Management of missing air pollution data within urban environments using machine learning regressions: a case study...

Machine learning models are increasingly used in air pollution research, yet their application to long-term reconstruction of missing air-quality records remains limited. Pollutants such as PM_2.5, PM₁₀, O₃, NO₂, SO₂, and CO pose major environmental challenges in megacities like Delhi, but accurate assessment requires long, complete, and spatially distributed monitoring records that are often unavailable. Delhi operated only four stations in 2014, expanding to 45 by 2024, producing fragmented datasets that hinder trend analysis and policy evaluation. This study develops an iterative, multi-model machine learning workflow to reconstruct missing daily pollution data for all six pollutants across 45 stations from 2014-2024. For each missing value, the algorithm identifies the four most correlated stations-assumed to share similar environmental conditions-and uses their observations as predictors in MATLAB's Regression Learner to evaluate 35 ML models. The best-performing model is selected based on reconstruction accuracy, and the procedure iterates until all gaps are filled. Results show that models such as Fine Tree, Bagged Trees, Optimizable Ensemble, Fine Gaussian SVM, Rational Quadratic, and Exponential kernels consistently outperform multilinear regression. The reconstructed datasets enable computation of ensemble-mean pollutant records, yielding a more realistic and bias-corrected representation of Delhi's air-quality evolution, which indicates modest improvement over the decade. Validation through artificial-gap experiments across pollutants, stations, and years demonstrates good performance under the tested conditions for PM_2.5 and other regionally driven pollutants, with reduced accuracy for species dominated by local variability. These findings support the usefulness of the approach for long-term regional reconstruction while also highlighting its limitations for pollutants with strong local emission signatures.