credit.ensemble.color

credit.ensemble.color#

Classes#

ColorNoise

2D spatially correlated noise generator for lat/lon grids.

Functions#

apply_noise_perturbation_step(→ Tuple[torch.Tensor, ...)

Red noise perturbation in physical units, with per-channel control.

Module Contents#

class credit.ensemble.color.ColorNoise(amplitude: float = 0.05, reddening: int = 2)#

2D spatially correlated noise generator for lat/lon grids.

Generates colored noise with controllable spatial correlation using power-law scaling in the frequency domain. The noise characteristics are determined by the reddening parameter:

  • reddening = 0: White noise (uncorrelated, flat power spectrum)

  • reddening = 1: Pink noise (1/f power spectrum)

  • reddening = 2: Brown/Brownian/Red noise (1/f² power spectrum)

  • reddening > 2: Higher-order red noise (1/f^n power spectrum)

Higher reddening values produce smoother, more spatially coherent patterns, which are often more realistic for geophysical applications.

Parameters:

  • amplitude (float, optional) – Scaling factor for the generated noise. Defaults to 0.05.

  • reddening (int, optional) – Power-law exponent controlling spatial correlation. Higher values create smoother, more correlated noise patterns. Defaults to 2 (Brown noise).

amplitude = 0.05#
reddening = 2#
__call__(x: torch.Tensor) → torch.Tensor#

Generate spatially correlated noise matching input tensor dimensions.

Parameters:

x (torch.Tensor) – Reference tensor whose shape determines the output noise dimensions. The last two dimensions are treated as the spatial lat/lon grid.

Returns:

Spatially correlated noise tensor with the same shape as the input, scaled by the amplitude parameter.

Return type:

torch.Tensor

_create_correlated_noise(shape: tuple[int, Ellipsis], device: torch.device) → torch.Tensor#

Generate spatially correlated noise using frequency domain filtering.

Creates colored noise by: 1. Starting with white noise (uniform power spectrum). 2. Transforming to frequency domain. 3. Applying power-law scaling (1/f^reddening). 4. Transforming back to spatial domain.

Parameters:

  • shape (tuple[int, ...]) – Shape of the output noise tensor.

  • device (torch.device) – Device to generate tensors on.

Returns:

Spatially correlated noise with unit variance (before amplitude scaling).

Return type:

torch.Tensor

__repr__() → str#
credit.ensemble.color.apply_noise_perturbation_step(x: torch.Tensor, delta_prev: torch.Tensor | None, forecast_step: int, rho: float = 0.9, perturbation_std: float | torch.Tensor = 0.05) → Tuple[torch.Tensor, torch.Tensor]#

Red noise perturbation in physical units, with per-channel control.