Research map / Image generation / Autoregressive and latent generation
AirRadar: Inferring Nationwide Air Quality in China with Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AirRadar, a masked-token spatial-causal neural network that infers PM2.5 at unmonitored locations across China from nearby stations, outperforming STFNN and other baselines.
Problem
Real-time air quality inference at locations without monitoring stations, at nationwide scale with thousands of stations, where complex spatial correlations and spatial heterogeneity must be handled.
Approach
Learnable mask tokens replace unobserved station features; a Spatial Learning Module combines dartboard-projection local attention (wind-aligned) with FFT-based global mixing, and a Causal Learning Module applies backdoor adjustment with learned, dynamically weighted context encoders to handle spatial heterogeneity.
Key results
- AirRadar achieves MAE 6.41 at 25% missing ratio vs. 11.14 for STFNN (about 40% reduction), and 6.79 at 50% (-44.9%).
- At 75% missing ratio AirRadar reaches MAE 8.11 vs. 11.27 for STFNN (-28.0%).
- Ablations show removing the global FNO module degrades MAE more than removing the local dartboard module (e.g., 11.68 vs. 8.11 at 75%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AirRadar instead of STFNN | Nationwide air quality inference · Nationwide China, missing ratio 25% | MAE: 11.14 -> 6.41 | Helped |
| use AirRadar instead of STFNN | Nationwide air quality inference · Nationwide China, missing ratio 50% | MAE: 11.32 -> 6.79 | Helped |
| use AirRadar instead of STFNN | Nationwide air quality inference · Nationwide China, missing ratio 75% | MAE: 11.27 -> 8.11 | Helped |
| add the local spatial component (ablation) | AirRadar · Nationwide China, missing ratio 25% | MAE: 8.67 -> 6.41 | Helped |
| add the local spatial component (ablation) | AirRadar · Nationwide China, missing ratio 50% | MAE: 7.45 -> 6.79 | Helped |
| add the local spatial component (ablation) | AirRadar · Nationwide China, missing ratio 75% | MAE: 9.09 -> 8.11 | Helped |
| add the global spatial component (ablation) | AirRadar · Nationwide China, missing ratio 25% | MAE: 8.43 -> 6.41 | Helped |
| add the global spatial component (ablation) | AirRadar · Nationwide China, missing ratio 50% | MAE: 9.22 -> 6.79 | Helped |
| add the global spatial component (ablation) | AirRadar · Nationwide China, missing ratio 75% | MAE: 11.68 -> 8.11 | Helped |
| include historical data (ablation) | AirRadar · Nationwide China, all missing ratios | performance | Helped |
| include the Causal Learning Module (ablation) | AirRadar · Nationwide China | performance | Helped |
| include adaptive dynamic weight allocation (ablation) | AirRadar Causal Learning Module · Nationwide China | performance | Helped |
Surprising findings
- Performance is not strictly proportional to the number of causal contexts; C=4 chosen as default.
Limitations
- Evaluated on a single year (2018) of data from one nationwide dataset.
- Comparison to STFNN is the only strong baseline; reported gains are relative to it on one dataset.
Experimental setup
Datasets: nationwide China air quality dataset (2018, 1,085 stations)
Benchmarks: nationwide China air quality dataset
Metrics: MAE RMSE MAPE
Baselines: KNN RF SGNP STGNP VAE GAE GraphMAE MCAM STFNN
Builds on: STFNN MAE GraphMAE Fourier Neural Operator dartboard projection
Closest papers by meaning
- AirPhyNet: Harnessing Physics-Guided Neural Networks for Air Quality Prediction (ICLR 2024 · similarity 0.87)
- Air Quality Prediction with Physics-Guided Dual Neural ODEs in Open Systems (ICLR 2025 · similarity 0.83)
- Indoor Air Quality Dataset with Activities of Daily Living in Low to Middle-income Communities (NeurIPS 2024 · similarity 0.78)
- HR-Extreme: A High-Resolution Dataset for Extreme Weather Forecasting (ICLR 2025 · similarity 0.78)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.78)
- SNN-PDE: Learning Dynamic PDEs from Data with Simplicial Neural Networks (AAAI 2024 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.