ORCID

https://orcid.org/0009-0002-2770-6751

Date of Award

Spring 2025

Language

English

Embargo Period

4-30-2026

Document Type

Master's Thesis

Degree Name

Master of Science (MS)

College/School/Department

Department of Atmospheric and Environmental Sciences

Program

Atmospheric Science

First Advisor

Scott Miller

Second Advisor

Cheng-Hsuan Lu

Committee Members

Scott Miller, Cheng-Hsuan Lu

Keywords

low-cost sensor, air quality, calibration, mesonet

Subject Categories

Other Chemical Engineering | Other Environmental Sciences

Abstract

The University at Albany designed and manufactured 59 low-cost air quality sensor packages to continuously measure PM₂.₅, CO, O₃, NO₂, and NO at 38 New York State Mesonet (NYSM) sites located in the New York City Metropolitan Area. Prior to use for monitoring, low-cost sensors require calibration to correct for environmental sensitivities. Calibration models can be developed using data collected from co-location periods in which low-cost sensors are installed at sites with Federal Reference Methods and/or Federal Equivalent Methods instruments. In this study, packages were periodically co-located (calibrated) for 18 to 162 days at the New York State Department of Environmental Conservation Queens College site. Two types of calibration models, a multiple linear regression (MLR) and a Random Forest-MLR hybrid model, were initially developed using the “traditional” calibration approach in which individual models are trained for each sensor and package (i.e., 59 CO models). Both models produced accurate concentrations during short-term evaluation (< 2 months), with the hybrid model improving accuracy at low concentrations for O₃, PM₂.₅, NO₂, and NO. Extending evaluation over a 1-year period to reflect realistic network operations revealed drift reducing data quality for NO₂ and O₃, with O₃ developing a negative bias only 3 months after calibration. Periodically recalibrating packages was expensive and ineffective at improving model performance, with most packages failing to be recalibrated prior to the onset of NO₂ and O₃ drift. To supplement the need for recalibration, the Network Calibration Algorithm was developed by training a single MLR ( CO) or hybrid model (O₃, PM₂.₅, NO₂, NO) per pollutant using 16 months of continuous co-located data from two packages permanently deployed at the calibration site. The algorithm successfully increased long-term accuracy and stability, enabling long-term deployments of low-cost sensors without the need for recalibration.

Available for download on Thursday, April 30, 2026

Share

COinS