Can wetlands be an effective option to reduce the particulate matter pollution in the air in urban spaces?

The problem of particulate matter concentration, especially PM2.5 and PM10, is a major concern for all million-plus cities worldwide. Depending upon the scale of implementation of the available technologies and emission reduction policies, the levels of these particulate matter vary over time and space. In this study, we evaluated the effect of wetlands in reducing the concentration of PM2.5 in the air over one of the highest polluted cities in the world, New Delhi, in the Indian region. The PM2.5 was modeled considering the distance-to-wetlands - from the in-situ pollution monitoring stations - as an influential factor, including other determining environmental covariates such as meteorological parameters, atmospheric optical parameters, surface greenness, and land use and land cover (LULC) type (Experiment Set 1). We also conducted a similar experimental setup to build a predictive model excluding the variable distance-to-wetlands (Experiment Set 2). The data of PM2.5 from 21 monitoring stations and all other covariates corresponding to these stations were collected at a daily temporal scale from January 2016 to August 2019. Due to the complexity of the relationships as well as the distribution patterns of all independent variables, a series of machine learning (ML) and artificial intelligence (AI) based analytics, such as random forest (RF), gradient boosting (GB), support vector machine (SVM), and artificial neural net (ANN) regression, were used to model the PM2.5 at monthly and seasonal time scale spatially. All these AI/ML models were trained on 70% of the observations through a random selection. The remaining 30% of the data was used for evaluating the models’ performance. The performances of the models were then compared for both sets of experimental setups. The statistics for model performance diagnostic shows a higher R² for RF-regression than other AI/ML regression models at the training stage under both sets of experimental setups (R² ≥ 0.69 for Experiment Set 1; R² ≤ 0.66 for Experiment Set 2 under RF regression; R² ≤ 0.64 for other models under both experimental setups). The variable influence score (VIS) under RF-regression manifests that the proximity of wetlands is important than the variation of precipitation and LULC type (VIS: 4.12% for distance-to-wetlands, VIS: 2.61%, and 0.24% for precipitation and LULC type, respectively). The predictability of the RF-regression model, while evaluated with the test data, shows R² ~ 0.66, with RMSE of 80.4 µg m^-3 for the Experimental Set 1, and R² ~ 0.63 with RMSE of 83.7 µg m^-3 for the Experimental Set 2. It was noticed from the analysis that within a 1000 m buffer distance from the wetlands, the concentration of PM2.5 remains relatively lower than a distance greater than 1000 m. Such difference is benign, yet the key factor behind such a benign effect is related to the variation in surface area of the wetlands. Larger wetlands may have a distant impact on keeping the PM2.5 low. The study, thus, concludes that restoring the wetlands might be one of the practical solutions to keep the PM level within the ambit of NAAQ standards.