Spatial AI should be built to complement classical spatial statistics, rather than replace them, according to a new preprint. Its central argument is that flexible learning and stochastic spatial models should have distinct jobs within the same system.
The paper offers a guide to the field’s development rather than an exhaustive history. It ranges across spatial data observed at points, over regions, on images, through tensors and across networks.
The map is part of the question
The review’s most practical warning is that a finer prediction map is not automatically a more relevant scientific answer. Changes in spatial support—the scale or form represented by a result—must be made explicit.
Kriging illustrates the statistical principle the authors want to preserve: prediction and prediction uncertainty derive from the same stochastic model. The estimate and its uncertainty are built together, rather than treated as unrelated outputs.
Different lenses on spatial patterns
A spectral view offers another way to read a spatial pattern, sorting dependence by frequency. Low frequencies represent broad regional gradients, while high frequencies represent local roughness.
When patterns change from place to place—a condition called nonstationarity—the review says local covariance behavior needs diagnostics and regularization. In ordinary language, local changes need to be checked and constrained before they are treated as reliable structure.
Bayesian spatial modeling separates observation error, variation in the underlying process, uncertainty in the model’s parameters and posterior prediction.
When data from different spatial sources are combined, the review frames the task around a common underlying process, explicit changes in spatial support, correction for bias and uncertainty propagation—carrying uncertainty through the combined analysis.
The costs of ignoring structure
Sampling design can quietly change the picture. The review warns that when informative sampling is ignored, predictions may be distorted and uncertainty understated in regions with few observations.
Computation sets another boundary. The paper says a dense Gaussian-process calculation using Cholesky factorization, a standard matrix decomposition, requires O(n^3) operations and O(n^2) memory. In plain language, the work grows cubically and the memory demand quadratically as the number of observations rises.
How the next generation should be tested
In the proposed hybrid, the learning component handles complex mean structure, while the stochastic component handles residual dependence and uncertainty.
In joint neural-spatial estimation, uncertainty in the fitted mean can be included in spatial prediction, but the learned mean may compete with residual spatial structure. The paper describes this as an identifiability challenge—difficulty separating which component explains what.
Validation, the review says, should match the intended use. It recommends evaluating uncertainty coverage at the scale of use—whether stated ranges contain the values they are meant to contain—and using spatially blocked or regional holdouts rather than random splits when predicting in new locations.
The conclusion is a direction for the field, not a performance ranking: the authors characterize classical stochastic modeling and Spatial AI as complementary, with flexible learning, graph models and spatial processes serving distinct roles.
The arXiv version 1 preprint, dated 20 Aug 2026, therefore reads as a conceptual roadmap rather than evidence that one method outperforms another in a specific application.
Paper data and sources
Original title: From Kriging to Spatial AI: Fifty Years of Spatial Statistics for Complex Dependent Data
Authors: Montserrat Fuentes, Veronica B. Patterson
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text