No, in general you can not treat time series data as cross sectional.
Furthermore, you can not pool different levels together and treat them as one random sampling. This, at the very least, will give you a biased estimate (except for when the regressor you are looking at is somehow uncorrelated to the identity of the city, which doesn't make sense).
Think about it this way: There is a distinct effect for the samples from the same city: The actual influence of being from the same city.
With pooling, you eliminate the portion our of your model which would take this effect into account. So now the rest of your model will inherit this variation in the data (Your model now assumes there is no difference between the samples as it comes to city). However this biases your estimates: The estimate now includes the true effect (whatever you are researching) as well as the inherited effect which is actually just from sampling from a certain city.
You can no longer distinguish between cities - and any influence that comes from drawing from a certain city will now be attributed to your other factors - if those factors are also correlated with being from a certain city. Which they most likely are.
You'll be introducing two errors. First the bias from just not acknowledging the existence of any effect of drawing from the same city, second the problems you get with just OLSssing over a time-series.
Samples from a time series, for example city A, are serially correlated as the previous year will have some sort of relationship with the current year. But this can introduce problems such as autocorrelation, which may make your assumption of exogenity invalid. In other words your error terms may end up (contemporarily) correlated with your regressor / explanatory variables. Then, once again, your estimate is no longer correct.