Air Quality Prediction with A Meteorology-Guided Modality-Decoupled Spatio-Temporal Network
Hang Yin, Yan-Ming Zhang, Jian Xu, Jian-Long Chang, Yin Li, Cheng-Lin Liu · 2025
Read Air Quality Prediction with A Meteorology-Guided Modality-Decoupled Spatio-Temporal Network by Hang Yin, Yan-Ming Zhang, Jian Xu, Jian-Long Chang, Yin Li, Cheng-Lin Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-10014.txt.
Introduction
Hang Yin and co-authors gave a network for air quality prediction that treats weather as its own input stream in two thousand twenty-five. Air quality depends on emissions and on weather, which spreads and transforms pollutants. Past models underused detailed weather data, the authors argue.
Their encoder-decoder network, MDSTNet, models air readings and atmospheric data as separate modalities that interact. It uses weather at many pressure levels plus forecasts. They also built ChinaAirNet, a nationwide dataset pairing air quality with layered weather records. On that data the model cut forty-eight-hour prediction error by about seventeen percent versus the best prior model. Code and data will be shared.
Catalog record
| CREATOR | Hang Yin, Yan-Ming Zhang, Jian Xu, Jian-Long Chang, Yin Li, Cheng-Lin Liu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-10014
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Air Quality Prediction with A Meteorology-Guided Modality-Decoupled Spatio-Temporal Network Hang Yin, Yan-Ming Zhang, Jian Xu, Jian-Long Chang, Yin Li, Cheng-Lin Liu Abstract Air quality prediction plays a crucial role in public health and environmental protection. Accurate air quality prediction is a complex multivariate spatiotemporal problem, that involves interactions across temporal patterns, pollutant correlations, spatial station dependencies, and particularly meteorological influences that govern pollutant dispersion and chemical transformations. Existing works underestimate the critical role of atmospheric conditions in air quality prediction and neglect comprehensive meteorological data utilization, thereby impairing the modeling of dynamic interdependencies between air quality and meteorological data. To overcome this, we propose MDSTNet, an encoder-decoder framework that explicitly models air quality observations and atmospheric conditions as distinct modalities, integrating multi-pressure-level meteorological data and weather forecasts to capture atmosphere-pollution dependencies for prediction. Meantime, we construct ChinaAirNet, the first nationwide dataset combining air quality records with multi-pressure-level meteorological observations. Experimental results on ChinaAirNet demonstrate MDSTNet’s superiority, substantially reducing 48-hour prediction errors by 17.54% compared to the state-of-the-art model. The source code and dataset will be available on github. † † impactstatement: Air quality prediction serves as a critical tool for preventive public health management by enabling early warnings to reduce pollution-induced disease burdens. As a multivariate spatiotemporal forecasting task, it requires modeling multi-location pollutant concentrations through time. Existing predictive models mostly underestimate the critical influence of meteorological conditions. This paper proposes the Meteorology-Guided Modality-Decoupled Spatio-Temporal Network (MDSTNet). MDSTNet decouples air quality observations and meteorological conditions as two distinct modalities while integrating weather forecast as dynamic input prompts to guide air quality prediction. MDSTNet is supposed to effectively reduce prediction errors, establishing a new paradigm for meteorology-guided air quality prediction. Meanwhile, we also built a nationwide dataset named ChinaAirNet, which contains two years air quality observations integrated with multi-pressure level meteorological data. As an open-access resource, ChinaAirNet enhances research reproducibility while providing the first unified benchmark for systematically modeling multi-pressure level meteorological factors and air pollution interactions. I Introduction The urbanization and industrialization of the world have inevitably led to serious air pollution problems. Exposure to air pollution is directly linked to the prevalence of non-communicable diseases such as heart disease, asthma, and lung cancer [ 1 , 2 ] , and the high cost of treatment places a massive burden on governments and individuals. This situation can be greatly mitigated if appropriate countermeasures, such as reminding people to wear masks, limiting factory emissions, and restricting vehicles, are taken in advance. Therefore, accurate air quality prediction has become a research hotspot for public health, national economy, and urban management. Air quality prediction as a multivariate spatiotemporal prediction problem involves jointly forecasting the concentrations of various pollutants at multiple locations for a specified time range. The COVID-19 lockdowns demonstrated that even with drastic anthropogenic emission reductions, air quality improvements remained limited. This indicated that air quality variations were predominantly influenced by meteorological conditions [ 3 ] . Studies [ 4 , 5 , 6 ] highlight that the interactions between meteorological factors and air pollution exhibit complex temporal and spatial dynamics, with nonlinear effects playing a critical role in pollutant formation and dispersion. Therefore, the main difficulty of accurate air quality prediction lies in comprehensive modeling of the complex spatiotemporal correlations between meteorological factors and air pollution, which primarily stems from two challenges. First, air quality is characterized by pollutant dispersion patterns while weather follows pressure-driven fluid dynamics, which exhibit distinct spatiotemporal patterns. Therefore, air quality and meteorological conditions constitute distinct modalities that require separate modeling to prevent cross-modal interference. Second, AI weather models [ 7 , 8 , 9 , 10 ] provide high-accuracy weather forecasts that could significantly enhance air quality prediction. The temporal gap between weather forecasts and historical air quality observations complicates their integration. Despite the numerous efforts in advancing air quality prediction research in recent years [ 11 , 12 , 13 , 14 , 15 , 16 , 17 ] , existing approaches predominantly frame this task as a multivariate spatiotemporal prediction problem. These conventional methods [ 18 , 19 ] focus on modeling either the temporal dynamics inherent in air quality data or the spatial dependencies between monitoring stations, while frequently underestimating the critical influence of meteorological conditions that fundamentally govern pollutant dispersion and chemical transformations. On the one hand, all these studies [ 11 , 12 , 13 ] are confined to utilizing only historical air quality observations or integrating limited surface-level meteorological features for forecasting, while disregarding the predictive value of upper-air pressure levels meteorological data and weather forecasting data. On the other hand, conventional approaches [ 14 , 16 , 17 ] often naively concatenate air quality with meteorological features as homogeneous input vectors. However, air quality exhibit starkly dissimilar spatiotemporal patterns compared to meteorological fluid dynamics. Such naive fusion forces both modalities into a unified spatiotemporal dependency and fundamentally misrepresents their heterogeneous modalities, impairing the model’s capacity to capture cross-modal dependencies and meteorology-guided pollutant dispersion processes. To overcome the above shortcomings, in this paper, we propose Meteorology-Guided M odality- D ecoupled S patio- T emporal Net work (MDSTNet), a novel deep learning framework that explicitly and carefully models meteorological conditions for air quality prediction through cross-modal interaction. MDSTNet employs an encoder-decoder framework, where the encoder formalizes air quality observations and meteorological conditions as two distinct modalities. The decoder adaptively aligns weather forecast with historical air quality representations to resolve temporal asynchrony. Specifically, the encoder integrates stacked transformer layers, with each layer consisting three decoupled parallel attention branches to comprehensively capture historical patterns from three views: spatial correlations within air quality monitoring stations, inter-variable dependencies among diverse pollutants, and cross-modal interactions between pollutants and meteorological features. The decoder utilizes multi-step future weather forecast data as contextual prompts to dynamically filter these learned historical representations through attention mechanisms in spatial and variable dimensions, ultimately generating predictive representations for accurate air quality forecasting. Along with the proposed architecture, we also built a new large-scale dataset named ChinaAirNet, which contains two years of nationwide air quality observations from 1,628 monitoring stations in China. Distinct from existing datasets, ChinaAirNet pioneers the comprehensive integration of air quality records with multi-pressure-level meteorological data (56 meteorological variables from 1000 hPa to 850 hPa), aligning with air quality data at location and time. Our research contributions can be summarized as follows. \bullet We present the first systematic incorporation of multi-pressure-level meteorological data and weather forecast information into air quality prediction, empirically unveiling the critical role of meteorological conditions. \bullet We formalize air quality observations and meteorological conditions as two distinct modalities and design a novel multi-modal encoder-decoder architecture named MDSTNet, where the encoder models different types of dependencies with decoupled parallel branches and an efficient cross-attention technique, while the decoder incorporates weather forecast features as dynamic input prompts to guide air quality prediction. \bullet We construct the first large-scale open-source air quality dataset, ChinaAirNet, which integrates air quality records and multi-pressure-level meteorological observations. \bullet Extensive experimental results demonstrate the effectiveness of our method. Specifically, MDSTNet achieves 17.54% improvement in 48-hour prediction compared to the state-of-the-art model on ChinaAirNet. II Related Work II-A Air Quality Prediction During the past decades, air quality prediction methodologies have evolved through three principal paradigms: physics-based