Source-linked AI summary
Predicting Spatiotemporal Mobile Sensing-Based PM2.5 Concentrations Using Low-Rank Adapted Spatially Attentive Graph Neural Network
Om Chiddarwar, Priyanka Mandal, Praveen Kumar Chandaliya, Shriniwas Arkatkar
TL;DR
Urban transport corridors require high-resolution PM2.5 monitoring because pollution varies across space and time, while mobile sensing provides detailed but complex observations. The paper develops SA-GNN-LoRA using clustered graph nodes, temporal modeling, and spatial attention, and reports stronger forecasting performance than baseline models on mobile-sensing data from Surat.
Problem
Rapidly changing corridor pollution requires fine-grained monitoring, but fixed stations provide limited spatial resolution and mobile sensing introduces irregular, noisy observations.
Method
SA-GNN-LoRA represents mobile observations with fixed-segment or DBSCAN-derived graph nodes, models temporal dependencies with GRU-based modules, and learns spatial interactions with graph attention.
Results
SA-GNN-LoRA achieves the strongest reported performance, with R2=0.95, RMSE=6.84, CSI: 95.79, POD: 98.88, and FAR: 3.15.
Takeaways & Limitations
The framework supports fine-grained real-time PM2.5 forecasting and is presented as useful for public health, environmental policy, and urban sustainability applications.
Abstract
from arXiv · showhide
Urban air quality can vary significantly along transit corridors, necessitating high-resolution monitoring. This work introduces a novel mobile-sensing dataset from Surat, Gujarat, India, comprising PM$*{2.5}$ concentrations, meteorological variables (temperature, humidity, wind speed, wind direction), and land-use features. To represent the spatiotemporal data as a graph, two node-definition strategies were used: (i) uniform segmentation (200--400~m intervals) and (ii) DBSCAN clustering to adaptively group dense observations. For each node, rolling mean and standard deviation of meteorological variables were computed. To model this high-dimensional data, we propose a SA-GNN for fine-grained, short-term PM$*{2.5}$ forecasting and hotspot identification. We compared SA-GNN with LSTM, RNN, GRU, and ANN models. These models performed well on low-resolution data but had difficulty capturing rapidly changing patterns in urban air quality. SA-GNN employs cluster-specific GRUs to capture localized temporal dependencies and a Graph Attention Network to learn spatial heterogeneity. This hybrid architecture effectively models rapid fluctuations and complex spatial interactions. On our dataset, SA-GNN achieved $R^2 = 0.95$, RMSE $= 6.8$, and MAE $= 4.2~\si{\micro\gram\per\meter\cubed}$, outperforming all baseline models. Combining spatial clustering with adaptive attention significantly improves forecasting, enabling real-time, fine-grained monitoring and supporting personalized exposure tracking and timely alerts for healthier cities.
1. Introduction
Urban PM2.5 varies sharply across transport corridors, while fixed stations and conventional models have limited spatial or spatiotemporal coverage. The paper addresses this gap with mobile sensing and SA-GNN-based fine-grained forecasting.
- Motivation: PM2.5 exposure is a major health concern, particularly in rapidly motorizing regions where corridor emissions disproportionately affect commuters and pedestrians.Surat is identified as an increasingly vulnerable industrial and coastal city with elevated winter PM2.5.
- Motivation: Fine-scale PM2.5 concentrations can change rapidly because of local emissions, meteorology, geography, traffic, and land-use patterns.Static monitoring stations miss micro-level fluctuations, while mobile sensing introduces non-uniform sampling, GPS noise, and missing values.
- Research Gap: Traditional linear and machine-learning models capture limited relationships but do not reliably model sequential dependencies, spatial generalizability, or spatial heterogeneity.Deep learning models learn temporal structure but generally fail to capture non-uniform urban spatial variation.
- Approach: SA-GNN constructs dynamic pseudo-station graphs from fixed-length trajectory segments and DBSCAN clusters, combining GRUs for temporal relations with GAT-based spatial weighting.The framework was implemented using a bike-mounted mobile sensing platform that collected more than 14 km of data in Surat.
- Contributions: R2=0.95, RMSE=6.84 µg/m3, and MAE=4.19 µg/m3 were reported for SA-GNN, outperforming LSTM, RNN, and physics-informed GNN baselines.The framework targets fine-grained, real-time forecasting in dynamic mobile-sensing environments.
- Implications: The paper highlights mobile sensing and graph neural networks as a basis for scalable, personalized air-quality forecasting and exposure-aware mobility planning.The paper is organized around prior work, methodology, data and experiments, architecture, analyses, and conclusions.
2. Related Work
Air-quality forecasting has progressed from linear time-series and machine-learning models to deep learning and graph-based approaches. However, existing methods often remain limited by sequential, spatial, grid, fixed-station, or mobility-induced topology constraints.
- Classical and Machine-Learning Methods: ARIMA and Kalman filters model linear temporal patterns efficiently, but stationarity and linearity assumptions restrict their representation of complex PM2.5 relationships.These limitations motivated later nonlinear machine-learning approaches.
- Classical and Machine-Learning Methods: Support Vector Machines, Random Forests, and Gradient Boosting Machines introduce nonlinear modeling and improve short-term accuracy, but remain limited in sequential and spatial modeling.Their use of traffic and land-use features does not fully resolve spatiotemporal dependencies.
- Deep Learning Methods: LSTM and GRU recurrent models improve time-varying pollution modeling, but conventional RNNs can suffer vanishing gradients and fail to exploit spatial relationships between stations.ConvLSTM, 3D-CNN, and attention-augmented CNN models jointly model space and time but often assume uniform grids or fixed sensor placement.
- Graph Neural Network Methods: GNNs naturally represent irregular spatial structures and combine graph operations with temporal modeling, meteorology, land use, emissions, and physical knowledge.Examples include DCRNN, GC-LSTM, PM2.5-GNN, and physics-aware graph networks.
- Proposed Direction: The proposed framework uses DBSCAN for dynamic mobile-sensor clustering, irregular-sequence handling, GRU-based temporal modeling, and GAT-based adaptive spatial influence learning.This design targets localized and personalized prediction from mobile and meteorological data.
3. Problem Formulation
The paper formulates fine-grained mobile-sensing PM2.5 forecasting as spatiotemporal learning on a time-evolving graph. Spatial segments or DBSCAN clusters become nodes, with graph edges encoding proximity or travel distance and the SA-GNN producing future node-level forecasts.
- Graph Representation: Each spatial segment or dynamically formed mobile-sensor cluster is treated as a graph node rather than relying only on fixed air-quality monitoring stations.This representation targets localized pollution variations in heterogeneous urban environments.
- Graph Representation: The time-evolving graph uses fixed-size segmentation or DBSCAN clustering to define nodes and geographical proximity or travel distance to encode edges.Travel-distance edges are intended to better reflect human exposure patterns.
- Inputs and Forecasting Window: Historical PM2.5 concentrations across all n nodes are supplied over a lookback window of length r.The notation defines x(t−l) as concentrations at all nodes at the earlier time step t−l.
- Inputs and Forecasting Window: Auxiliary feature sequences, including meteorological variables, are incorporated alongside the node-level PM2.5 history.The feature representation contains n nodes and d features, while u denotes the number of future steps predicted.
- Forecasting Objective: The objective is to learn a parameterized function fW that forecasts PM2.5 concentrations for all spatial segments or clusters over future timestamps.The proposed function is the Spatially Attentive Cluster-based Graph Neural Network.
- Forecasting Objective: The architecture overview shows historical PM2.5 and meteorological inputs being used to learn spatial and temporal patterns across clustered graph nodes before predicting future concentrations.The formulation is intended to support high-resolution, personalized mobile-sensing forecasts.
4. Dataset Description and Preprocessing
The study builds a high-resolution mobile-sensing dataset along Surat’s 14 km Udhna–Bhestan corridor to characterize commuter PM2.5 exposure across heterogeneous traffic, land-use, temporal, and meteorological conditions. Descriptive patterns show substantial pollution variability, motivating fine-grained preprocessing and feature selection.
- Study corridor: The dataset covers Surat’s 14 km bidirectional Udhna–Bhestan arterial, a heterogeneous corridor spanning residential, commercial, industrial, and transport settings.The route includes seven major intersections and heavy mixed traffic, including two-wheelers, cars, buses, and trucks.
- Monitoring design: 53 repeated runs sampled morning, afternoon, and evening periods across post-monsoon, winter, and pre-monsoon seasons.Runs covered both travel directions, capturing temporal and traffic-related variability.
- Mobile sensing and data collection: The MTW-mounted platform collected PM2.5, meteorological measurements, and GPS-referenced observations to replicate commuter-level exposure.The platform used a calibrated optical particle counter, meteorological probes, and GPS; data were logged every 3 seconds.
- Observed PM2.5 patterns: 100.5 µg/m³ was the mean PM2.5 concentration, with values ranging from 12 to 930 µg/m³ and a heavy upper tail.Over 70% of observations exceeded WHO guidelines, while nearly 40% exceeded national standards.
- Temporal and seasonal variability: 106.8 µg/m³ was the evening mean and 101.7 µg/m³ the morning mean, compared with 93.3 µg/m³ in the afternoon.Weekdays averaged 108.4 µg/m³ versus 64.2 µg/m³ on weekends, while January–February exceeded 120 µg/m³ and April–May averaged about 85 µg/m³.
- Spatial and land-use variability: 119.8 µg/m³ was the highest segment mean, recorded on the 4600–4800 m intersection stretch; industrial areas averaged 95.1 µg/m³, versus 61.3 µg/m³ across residential land use.Hotspots were associated with congestion, idling queues, and industrial frontage.
- Feature selection: Feature selection retained meteorological variables, vehicle speed, and land-use attributes using multiple correlation and relational analyses.Wind direction was retained despite weaker Pearson and Spearman correlations because of its directional influence on pollutant transport.
5. Network Architecture
SA-GNN combines transport-aware spatial modeling, temporal sequence processing, efficient attention, and regularized training for multi-step PM2.5 forecasting.
- Spatiotemporal architecture: The architecture jointly models spatial and temporal dependencies through graph-based attention and recurrent or convolutional temporal structures.The spatiotemporal module represents stations as a directed graph, while the temporal module captures short- and long-term sequence patterns.
- Pollutant transport modeling: Each node combines PM2.5 measurements with meteorological variables, while an advection coefficient represents wind-driven transport from source node j to sink node i.The coefficient uses wind speed, wind direction relative to the node pair, and geographical distance, with ReLU enforcing nonnegative flow.
- Spatial attention: Transport-aware node features pass through two GAT layers to learn hierarchical spatial dependencies and higher-order neighborhood interactions.The resulting spatially attended features are fused with transport embeddings and original node features.
- Temporal module: A 1D CNN extracts local temporal patterns, stacked GRUs model longer dependencies, and an MLP produces temporally encoded forecasting features.These temporal features are concatenated with graph-derived spatiotemporal features before final attention-based prediction.
- Efficient attention and adaptation: Linear Attention uses a kernel-based approximation, while LoRA adapts query, key, and value projections with reduced parameter redundancy.The refined representation is passed to a final MLP to generate PM2.5 forecasts over the prediction horizon.
- Training objective: The training objective combines Mean Squared Error with L2 regularization to balance forecasting accuracy and model complexity.The MSE penalizes prediction errors, while L2 regularization constrains large parameter values to reduce overfitting risk.
6. Results
The evaluation compares SA-GNN variants with sequence and feedforward baselines using forecasting, detection, interpretability, and visualization analyses. SA-GNN with Linear Attention and LoRA achieves the strongest reported overall performance, while contextual features materially improve predictions.
- Baseline comparison: Sequence baselines outperform ANN, with LSTM achieving the lowest baseline MAE (3.75) and RNN recording the highest baseline R2 (0.87).The comparison attributes the sequence-model advantage to their ability to capture temporal dependencies in air-quality time series.
- Quantitative comparison: SA-GNN with Linear Attention and LoRA achieves the highest R2 (0.95), lowest RMSE (6.84), CSI of 95.79, POD of 98.88, and FAR of 3.15.The configuration consistently outperforms the baseline models and combines efficient attention with parameter-efficient tuning.
- Ablation study: Removing Linear Attention significantly degrades performance, underscoring the complementary roles of temporal and spatial attention mechanisms.The ablation result is particularly evident in variants relying solely on GAT layers.
- Ablation study: The 12-head Multi-Head Attention variant reaches R2 (0.92), while the Adapter Layer variant records the lowest MAPE (4.64).These variants demonstrate strong performance and different trade-offs across detection accuracy, false alarms, feature interaction, and generalization.
- Feature analysis: Adding five key external factors raises R2 from 0.87 to 0.95 while reducing both MAE and RMSE.The analysis combines LIME and LRP with traffic and temporal features to assess contextual contributions.
- Qualitative evaluation: The visual analyses show agreement between actual and predicted concentrations, training and validation loss curves, and station-level forecasts across the network.The reported station examples include short-term fluctuations, sharp pollutant spikes, and rapid variations at Station 7.
7. Conclusion
SA-GNN-LoRA combines spatial attention, cluster-specific temporal modeling, and mobile sensing to forecast PM2.5 along a 14 km Surat route. It outperformed traditional baselines in short-term prediction.
- SA-GNN-LoRA uses mobile sensing data from 53 runs along a 14 km urban route in Surat, India.
- The model captures spatial dependencies with two Graph Attention Networks and temporal dependencies with a hybrid 1D-CNN and GRU module.
- Cluster-based grouping assigns monitoring points with similar pollution dynamics to dedicated GRU encoders for localized temporal modeling.
- The model uses 20 historical time steps covering 60 seconds to predict 8 future steps covering 24 seconds.
- RMSE 6.8, SA-GNN-LoRA significantly outperformed LSTM, GRU, ANN, and RNN baselines.