<p>Air pollution monitoring is essential for assessing air quality in urban areas, supporting public health prevention and guiding targeted interventions. Pollutant data are typically collected using fixed-location or mobile sensors. While high-precision sensors offer accurate measurements, their production requires advanced expertise and dedicated hardware, often implying substantial costs. Consequently, low-cost sensors are widely adopted in urban environments, despite their limited estimation capabilities. To address this limitation, many Machine Learning (ML) algorithms have been employed to post-process sensor data to better align them with high-precision sensor measurements. Although state-of-the-art ML models often achieve high accuracy, they are frequently opaque, limiting transparency and hindering the interpretation of model decisions, which is a critical aspect in environmental monitoring. In this study, we evaluate and compare both black-box and interpretable approaches for improving the estimation of PM10 concentrations (i.e., airborne particulate matter) in Trieste, Friuli-Venezia Giulia, Italy. In particular, we analyze the effectiveness of Symbolic Regression (SR), an evolutionary algorithm capable of generating compact and potentially interpretable mathematical expressions that explicitly describe relationships between atmospheric features and PM10 levels. Results show that SR provides interpretable models that also exhibit better accuracy compared to the other glass-box strategies in the calibration environment, while offering insights into the influence of environmental variables on air quality. On the other hand, in our specific setting, SR suffers from transferability issues into low-pollution areas. Our study presents the trade-off levels associated with using black-box and glass-box approaches, with particular emphasis on the advantages and limitations of SR compared to other strategies in the evaluated environments.</p>