Advancing the Frontiers of Urban Mobile Sources for Air Pollution Prediction and Monitoring: Integrating Deep Learning, Quantum Computing, and Enhanced Sensing

The relentless challenge of air pollution, driven by industrialization and urbanization, demands increasingly sophisticated tools for accurate prediction, source attribution, and comprehensive monitoring. Protecting public health and informing effective mitigation policies hinges on our ability to understand and forecast the complex dynamics of atmospheric pollutants. This Special Issue showcases cutting-edge research pushing the boundaries in these domains, leveraging advancements in deep learning (DL), quantum computing, signal processing, and sensing technologies. The contributions highlight a shift towards more integrated, explainable, and computationally efficient approaches.

Zhivkov et al. [1] pioneered a digital twin framework integrating mobile/stationary sensors and machine learning to transform urban air quality assessment. Our two-step calibration dramatically enhanced low-cost PM sensor accuracy (R²: 0.29→ 0.87–0.95), enabling scalable networks. Mobile deployments revealed critical pollution hotspots—300% concentration spikes near traffic corridors during rush hours, overlooked by stationary monitors. Graph neural network analysis quantified sources; for example, it was demonstrated that 65% of roadside PM_2.5 come from vehicles, while residential zones showed dust dominance. Urban greenspaces reduced PM by 30%, yet low-emission zones initially showed negligible impact. This paradigm provides actionable intelligence for targeted mitigation and evidence-based urban planning, advancing healthier city development.

Traditional air quality models often struggle to fully capture the intricate interplay between spatial dispersion patterns and temporal evolution. Wu et al. [2] address this head-on with their Comprehensive Scale Spatiotemporal Fusion Air Quality Predictor (CSST-AQP). By introducing a novel dual-branch architecture, CSST-AQP uniquely combines multi-scale spatial correlation analysis with adaptive temporal modeling. This allows it to explicitly capture the complex interactions governing pollutant dispersion, moving beyond isolated feature analysis. The hierarchical fusion engine integrating features at individual sites and regional scales underpins its state-of-the-art performance (RMSE 6.11–9.13 µg/m³, R²: 0.91–0.93 across 14 Chinese regions) and robust 60-hour forecasting capability for diverse pollutants.

Computational costs remain a significant bottleneck in high-fidelity numerical weather prediction (NWP) models like the Weather Research and Forecasting (WRF) model. Paulo et al. [3] demonstrate the transformative potential of Fourier Neural Operators (FNOs) in accelerating transient wind forecasting. Their key insight was tailoring FNO models to specific atmospheric altitudes, acknowledging that dominant physical processes and optimal input features vary with height. This altitude-specific approach significantly outperformed a single global FNO model and persistence baselines (achieving nRMSE of 27.64% and 20.46% at surface and middle altitudes), providing highly optimized initial conditions that can drastically reduce NWP convergence times.

Exploring the nascent field of quantum machine learning (QML), Victor et al. [4] present a groundbreaking quantum-classical hybrid architecture for ozone prediction in Houston, Texas. By integrating a graph-based DL structure with a quantum neural network (QNN), they achieved remarkable forecasting skill (FS: 31.01% at 1 h, 48.01% at 3 h, 57.46% at 6 h). Their work provides crucial practical insights: smoother classical-to-quantum transition layers, min-max normalization, and increased ansatz repetitions significantly enhanced hybrid model performance, validating QML as a competitive tool for specific air quality forecasting challenges.

Effectively capturing non-stationarity and nonlinear dependencies in pollutant time series is critical. Wang et al. [5] innovate by integrating Wavelet Transform (WT) with a Multilayer Perceptron (MLP), creating the WTMP model. The Overlapping Discrete Wavelet Transform (ODWT) decomposes pollutant data, extracting multi-resolution features that are then fused with static covariates (e.g., time of day) within an improved MLP architecture. This approach, validated on data from Ruian City, China, effectively handles complex temporal patterns leading to minimal prediction errors, demonstrating the power of combining classical signal processing with DL.

Accurate emission factor (EF) estimation is fundamental for source apportionment and inventory development. Xu et al. [6] tackle the limitation of standard models like COPERT in utilizing real-world OBD data. Their novel Two-Stream Network ingeniously fuses raw time-series features (organized in an information matrix capturing historical dependencies) with time-frequency features derived via Continuous Wavelet Transform (CWT) into a multi-channel matrix. Processed through a parallel ResNet50 and Convolutional Block Attention Module (CBAM) stream, this model significantly enhances the accuracy of NOx EF prediction for heavy-duty diesel vehicles, providing a valuable tool for refining transport emission inventories.

While DL models often achieve high accuracy, their “black-box” nature limits interpretability and trust. Pan et al. [7] propose a Hybrid Autoformer Network explicitly designed to predict air pollution while establishing interpretable relationships with external factors (e.g., O₃, wind, temperature). They integrate a Genetic Algorithm (GA) featuring an elite variable voting operator for refined feature selection and an archive storage operator to mitigate initialization sensitivity. This approach not only boosted prediction accuracy by 2–8% on the Ma’anshan dataset but also provided clear insights into the most influential external drivers of pollution events.

Accurate monitoring is the bedrock of assessment and prediction. Balendra et al. [8] provide a comprehensive review of the integration of deep learning (DL) with airborne particulate matter (PM) sensing, highlighting the promising role of Surface Plasmon Resonance (SPR). They detail how DL techniques (CNNs, RNNs, LSTMs, autoencoders) enhance tasks like PM estimation from satellite imagery, sensor calibration, and air quality prediction. The review critically addresses challenges: handling noisy data from low-cost sensors, ensuring model generalizability, and improving interpretability. It emphasizes the potential of hybrid sensing systems coupled with robust DL pipelines, including deep reinforcement learning and advanced data fusion, for future environmental monitoring.

Focusing on specific harmful compound classes, Gao et al. [9] review analytical techniques for atmospheric carbonyl compounds (e.g., formaldehyde, acetaldehyde). Their analysis contrasts direct and indirect sampling methods, offline (e.g., DNPH-HPLC/GC) and online techniques (e.g., PTR-MS, CRDS), highlighting strengths, limitations (e.g., species coverage, stability, time resolution), and optimal applications (e.g., PTR-MS + GC-MS for smog chamber studies). The review underscores the ongoing need for improvements in environmental adaptability, automation, sensitivity, selectivity, and the development of novel sampling systems to reduce chamber interference.

The advancements documented here–spanning highly accurate spatiotemporal prediction, accelerated NWP initialization, quantum-classical hybrids, wavelet-enhanced time series analysis, refined emission factor estimation, explainable models, DL-enhanced sensing, and sophisticated speciation techniques—provide powerful new tools for atmospheric scientists, environmental agencies, and policymakers. The integration of spatial and temporal dynamics, the pursuit of explainability alongside accuracy, and the embrace of novel computational and sensing approaches define the current frontier. As these fields continue to converge and evolve, we move closer to achieving the comprehensive, real-time understanding of air quality necessary to safeguard public health and planetary well-being effectively.