Spatial assessment of PM10 hotspots using Random Forest, K-Nearest Neighbour and Naïve Bayes

Abstract

Spatial modelling and analysis can assist in improving the decision-making process of mitigating bad air quality. One of Malaysia's most harmful air pollutants is particulate matter (PM), which has been used to denote the Air Pollutant Index (API) for over 20 years. The spatial prediction of particulate matter less than 10 μm (PM10) hotspots is crucial to be assessed as it adversely affects human health and the environment. Advanced prediction of PM10 hotspots can ensure adequate preparedness for air quality management and minimize its effects. The PM10 data acquired for 2012–2016 in Malaysia's urbanized and populated state, Selangor, was used for the modelling. The PM10 was modelled using remote sensing data such as elevation, slope, road density, Soil Adjusted Vegetation Index, Normalized difference Vegetation Index, built-up index, land surface temperature, and wind speed. Spatial modelling of the PM10 hotspot was done using a Naïve Bayes (NB), Random Forest (RF), and K-Nearest Neighbour (KNN) algorithms. Results revealed a good prediction of PM10 hotspot with model performance for KNN, RF, and NB are in terms of specificity: 0.98, 0.99, 0.92; precision: 0.98, 0.99, 0,92; recall: 0.94, 0.98, 0.91; and the overall accuracy is 0.96, 0.98, 0.91, respectively. The PM10 hotspot map produced by the RF model indicates that urbanized and industrialized areas have high PM10 concentration, which characterizes the harmful effect of air pollutants in urbanized regions. These models can therefore be used for spatial assessment of PM supporting the Sustainable Development Goal (SDG) 11 on Sustainable Cities and Communities.