A spatiotemporal machine learning framework for high-resolution PM_2.5 estimation using reconstructed satellite...

Accurate monitoring of fine particulate matter (PM_2.5) in North-East India is severely constrained by sparse ground-based observations, complex terrain, high humidity and persistent cloud cover, particularly during the monsoon season, when satellite aerosol retrievals are frequently missing or unreliable. To address these limitations, this study develops an integrated multi-sensor satellite and machine learning (ML) framework for estimating daily surface PM_2.5 concentrations at 1 km spatial resolution over the Guwahati-Kamrup region. A key contribution of this study is the integration of reconstructed multi-sensor aerosol optical depth within a high-resolution PM_2.5 estimation framework tailored for a cloud-dominated, data-scarce region of North-East India. The reconstructed aerosol optical depth is combined with meteorological variables, land-surface characteristics and limited CPCB ground measurements within a unified modelling architecture. Multiple ML models are evaluated, and a multi-layer perceptron is identified as the best-performing estimator, achieving a coefficient of determination (R²) of 0.82 with substantially reduced prediction errors compared to models driven by raw satellite inputs. The resulting PM_2.5 fields reveal pronounced urban-rural gradients, persistent transport-aligned hotspots and strong seasonal modulation consistent with the influence of boundary-layer dynamics and wet scavenging processes. Winter and post-monsoon periods exhibit the highest pollution accumulation, while monsoon conditions consistently suppress near-surface concentrations. Long-term hotspot analysis and statistical projections indicate a gradual intensification of urban PM_2.5 exposure under observed emission trajectories.