Mohammad Javad Nematollahi

Meteorological observations are fundamental to environmental modeling, yet data gaps frequently compromise their utility. This study benchmarked five machine-learning algorithms (Random Forest, Extra Trees, Gradient Boosting, Histogram-based Gradient Boosting, and Ridge regression) for hourly temporal imputation of five key meteorological variables—temperature, humidity, dew point, atmospheric pressure, and wind speed—across 17 airport-based synoptic stations in Iran. A comprehensive feature-engineering approach generated 43 predictors capturing calendar, cyclical, continuous composite, binary regime, and cross-variable patterns. Performance was assessed using R², RMSE, and MAE with time-series cross-validation, bootstrap resampling, split and Mondrian conformal prediction, three robustness stress tests, and comparison against persistence and climatological baselines. Results revealed a striking performance dichotomy. Thermodynamically coupled variables achieved exceptional accuracy, with mean cross-validated R² values of 0.994 ± 0.003 (temperature), 0.997 ± 0.001 (humidity), and 0.988 ± 0.009 (dew point), with correspondingly low RMSEs (0.74 ± 0.22 °C, 1.10 ± 0.21%, and 0.60 ± 0.19 °C). Feature-ablation confirmed that this skill was driven primarily by inter-variable coupling, with the four cross-variable predictors accounting for 96.3–100% of importance; removing them caused the dew point R² to collapse to 0.375. In contrast, dynamically controlled variables exhibited substantially lower skill: atmospheric pressure achieved a mean R² of 0.44 ± 0.25 (RMSE 3.87 ± 0.86 hPa), while wind speed performed poorest (mean R² 0.13 ± 0.22, RMSE 2.20 ± 0.50 m/s), with negative R² at several stations and no significant improvement over persistence. Their diffuse feature importance profiles and wider conformal prediction intervals indicated that temporal and cross-variable features alone cannot capture advective and synoptic-scale influences. Pressure proved less reconstructable than the thermodynamically coupled variables, reflecting its dependence on large-scale atmospheric dynamics rather than on local site conditions. This study demonstrates that ensemble tree-based models excel at reconstructing variables with strong inter-variable coupling but require spatial and physics-informed integration for dynamic parameters in regions with complex synoptic-scale dynamics.

Feature engineering (FE) is widely employed to enhance machine learning (ML) performance, yet its effectiveness across diverse meteorological variables and climatic regions remains poorly understood. This study systematically evaluates fifteen FE strategies—including temporal, cyclic, lagged, rolling, and interaction features—using a Random Forest model to predict eight meteorological variables (temperature, humidity, pressure, wind speed, clouds, precipitation, T-max, and T-min) at daily resolution across 1,634 synoptic stations in Asia (2000–2024). Our results demonstrate that FE is not universally beneficial: statistically significant improvements (p < 0.05) in R² occurred in only 42.7% of station-target combinations. Scenario 10 (2-way interactions) was the most frequently optimal strategy (16.0%), followed by Scenario 2 (temporal features; 13.1%). Pressure showed exceptional responsiveness to comprehensive feature sets (Scenario 15), achieving continent-wide RMSE reductions of 41–42%. Temperature and clouds benefited most from interaction features in tropical regions, while simpler temporal features sufficed at higher latitudes. Humidity and wind speed required minimal FE. Critically, for extreme temperatures (T-max/T-min), the control scenario (no FE) outperformed engineered alternatives in 37% and 49% of cases, respectively, with feature augmentation actively degrading performance. Spatial analysis revealed a distinct latitudinal gradient, wherein the utility of complex features diminishes from tropics northward. These findings provide conclusive evidence that optimal FE strategies are neither universal nor intuitive, but rather demand a bespoke, variable-specific, and geographically-aware approach. This study offers a comprehensive, continent-scale roadmap for strategic feature engineering in meteorological ML, identifying not only what works, but crucially, what does not, and where.

Accurate prediction of Carbon monoxide (CO) concentration from co-pollutants and meteorological variables is challenging due to predictor collinearity, nonlinear interactions, and measurement variability, which often lead to unstable generalization in practical air-quality applications. This study proposes a leakage-free out-of-fold (OOF) stacking framework using ElasticNetCV as a sparse meta-learner to improve robustness and reliability in CO regression. In the proposed approach, meta-features are constructed exclusively from cross-validated base-model predictions, thereby preventing information leakage during ensemble training. At the same time, ElasticNet regularization controls ensemble complexity and prunes redundant learners. The method is evaluated on a multi-pollutant and meteorological dataset including O3, NO2, NO, temperature, dew point, humidity, wind speed, and air Pressure as predictors. Beyond conventional accuracy metrics (MAE, RMSE, and R2), we assess generalization through train–test gap analysis, learning curves, meta-weight stability evaluation, and paired statistical tests. Results demonstrate that the proposed OOF stacking approach consistently reduces generalization gaps, mitigates performance variability, and achieves statistically significant improvements over individual base models. These findings indicate that OOF stacking with ElasticNet meta-learning behaves as a principled regularization and model-selection mechanism, improving the stability and reliability of generalization for CO prediction in heterogeneous environmental datasets.

Wastewater treatment plants (WWTPs) are routes through which microplastics (MPs) enter the environment. This research assesses MP occurrence in 2 urban and 5 industrial WWTPs, in a city of 1.6 M inhabitants (Tabriz, Iran). The sampling schedule helps quantify the city’s MP release into surface water. One of the urban WWTPs in the study used tertiary treatment, whereas the rest of the plants only had primary and secondary treatments. MPs were not efficiently removed in the urban WWTPs: 255 ± 55 MP/L (influent) and 240 ± 50 MPs/L (effluent). In the industrial WWTPs, there was some elimination of MPs: 246 ± 169 MP/L (influent) and 94 ± 72 MPs/L (effluent). The lowest MP removal, < 10%, corresponds to urban treatment. The influents had > 1–5 mm MP particles at 14% of the total MPs in the incoming sewage, and such fraction was just < 2% of the total MPs in the effluents. In contrast, smaller MPs, especially those < 250 μm (the smallest detectable and confirmed MP was 50 μm), were the most abundant MP fraction in the treated effluents. Fibres, black MPs and particles ranging from 100 μm to < 250 μm prevailed in influents and effluents of both urban and industrial WWTPs. In the wastewater of both types of plants, the MPs were mainly composed of polyamide and polycarbonate. This research confirms that urban and industrial wastewater are important MP contributors to receiving water and calls for improvements in wastewater treatments to decrease the uncontrolled entry of MPs into the environment.

Air pollution is a significant challenge in metropolitan areas, where increasing amounts of air pollutants threaten public health and environmental safety. The present study aims to forecast the concentrations of various air pollutants, including CO, O₃, NO₂, SO₂, PM₁₀, and PM_2.5, from 2013 to 2023 in the Tehran megacity, Iran, via deep learning (DL) models and evaluate their effectiveness over conventional machine learning (ML) methods. Key driving variables, including temperature, relative humidity, dew point, wind speed, and air pressure, were considered. R-squared (R²), root-mean-square error (RMSE), mean absolute error (MAE), and mean-square error (MSE) were used to assess and compare the models. This research demonstrated that DL models typically outperform ML models in forecasting air pollution. Gated recurrent units (GRUs), fully connected neural networks (FCNNs), and convolutional neural networks (CNNs) recorded R² and MSE values of 0.5971 and 42.11 for CO, 0.7873 and 171.40 for O₃, and 0.4954 and 25.17 for SO₂, respectively. Consequently, the FCNN and GRU presented remarkable performance in predicting NO₂ (R² = 0.6476 and MSE = 75.16), PM₁₀ (R² = 0.8712 and MSE = 45.11), and PM_2.5 (R² = 0.9276 and MSE = 58.12) concentrations. In terms of operational speed, the FCNN model exhibited the most efficiency, with a minimum and maximum runtime of 13 and 28 s, respectively. The feature importance analysis suggested that CO, O₃ and NO₂, SO₂ and PM₁₀, and PM_2.5 are most affected by temperature, humidity, PM_2.5, and PM₁₀, respectively. Thus, temperature and humidity were the primary factors affecting the variability in pollutant concentrations. The conclusions confirm that the DL models achieve significant accuracy and serve as essential instruments for managing air pollution, providing practical insights for decision-makers to adopt efficient air quality control strategies.

As a significant global concern, air pollution triggers enormous challenges in public health and ecological sustainability, necessitating the development of precise algorithms to forecast and mitigate its impacts, which has led to the development of many machine learning (ML)-based models for predicting air quality. Meanwhile, overfitting is a prevalent issue with ML algorithms that decreases their efficacy and generalizability. The present investigation, using an extensive collection of data from 16 sensors in Tehran, Iran, from 2013 to 2023, focuses on applying the Least Absolute Shrinkage and Selection Operator (Lasso) regularisation technique to enhance the forecasting precision of ambient air pollutants concentration models, including particulate matter (PM_2.5 and PM₁₀), CO, NO₂, SO₂, and O₃ while decreasing overfitting. The outputs were compared using the R-squared (R²), mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and normalised mean square error (NMSE) indices. Despite the preliminary findings revealing that Lasso dramatically enhances model reliability by decreasing overfitting and determining key attributes, the model’s performance in predicting gaseous pollutants against PM remained unsatisfactory (R²_PM2.5 = 0.80, R²_PM10 = 0.75, R²_CO = 0.45, R²_NO2 = 0.55, R²_SO2 = 0.65, and R²_O3 = 0.35). The minimal degree of missing data presumably explained the strong performance of the PM model, while the high dynamism of gases and their chemical interactions, in conjunction with the inherent characteristics of the model, were the primary factors contributing to the poor performance of the model. Simultaneously, the successful implementation of the Lasso regularisation approach in mitigating overfitting and selecting more important features makes it highly suggested for application in air quality forecasting models.