A Non-stationary, Amortized, Transfer Learning Approach for Modeling Italian Air Quality

2026-04-20

AI Summary

This study addresses the challenge of sparse and unevenly distributed air quality monitoring stations in Italy, coupled with the inability of chemical transport models to capture fine-scale pollution variability in complex terrain. To overcome this, the authors propose a spatial transfer learning framework that integrates ground observations with model outputs to produce high-resolution, uncertainty-quantified daily NO₂ concentration maps across Italy for 2023. The approach builds upon the LatticeKrig geostatistical model, incorporating a nonstationary, anisotropic covariance structure whose spatially varying parameters are estimated via an image-to-image neural network. Computational efficiency is achieved through compactly supported basis functions and sparse precision matrices, enabling scalable cross-resolution modeling. Experiments demonstrate that the method consistently outperforms traditional stationary models at both station and grid scales, substantially improving prediction accuracy in geographically complex regions and offering robust support for epidemiological research and environmental policymaking.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Remote Sensing / Geospatial AI
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Air quality monitoring in Italy relies on sparse, irregular, ground-based stations that provide high-quality but incomplete measurements of pollution. Chemical transport models (CTMs) offer full spatial and temporal coverage but smooth over local variability. We develop a spatial transfer-learning framework that integrates these two data sources to produce daily, fine-grid predictions of nitrogen dioxide (NO$_2$) concentrations across Italy for 2023, with uncertainty quantification. The resulting maps provide a resource for decision making in downstream applications such as epidemiology and environmental policy. Our approach builds on the geostatistical LatticeKrig framework, which uses compactly supported basis functions and coefficients governed by a sparse precision matrix. We learn a nonstationary, anisotropic correlation structure from the gridded CTM outputs using an image-to-image neural architecture that estimates millions of spatially varying parameters in a matter of seconds. The basis-function representation enables this covariance structure to be transferred to the point-level station data and projected onto a finer prediction grid, a key extension for handling the change of support between data sources. A likelihood-based refinement step then adjusts the correlation range to recover fine-scale variability smoothed out by the gridded data. The proposed methodology results in a flexible, non-stationary, and anisotropic representation of the spatial process, better accommodating the complex geography of Italy. Performance is assessed through experiments on both gridded CTM outputs and point-level station measurements, demonstrating improvements over the stationary formulation.

Problem

• air quality modeling
• data fusion
• non-stationarity
• transfer learning
• spatial prediction

Innovation

• non-stationary spatial modeling
• amortized inference
• transfer learning
• LatticeKrig
• anisotropic covariance