and the resolution components of its affine transform property no longer apply to the new arrays.

In the example below note that:

1. I use a contextmanager so the DatasetReader and MemoryFile objects get cleaned up automatically. This is why I use yield not return in the function
2. I had to change the order of indexes in raster.read as arrays are (band, row, col) order not (row, col, band) like you used in your snippet.

# Example licensed under cc by-sa 3.0 with attribution required

from contextlib import contextmanager  

import rasterio
from rasterio import Affine, MemoryFile
from rasterio.enums import Resampling

# use context manager so DatasetReader and MemoryFile get cleaned up automatically
@contextmanager
def resample_raster(raster, scale=2):
    t = raster.transform

    # rescale the metadata
    transform = Affine(t.a / scale, t.b, t.c, t.d, t.e / scale, t.f)
    height = raster.height * scale
    width = raster.width * scale

    profile = src.profile
    profile.update(transform=transform, driver='GTiff', height=height, width=width)

    data = raster.read( # Note changed order of indexes, arrays are band, row, col order not row, col, band
            out_shape=(raster.count, height, width),
            resampling=Resampling.bilinear,
        )

    with MemoryFile() as memfile:
        with memfile.open(**profile) as dataset: # Open as DatasetWriter
            dataset.write(data)
            del data

        with memfile.open() as dataset:  # Reopen as DatasetReader
            yield dataset  # Note yield not return     


with rasterio.open('path/to/raster') as src:
    with resample_raster(src) as resampled:
        print('Orig dims: {}, New dims: {}'.format(src.shape, resampled.shape))
        print(repr(resampled))

Orig dims: (4103, 4682), New dims: (8206, 9364)
<open DatasetReader name='/vsimem/95befda0-2061-4294-982b-20e46f127066.' mode='r'>

I wrote a wrapper for rasterio.mask.mask that accepts numpy arrays as inputs.

def mask_raster_with_geometry(raster, transform, shapes, **kwargs):
    """Wrapper for rasterio.mask.mask to allow for in-memory processing.

    Docs: https://rasterio.readthedocs.io/en/latest/api/rasterio.mask.html

    Args:
        raster (numpy.ndarray): raster to be masked with dim: [H, W]
        transform (affine.Affine): the transform of the raster
        shapes, **kwargs: passed to rasterio.mask.mask

    Returns:
        masked: numpy.ndarray or numpy.ma.MaskedArray with dim: [H, W]
    """
    with rasterio.io.MemoryFile() as memfile:
        with memfile.open(
            driver='GTiff',
            height=raster.shape[0],
            width=raster.shape[1],
            count=1,
            dtype=raster.dtype,
            transform=transform,
        ) as dataset:
            dataset.write(raster, 1)
        with memfile.open() as dataset:
            output, _ = rasterio.mask.mask(dataset, shapes, **kwargs)
    return output.squeeze(0)

This is done via first writing to memfile and then reading from it before the function call to rasterio.mask.mask(). The MemoryFile will be discarded after exiting the with statement.

Unit test:

@pytest.mark.parametrize(
    'raster,transform,shapes,expected',
    [
        pytest.param(
            # the raster is positioned between x in (5, 8) and y in (-14, -10)
            np.ones((4, 3), dtype=np.float32),
            rasterio.transform.Affine(1, 0, 5, 0, -1, -10),
            # mask out two pixels
            # box params: xmin, ymin, xmax, ymax
            [shapely.geometry.box(6.1, -12.9, 6.9, -11.1)],
            np.array([
                [0, 0, 0],
                [0, 1, 0],
                [0, 1, 0],
                [0, 0, 0],
            ], dtype=np.float32),
        ),
    ],
)
def test_mask_raster_with_geometry(raster, transform, shapes, expected):
    np.testing.assert_array_equal(
        mask_raster_with_geometry(raster, transform, shapes),
        expected)