[Paper Review] Calibration of Spatial Forecasts from Citizen Science Urban Air Pollution Data with Sparse Recurrent Neural Networks

Matthew Bonas, Stefano Castruccio|arXiv (Cornell University)|May 6, 2021

Air Quality Monitoring and Forecasting37 references4 citations

TL;DR

This paper proposes a sparse recurrent neural network with a spike-and-slab prior to calibrate high-resolution spatial forecasts of urban air pollution from citizen science data in San Francisco. By enforcing sparsity and enabling fast marginal and spatial forecast calibration, the model reduces mean squared error by 35.7% over standard time series methods for up to 5-day predictions.

ABSTRACT

With their continued increase in coverage and quality, data collected from personal air quality monitors has become an increasingly valuable tool to complement existing public health monitoring system over urban areas. However, the potential of using such `citizen science data' for automatic early warning systems is hampered by the lack of models able to capture the high-resolution, nonlinear spatio-temporal features stemming from local emission sources such as traffic, residential heating and commercial activities. In this work, we propose a machine learning approach to forecast high-frequency spatial fields which has two distinctive advantages from standard neural network methods in time: 1) sparsity of the neural network via a spike-and-slab prior, and 2) a small parametric space. The introduction of stochastic neural networks generates additional uncertainty, and in this work we propose a fast approach for forecast calibration, both marginal and spatial. We focus on assessing exposure to urban air pollution in San Francisco, and our results suggest an improvement of 35.7% in the mean squared error over standard time series approach with a calibrated forecast for up to 5 days.

Motivation & Objective

• To address the challenge of forecasting high-resolution, nonlinear spatio-temporal air pollution patterns in urban areas using low-cost, citizen-collected sensor data.
• To overcome limitations of standard time series models in capturing local emission sources such as traffic and residential heating.
• To develop a machine learning framework with a small parametric space and inherent uncertainty quantification for reliable long-term forecasting.
• To introduce a fast calibration method that improves both marginal and spatial forecast reliability for urban air quality monitoring.

Proposed method

• The model employs a sparse recurrent neural network architecture with a spike-and-slab prior to enforce sparsity and reduce model complexity.
• Stochasticity in the network introduces uncertainty, which is explicitly modeled and calibrated using a fast, scalable calibration procedure.
• The calibration method is applied jointly to marginal distributions and spatial dependence structures of forecast outputs.
• The approach is trained and evaluated on high-frequency, spatially dense air pollution data collected from personal monitors in San Francisco.
• Model performance is benchmarked against standard time series forecasting methods using mean squared error (MSE) as the primary metric.
• The framework is designed to handle sparse, irregularly sampled urban air quality data while preserving local emission source dynamics.

Experimental results

Research questions

• RQ1 Can a sparse recurrent neural network with spike-and-slab priors effectively model high-resolution spatio-temporal air pollution patterns from citizen science data?
• RQ2 How does the proposed calibration method improve forecast reliability compared to standard time series approaches?
• RQ3 To what extent does the model reduce prediction error in multi-day forecasts across urban microenvironments?
• RQ4 Can the model capture nonlinear, localized pollution sources such as traffic and heating emissions more effectively than traditional models?

Key findings

• The proposed model achieves a 35.7% reduction in mean squared error compared to standard time series forecasting methods for up to 5-day spatial forecasts.
• The integration of spike-and-slab priors enables a small parametric space while maintaining high predictive accuracy and interpretability.
• The fast calibration procedure effectively improves both marginal and spatial forecast reliability, enhancing forecast trustworthiness.
• The model successfully captures complex, nonlinear spatio-temporal dynamics driven by local urban emission sources such as traffic and residential heating.
• The approach demonstrates strong generalization on high-frequency, sparse urban air quality monitoring data from personal sensors.
• The results confirm that calibrated forecasts from sparse stochastic networks outperform conventional models in urban air pollution prediction tasks.