Uncertainty

DepthDif estimates sampling uncertainty by running multiple stochastic reverse diffusion trajectories for the same condition.

For samples \(\hat{x}^{(1)},\ldots,\hat{x}^{(N)}\), the implementation returns the population standard deviation (torch.std(..., unbiased=False)):

\[ \sigma = \sqrt{\frac{1}{N}\sum_{i=1}^{N} \left(\hat{x}^{(i)}-\bar{x}\right)^2}. \]

This is sample dispersion under the configured sampler and checkpoint. It is not a calibrated predictive interval and does not include data or model uncertainty outside those stochastic trajectories.

Export modes

• Repository global exporters use 20 samples by default and preserve depth channels unless --uncertainty-collapse-depth is passed.
• The public weekly API uses 20 samples by default and exports one collapsed uncertainty raster for its legacy checkpoint.
• uncertainty_only=True or --uncertainty-only skips the ordinary prediction export and runs only the uncertainty ensemble.
• The uncertainty sampler and DDIM step count can be overridden independently from the main reconstruction sampler.

from depth_recon import run_week_inference

run_week_inference(
    year=2015,
    iso_week=25,
    export_uncertainty=True,
    uncertainty_num_samples=20,
    uncertainty_sampler="ddim",
    uncertainty_ddim_num_timesteps=50,
)

For depth-resolved uncertainty and reliability data used by the spatial dashboard, use the repository global exporters described in Inference.