We collected a mobile‑sensing dataset in Surat, Gujarat, India, containing $PM_{2.5}$ concentrations, temperature, humidity, wind speed, wind direction and land‑use attributes. To cast the spatiotemporal series into a graph, two node‑definition schemes were used: (i) uniform segmentation at 200–400 m intervals; (ii) DBSCAN clustering to adaptively group dense observations. For each node we computed rolling mean and standard deviation of the meteorological variables to capture local short‑term fluctuations. To handle the high‑dimensional input we propose a Spatially Attentive Graph Neural Network (SA‑GNN) for fine‑grained short‑term $PM_{2.5}$ forecasting and hotspot detection. SA‑GNN employs cluster‑specific GRUs for localized temporal dependencies and a Graph Attention Network to learn spatial heterogeneity, jointly modeling rapid changes and complex spatial interactions. Experiments compare SA‑GNN with LSTM, RNN, GRU and ANN baselines; the latter perform well on low‑resolution data but struggle with the volatile urban air‑quality patterns. On our dataset SA‑GNN achieves $R^2 = 0.95$, $RMSE = 6.8$, $MAE = 4.2\ \mu g/m^3$, surpassing all baselines. The combination of spatial clustering and adaptive attention markedly improves forecast accuracy, enabling real‑time fine‑grained monitoring, personalized exposure assessment and timely health alerts.
Review: By coupling localized temporal modeling with spatial attention, the approach overcomes the limitations of conventional sequence models on high‑resolution air‑quality data, offering a practical pathway for smarter urban health management.