def _convolve_2d_cupy(data, kernel):
data = data.astype(cupy.float32)
out = cupy.empty(data.shape, dtype='f4')
out[:, :] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
_convolve_2d_cuda[griddim, blockdim](data, kernel, cupy.asarray(out))
return out
def _evi_cupy(nir_data, red_data, blue_data, c1, c2, soil_factor, gain):
griddim, blockdim = cuda_args(nir_data.shape)
out = cupy.empty(nir_data.shape, dtype='f4')
out[:] = cupy.nan
args = (nir_data, red_data, blue_data, c1, c2, soil_factor, gain, out)
_evi_gpu[griddim, blockdim](*args)
return out
def _run_cupy(data, azimuth, angle_altitude):
x, y = np.gradient(data.get())
x = cupy.asarray(x, dtype=x.dtype)
y = cupy.asarray(y, dtype=y.dtype)
altituderad = angle_altitude * np.pi / 180.
sin_altituderad = np.sin(altituderad)
cos_altituderad = np.cos(altituderad)
griddim, blockdim = cuda_args(data.shape)
arctan_part = cupy.empty(data.shape, dtype='f4')
_gpu_calc[griddim, blockdim](x, y, arctan_part)
slope = np.pi / 2. - np.arctan(arctan_part)
sin_slope = np.sin(slope)
sin_part = sin_altituderad * sin_slope
azimuthrad = (360.0 - azimuth) * np.pi / 180.
aspect = (azimuthrad - np.pi / 2.) - np.arctan2(-x, y)
cos_aspect = np.cos(aspect)
cos_slope = np.cos(slope)
cos_part = cupy.empty(data.shape, dtype='f4')
_gpu_cos_part[griddim, blockdim](cos_altituderad, cos_slope,
cos_aspect, cos_part)
shaded = sin_part + cos_part
out = (shaded + 1) / 2
out[0, :] = cupy.nan
out[-1, :] = cupy.nan
out[:, 0] = cupy.nan
out[:, -1] = cupy.nan
return out
def _terrain_gpu(height_map, seed, x_range=(0, 1), y_range=(0, 1)):
NOISE_LAYERS = ((1 / 2**i, (2**i, 2**i)) for i in range(16))
noise = cupy.empty_like(height_map, dtype=np.float32)
griddim, blockdim = cuda_args(height_map.shape)
for i, (m, (xfreq, yfreq)) in enumerate(NOISE_LAYERS):
# cupy.random.seed(seed+i)
# p = cupy.random.permutation(2**20)
# use numpy.random then transfer data to GPU to ensure the same result
# when running numpy backed and cupy backed data array.
np.random.seed(seed + i)
p = cupy.asarray(np.random.permutation(2**20))
p = cupy.append(p, p)
_perlin_gpu[griddim,
blockdim](p, x_range[0] * xfreq, x_range[1] * xfreq,
y_range[0] * yfreq, y_range[1] * yfreq, m, noise)
height_map += noise
height_map /= (1.00 + 0.50 + 0.25 + 0.13 + 0.06 + 0.03)
height_map = height_map**3
return height_map
def _run_cupy_binary(data, values):
values_cupy = cupy.asarray(values)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
_run_gpu_binary[griddim, blockdim](data, values_cupy, out)
return out
def convolve_2d(image, kernel, pad=True, use_cuda=True):
"""Function to call the 2D convolution via Numba.
The Numba convolution function does not account for an edge so
if we wish to take this into account, will pad the image array.
"""
# Don't allow padding on (1, 1) kernel
if (kernel.shape[0] == 1 and kernel.shape[1] == 1):
pad = False
if pad:
pad_rows = kernel.shape[0] // 2
pad_cols = kernel.shape[1] // 2
pad_width = ((pad_rows, pad_rows), (pad_cols, pad_cols))
else:
# If padding is not desired, set pads to 0
pad_rows = 0
pad_cols = 0
pad_width = 0
padded_image = np.pad(image, pad_width=pad_width, mode="reflect")
result = np.empty_like(padded_image)
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(padded_image.shape)
_convolve_2d_cuda[griddim, blockdim](result, kernel, padded_image)
else:
result = _convolve_2d(kernel, padded_image)
if pad:
result = result[pad_rows:-pad_rows, pad_cols:-pad_cols]
if result.shape != image.shape:
raise ValueError("Output and input rasters are not the same shape.")
return result
def _run_cupy(data: cupy.ndarray) -> cupy.ndarray:
data = data.astype(cupy.float32)
griddim, blockdim = cuda_args(data.shape)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
_run_gpu[griddim, blockdim](data, out)
return out
def savi(nir_agg, red_agg, soil_factor=1.0, name='savi', use_cuda=True, use_cupy=True):
"""Returns Soil Adjusted Vegetation Index (SAVI).
Parameters
----------
nir_agg : DataArray
near-infrared band data
red_agg : DataArray
red band data
soil_factor : float
soil adjustment factor between -1.0 and 1.0.
when set to zero, savi will return the same as ndvi
Returns
-------
data: DataArray
Notes:
------
Algorithm References:
- https://www.sciencedirect.com/science/article/abs/pii/003442578890106X
"""
_check_is_dataarray(nir_agg, 'near-infrared')
_check_is_dataarray(red_agg, 'red')
if not red_agg.shape == nir_agg.shape:
raise ValueError("red_agg and nir_agg expected to have equal shapes")
if soil_factor > 1.0 or soil_factor < -1.0:
raise ValueError("soil factor must be between (-1.0, 1.0)")
nir_data = nir_agg.data
red_data = red_agg.data
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(nir_data.shape)
soil_factor_arr = np.array([float(soil_factor)], dtype='f4')
out = np.empty(nir_data.shape, dtype='f4')
out[:] = np.nan
if use_cupy:
import cupy
out = cupy.asarray(out)
_savi_gpu[griddim, blockdim](nir_data,
red_data,
soil_factor_arr,
out)
else:
out = _savi(nir_agg.data, red_agg.data, soil_factor)
return DataArray(out,
name=name,
coords=nir_agg.coords,
dims=nir_agg.dims,
attrs=nir_agg.attrs)
def _run_cupy_bin(data, bins, new_values):
bins_cupy = cupy.asarray(bins)
new_values_cupy = cupy.asarray(new_values)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
_run_gpu_bin[griddim, blockdim](data, bins_cupy, new_values_cupy, out)
return out
def _mean_cupy(data, excludes):
out = cupy.zeros_like(data)
out[:, :] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
_mean_gpu[griddim, blockdim](data, cupy.asarray(excludes), out)
return out
def _run_cupy(data: cupy.ndarray, cellsize_x: Union[int, float],
cellsize_y: Union[int, float]) -> cupy.ndarray:
cellsize_x_arr = cupy.array([float(cellsize_x)], dtype='f4')
cellsize_y_arr = cupy.array([float(cellsize_y)], dtype='f4')
griddim, blockdim = cuda_args(data.shape)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
_run_gpu[griddim, blockdim](data, cellsize_x_arr, cellsize_y_arr, out)
return out
def sipi(nir_agg, red_agg, blue_agg, name='sipi', use_cuda=True, use_cupy=True):
"""Computes Structure Insensitive Pigment Index which helpful
in early disease detection
Parameters
----------
nir_agg : DataArray
near-infrared band data
green_agg : DataArray
green band data
Returns
-------
data: DataArray
Notes:
------
Algorithm References:
https://en.wikipedia.org/wiki/Enhanced_vegetation_index
"""
_check_is_dataarray(nir_agg, 'near-infrared')
_check_is_dataarray(red_agg, 'red')
_check_is_dataarray(blue_agg, 'blue')
if not red_agg.shape == nir_agg.shape == blue_agg.shape:
raise ValueError("input layers expected to have equal shapes")
nir_data = nir_agg.data
red_data = red_agg.data
blue_data = blue_agg.data
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(nir_data.shape)
out = np.empty(nir_data.shape, dtype='f4')
out[:] = np.nan
if use_cupy:
import cupy
out = cupy.asarray(out)
_sipi_gpu[griddim, blockdim](nir_data,
red_data,
blue_data,
out)
else:
out = _sipi(nir_data, red_data, blue_data)
return DataArray(out,
name=name,
coords=nir_agg.coords,
dims=nir_agg.dims,
attrs=nir_agg.attrs)
def _run_cupy_bin(data, bins, new_values):
# replace inf by nan to avoid classify these values as we want to treat them as outliers
data = cupy.where(data == cupy.inf, cupy.nan, data)
data = cupy.where(data == -cupy.inf, cupy.nan, data)
bins_cupy = cupy.asarray(bins)
new_values_cupy = cupy.asarray(new_values)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
_run_gpu_bin[griddim, blockdim](data, bins_cupy, new_values_cupy, out)
return out
def _run_cupy(data: cupy.ndarray, cellsize: Union[int, float]) -> cupy.ndarray:
cellsize_arr = cupy.array([float(cellsize)], dtype='f4')
# TODO: add padding
griddim, blockdim = cuda_args(data.shape)
out = cupy.empty(data.shape, dtype='f4')
out[:] = cupy.nan
_run_gpu[griddim, blockdim](data, cellsize_arr, out)
return out
def ebbi(red_agg, swir_agg, tir_agg, name='ebbi', use_cuda=True, use_cupy=True):
"""Computes Enhanced Built-Up and Bareness Index
Parameters
----------
red_agg : DataArray
red band data
swir_agg : DataArray
shortwave infrared band data
tir_agg : DataArray
thermal infrared band data
Returns
-------
data: DataArray
Notes:
------
Algorithm References:
https://rdrr.io/cran/LSRS/man/EBBI.html
"""
_check_is_dataarray(red_agg, 'red')
_check_is_dataarray(swir_agg, 'swir')
_check_is_dataarray(tir_agg, 'thermal infrared')
if not red_agg.shape == swir_agg.shape == tir_agg.shape:
raise ValueError("input layers expected to have equal shapes")
red_data = red_agg.data
swir_data = swir_agg.data
tir_data = tir_agg.data
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(red_data.shape)
out = np.empty(red_data.shape, dtype='f4')
out[:] = np.nan
if use_cupy:
import cupy
out = cupy.asarray(out)
_sipi_gpu[griddim, blockdim](red_data,
swir_data,
tir_data,
out)
else:
out = _sipi(red_data, swir_data, tir_data)
return DataArray(out,
name=name,
coords=red_agg.coords,
dims=red_agg.dims,
attrs=red_agg.attrs)
def _run_normalized_ratio(arr1, arr2, use_cuda=True, use_cupy=True):
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(arr1.shape)
out = np.empty(arr1.shape, dtype='f4')
out[:] = np.nan
if use_cupy:
import cupy
out = cupy.asarray(out)
_normalized_ratio_gpu[griddim, blockdim](arr1, arr2, out)
else:
out = _normalized_ratio(arr1, arr2)
return out
def _perlin_cupy(data: cupy.ndarray, freq: tuple, seed: int) -> cupy.ndarray:
# cupy.random.seed(seed)
# p = cupy.random.permutation(2**20)
# use numpy.random then transfer data to GPU to ensure the same result
# when running numpy backed and cupy backed data array.
np.random.seed(seed)
p = cupy.asarray(np.random.permutation(2**20))
p = cupy.append(p, p)
griddim, blockdim = cuda_args(data.shape)
_perlin_gpu[griddim, blockdim](p, 0, freq[0], 0, freq[1], 1, data)
minimum = cupy.amin(data)
maximum = cupy.amax(data)
data[:] = (data - minimum) / (maximum - minimum)
return data
def _run_cupy(d_data, azimuth, angle_altitude):
# Precompute constant values shared between all threads
altituderad = angle_altitude * np.pi / 180.
sin_altituderad = np.sin(altituderad)
cos_altituderad = np.cos(altituderad)
azimuthrad = (360.0 - azimuth) * np.pi / 180.
# Allocate output buffer and launch kernel with appropriate dimensions
output = cupy.empty(d_data.shape, np.float32)
griddim, blockdim = cuda_args(d_data.shape)
_gpu_calc_numba[griddim, blockdim](d_data, output, sin_altituderad,
cos_altituderad, azimuthrad)
# Fill borders with nans.
output[0, :] = cupy.nan
output[-1, :] = cupy.nan
output[:, 0] = cupy.nan
output[:, -1] = cupy.nan
return output
def _hotspots_cupy(raster, kernel):
if not (issubclass(raster.data.dtype.type, cupy.integer)
or issubclass(raster.data.dtype.type, cupy.floating)):
raise ValueError("data type must be integer or float")
data = raster.data.astype(cupy.float32)
# apply kernel to raster values
mean_array = convolve_2d(data, kernel / kernel.sum())
# calculate z-scores
global_mean = cupy.nanmean(data)
global_std = cupy.nanstd(data)
if global_std == 0:
raise ZeroDivisionError(
"Standard deviation of the input raster values is 0.")
z_array = (mean_array - global_mean) / global_std
out = cupy.zeros_like(z_array, dtype=cupy.int8)
griddim, blockdim = cuda_args(z_array.shape)
_run_gpu_hotspots[griddim, blockdim](z_array, out)
return out
def _run_cupy(data: cupy.ndarray,
cellsize_x: Union[int, float],
cellsize_y: Union[int, float]) -> cupy.ndarray:
cellsize_x_arr = cupy.array([float(cellsize_x)], dtype='f4')
cellsize_y_arr = cupy.array([float(cellsize_y)], dtype='f4')
pad_rows = 3 // 2
pad_cols = 3 // 2
pad_width = ((pad_rows, pad_rows),
(pad_cols, pad_cols))
slope_data = np.pad(data, pad_width=pad_width, mode="reflect")
griddim, blockdim = cuda_args(slope_data.shape)
slope_agg = cupy.empty(slope_data.shape, dtype='f4')
slope_agg[:] = cupy.nan
_run_gpu[griddim, blockdim](slope_data,
cellsize_x_arr,
cellsize_y_arr,
slope_agg)
out = slope_agg[pad_rows:-pad_rows, pad_cols:-pad_cols]
return out
def _gci_cupy(nir_data, green_data):
griddim, blockdim = cuda_args(nir_data.shape)
out = cupy.empty(nir_data.shape, dtype='f4')
out[:] = cupy.nan
_gci_gpu[griddim, blockdim](nir_data, green_data, out)
return out
def aspect(agg, name='aspect', use_cuda=True, pad=True, use_cupy=True):
"""Returns downward slope direction in compass degrees (0 - 360) with 0 at 12 o'clock.
Parameters
----------
agg : DataArray
Returns
-------
data: DataArray
Notes:
------
Algorithm References:
- http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-analyst-toolbox/how-aspect-works.htm#ESRI_SECTION1_4198691F8852475A9F4BC71246579FAA
- Burrough, P. A., and McDonell, R. A., 1998. Principles of Geographical Information Systems (Oxford University Press, New York), pp 406
"""
if not isinstance(agg, DataArray):
raise TypeError("agg must be instance of DataArray")
if has_cuda() and use_cuda:
if pad:
pad_rows = 3 // 2
pad_cols = 3 // 2
pad_width = ((pad_rows, pad_rows),
(pad_cols, pad_cols))
else:
# If padding is not desired, set pads to 0
pad_rows = 0
pad_cols = 0
pad_width = 0
data = np.pad(agg.data, pad_width=pad_width, mode="reflect")
griddim, blockdim = cuda_args(data.shape)
out = np.empty(data.shape, dtype='f4')
out[:] = np.nan
if use_cupy:
import cupy
out = cupy.asarray(out)
_horn_aspect_cuda[griddim, blockdim](data, out)
if pad:
out = out[pad_rows:-pad_rows, pad_cols:-pad_cols]
elif isinstance(agg.data, da.Array):
out = agg.data.map_overlap(_horn_aspect,
depth=(1, 1),
boundary=np.nan,
meta=np.array(()))
else:
out = _horn_aspect(agg.data)
return DataArray(out,
name=name,
dims=agg.dims,
coords=agg.coords,
attrs=agg.attrs)
def _ebbi_cupy(red_data, swir_data, tir_data):
griddim, blockdim = cuda_args(red_data.shape)
out = cupy.empty(red_data.shape, dtype='f4')
out[:] = cupy.nan
_ebbi_gpu[griddim, blockdim](red_data, swir_data, tir_data, out)
return out
def _focal_stats_func_cupy(data, kernel, func=_focal_max_cuda):
out = cupy.empty(data.shape, dtype='f4')
out[:, :] = cupy.nan
griddim, blockdim = cuda_args(data.shape)
func[griddim, blockdim](data, kernel, cupy.asarray(out))
return out
def convolve_2d(image: xr.DataArray,
kernel,
pad=True,
use_cuda=True) -> xr.DataArray:
"""
Calculates, for all inner cells of an array, the 2D convolution of
each cell via Numba. To account for edge cells, a pad can be added
to the image array. Convolution is frequently used for image
processing, such as smoothing, sharpening, and edge detection of
images by elimatig spurious data or enhancing features in the data.
Parameters:
----------
image: xarray.DataArray
2D array of values to processed and padded.
kernel: array-like object
Impulse kernel, determines area to apply
impulse function for each cell.
pad: Boolean
To compute edges set to True.
use-cuda: Boolean
For parallel computing set to True.
Returns:
----------
convolve_agg: xarray.DataArray
2D array representation of the impulse function.
All other input attributes are preserverd.
Examples:
----------
Imports
>>> import numpy as np
>>> import xarray as xr
>>> from xrspatial import convolution, focal
Create Data Array
>>> agg = xr.DataArray(np.array([[0, 0, 0, 0, 0, 0, 0],
>>> [0, 0, 2, 4, 0, 8, 0],
>>> [0, 2, 2, 4, 6, 8, 0],
>>> [0, 4, 4, 4, 6, 8, 0],
>>> [0, 6, 6, 6, 6, 8, 0],
>>> [0, 8, 8, 8, 8, 8, 0],
>>> [0, 0, 0, 0, 0, 0, 0]]),
>>> dims = ["lat", "lon"],
>>> attrs = dict(res = 1))
>>> height, width = agg.shape
>>> _lon = np.linspace(0, width - 1, width)
>>> _lat = np.linspace(0, height - 1, height)
>>> agg["lon"] = _lon
>>> agg["lat"] = _lat
Create Kernel
>>> kernel = focal.circle_kernel(1, 1, 1)
Create Convolution Data Array
>>> print(convolution.convolve_2d(agg, kernel))
[[ 0. 0. 4. 8. 0. 16. 0.]
[ 0. 4. 8. 10. 18. 16. 16.]
[ 4. 8. 14. 20. 24. 30. 16.]
[ 8. 16. 20. 24. 30. 30. 16.]
[12. 24. 30. 30. 34. 30. 16.]
[16. 22. 30. 30. 30. 24. 16.]
[ 0. 16. 16. 16. 16. 16. 0.]]
"""
# Don't allow padding on (1, 1) kernel
if (kernel.shape[0] == 1 and kernel.shape[1] == 1):
pad = False
if pad:
pad_rows = kernel.shape[0] // 2
pad_cols = kernel.shape[1] // 2
pad_width = ((pad_rows, pad_rows), (pad_cols, pad_cols))
else:
# If padding is not desired, set pads to 0
pad_rows = 0
pad_cols = 0
pad_width = 0
padded_image = np.pad(image, pad_width=pad_width, mode="reflect")
result = np.empty_like(padded_image)
if has_cuda() and use_cuda:
griddim, blockdim = cuda_args(padded_image.shape)
_convolve_2d_cuda[griddim, blockdim](result, kernel, padded_image)
else:
result = _convolve_2d(kernel, padded_image)
if pad:
result = result[pad_rows:-pad_rows, pad_cols:-pad_cols]
if result.shape != image.shape:
raise ValueError("Output and input rasters are not the same shape.")
return result
def _savi_cupy(nir_data, red_data, soil_factor):
griddim, blockdim = cuda_args(nir_data.shape)
out = cupy.empty(nir_data.shape, dtype='f4')
out[:] = cupy.nan
_savi_gpu[griddim, blockdim](nir_data, red_data, soil_factor, out)
return out
def _run_normalized_ratio_cupy(arr1, arr2):
griddim, blockdim = cuda_args(arr1.shape)
out = cupy.empty(arr1.shape, dtype='f4')
out[:] = cupy.nan
_normalized_ratio_gpu[griddim, blockdim](arr1, arr2, out)
return out
def _arvi_cupy(nir_data, red_data, blue_data):
griddim, blockdim = cuda_args(nir_data.shape)
out = cupy.empty(nir_data.shape, dtype='f4')
out[:] = cupy.nan
_arvi_gpu[griddim, blockdim](nir_data, red_data, blue_data, out)
return out
