Bayesian Change-of-Support Estimation for Resolution-Limited Spatial Images

Transport in Porous Media · Springer Nature · 2026-11-07 · DOI: 10.1007/s11004-026-10358-x

Аннотация

Regularized inversion returns images that are linear smoothings of an underlying field. The smoothing operator is known but nonstationary. It is the model resolution matrix of the inverse problem. When several such images are available, each with a different resolution matrix and each responding nonlinearly to a different combination of the same variables, classical change-of-support methods no longer apply. We formulate the problem in latent space. Point-support latent fields carry geostatistical priors and are observed through a petrophysical relationship composed with an image-specific resolution matrix. Sparse direct measurements enter as point observations, and iterated linear-Gaussian updating gives the posterior. Two properties are central, and both hold under linear relationships, Gaussian errors and a correctly specified model. The estimator is sufficient in the following sense. For every regularization weight, an image together with its resolution matrix and propagated error covariance yields the same posterior as the measurements used to produce the image. The estimator also provides a diagnostic. Pointwise likelihood identifiability follows from a rank condition on the stacked resolution-weighted Jacobian, a condition that can be evaluated before acquisition. Two further properties link the estimator to existing methods, and synthetic studies confirm each. Over an ensemble of realizations, latent joint estimation recovers porosity at point support with a coefficient of determination of $$0.77 \pm 0.08$$, against $$0.61 \pm 0.16$$ and $$0.38 \pm 0.20$$ for two established practices. The formulation applies wherever latent fields are observed through known, nonstationary smoothing operators. Its central consequence is that an inverted image reported with its resolution matrix is a sufficient statistic for the measurements behind it, whereas an image reported without one is strictly less informative.

Читать у издателя →

Похожие статьи