of which were not represented as inputs in these models. The line of poor
performance of the PM 2 .5 model at the northern tip of Manhattan correlates
well with a band of heavier taxi usage, as seen in Figure 5, and is also close to
where Interstate 95 cuts across the island. One possible explanation is that,
though taxi traffic does increase in that area, the increase is not sufficient to
encompass the huge increase in PM2 .5 from the 195 traffic. This is supported
by the lack of similar correlation between heavy taxi traffic and poor PM 2 .5
model performance in the southern part of the island, suggesting that the
increased error is not caused by overreliance on the taxi data.
Though the average R2 of the SO 2 models is similar to the error in computational models at similar grid sizes, it is the lowest of the pollutants modeled
here. SO 2 is also the only pollutant in this study that is not associated with
modern, gasoline-fueled passenger cars like taxis. SO 2 is emitted by other
vehicles found in cities, including very old passenger cars, diesel vehicles, and
industrial vehicles like buses and trucks. In many cases, these vehicles and
taxis are incentivized to avoid each other either spatially or temporally. For
example, taxi passengers are less likely to request a trip that mirrors one of
the predefined routes city buses take, and delivery trucks are likely to avoid
the congested roads of the urban center at their peak hours, which are exactly
when there are the most taxis. Therefore, the S0 2-emitting vehicles would
be expected to have different traffic patterns than taxis. The performance
of this taxi-based S02 models suggests that while these traffic patterns are
indeed sufficiently different than taxi patterns that taxis make an ifiperfect
substitute, they remain a functional proxy.
In general, the accuracy models suggest that this approach is an effective
and efficient one. Not only do the models predict all the pollutants as well
or better than computational models, they are relatively quick to produce
and run, even on an old laptop computer; the results from time trials on a
2013 Dell OptiPlex 9020 quad-core processor running at 3.1 GHz are shown
in Table 7. It would be possible to use the model in real-time to create
small-scale pollution predictions.