The trouble with making a higher res image is that when you make that lat/lon buckets smaller, you have fewer samples in each one, so the image gets noisier. For the best possible image you'd want to download all the years of data.

The process of making it was quite simple. I zeroed a 2d array of integers, then took all the pickup/dropoff points and incremented the nearest cell. The pixel values are based on the logarithm of the counts, since otherwise everything outside midtown would be pretty much black.

There are some artifacts, like the thin vertical line down the east river. I think that was because of how the data was rounded, i.e. the number of unique longitude values that map to a certain image column.

I wrote this myself with a few hundred lines of C++, though I'm sure there's GIS software out there that will do all this for you with a few clicks.