While ozone design values have decreased since 2000, the values measured in Denton Airport South (DEN), an exurban region in the northwest tip of the Dallas-Fort Worth (DFW) metroplex, remains above those measured in Dallas Hinton (DAL) and Fort Worth Northwest (FWNW), two extremely urbanized regions; in addition, all three sites remained in nonattainment of National Ambient Air Quality Standards (NAAQS) ozone despite reductions in measured NOx and CO concentrations. The region's inability to achieve ozone attainment is tied to its concentration of total non-methane organic compounds (TNMOC). The mean concentration of TNMOC measured at DAL, FWNW, and DEN between 2000 and 2018 were 67.4 ± 1.51 ppb-C, 89.31 ± 2.12 ppb-C, and 220.69 ± 10.36 ppb-C, respectively. Despite being the least urbanized site of the three, the TNMOC concentration measured at DEN was over twice as large as those measured at the other two sites. A factor-based source apportionment analysis using positive matrix factorization technique showed that natural gas was a major contributing source factor to the measured TNMOC concentrations at all three sites and the dominant source factor at DEN. Natural gas accounted for 32%, 40%, and 69% of the measured TNMOC concentration at DAL, FWNW, and DEN, respectively. The Barnett Shale region, an active shale gas region adjacent to DFW, is a massive source of unconventional TNMOC emissions in the region. Also, the ozone formation potential (OFP) of the TNMOC pool in DEN were overwhelmingly dominated by slow-reacting alkanes emitted from natural gas sources. While the air pollutant trends and characteristics of an urban airshed can be determined using long-term ambient air quality measurements, this is difficult in regions with sparse air quality monitoring. To solve the lack in spatial and temporal datasets in non-urban regions, various machine learning (ML) algorithms were used to train a computer cluster to predict air pollutant concentrations. Models built using certain ML algorithms performed significantly better than others in predicting air pollutants. The model built using the random forest (RF) algorithm had the lowest error. The performance of the prediction models was satisfactory when the local emission characteristics at the tested site were like the training site. However, the performance dropped considerably when tested against sites with significantly different emission characteristics or with extremely aggregated data points.