CatBoost Leads Chronologically Validated Air Quality Index Prediction Across Indian Cities

Air quality prediction in India has gained a new benchmark. Researchers at Vellore Institute of Technology have built an ensemble tree-based machine learning framework that predicts the Air Quality Index (AQI) from data collected by India’s Central Pollution Control Board, and their results show that careful handling of missing data and strict time-ordered validation can push prediction accuracy to an R² of 0.96. The study, published in Frontiers in Environmental Science, was conducted by Pulagam Maheswari Devi and Sumathy Subramanian.

The authors frame the work against rapid urbanisation and industrialisation, which have made air pollution a critical global health and environmental concern and made accurate AQI prediction crucial for effective prevention and control. They identify two persistent weaknesses in existing AQI prediction approaches: most lack temporal-based chronological splitting, which helps address bias in the data, and existing methods are inadequate in handling missing data prior to imputation.

The proposed framework integrates pre-processing and AQI prediction on the Central Pollution Control Board dataset. Before any modelling, the dataset was formally analysed using statistical mechanisms that classify missingness as MCAR (missing completely at random), MAR (missing at random), or MNAR (missing not at random). The researchers then applied imputation methods including mean substitution, k-nearest neighbours (KNN), MICE, GAIN, and linear interpolation, and ensured the statistical validity of the imputation.

Feature engineering formed the second pillar of the pipeline. Starting from 12 original features, the team expanded the feature set to 24 by adding rolling, lag, and temporal features. These engineered variables give the tree-based models additional temporal context derived from the pollution record itself.

For prediction, the study benchmarked four ensemble tree-based models: CatBoost, XGBoost, LightGBM, and Random Forest. The data were split 80:20 in strictly chronological order, meaning the models were trained on earlier observations and tested on later ones. Evaluation used three standard regression metrics: mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R²).

Among the four models, CatBoost achieved the strongest overall performance, with an MAE of 10.94, an RMSE of 16.93, and an R² of 0.96 on the held-out chronological test set.

The analysis also included a season-stratified evaluation, breaking performance down by time of year rather than pooling all conditions. This revealed that summer is the most predictable season, with an RMSE of 11.25 and an R² of 0.92.

According to the authors, the integration of imputation methods and the adoption of strict chronological validation for AQI prediction establishes a temporally validated benchmark across multiple cities in India. That combination distinguishes this work from earlier AQI studies that did not use chronological splitting or characterise missingness before imputation.

The authors state that the framework enables policymakers to implement season-specific emission controls, equips the public with timely health awareness, and supports environmental agencies in monitoring and regulating pollution. The reported metrics describe this CPCB dataset and benchmark, and the final formatted version of the article is forthcoming; the authors note that all claims are their own and do not necessarily represent those of their affiliated organisations or the publisher.

Subject of Research: Ensemble tree-based machine learning models for data-driven AQI prediction

Article Title: Ensemble tree-based machine learning models for data-driven AQI prediction

Article References: Pulagam, M. D., & Subramanian, S. (2026). Ensemble tree-based machine learning models for data-driven AQI prediction. Frontiers in Environmental Science, 14, Article 1952221. https://doi.org/10.3389/fenvs.2026.1952221

Image Credits: AI Generated

DOI: 10.3389/fenvs.2026.1952221

Keywords: Ensemble, tree-based, machine, learning, models, data-driven, prediction, scientific research