What's happening there is you are underexposing the shadows so much to preserve highlight detail (a difficult balance in a night scene) that you end up with a slight grain in the shadows - even that is not bad at all and you would never see that very light grain in a print or web-sized image (the two mediums most likely to be used but YMMV). There is also negligible color noise in that image. You can see that the properly exposed areas have essentially zero noise, as expected. Personally I would not NR that as you can see from your examples how horribly it destroys the detail, especially in the distance where there is precious little detail to begin with - in some cases it removes the windows from buildings almost entirely, and makes the whole image look like a watercolor painting unless viewed at web size, where the grain wouldn't show either. Entirely up to you of course, but I would personally suggest not using such heavy handed NR on an image with only a slight grain to it and almost no color noise. If you like the smoothed-over NR look (which is fine), there are also plugins you can get that will analyze the image and apply NR selectively to different areas of the photo (Like NIK DFINE2 which is very good) that you might find useful.
Another thing you can try is if the highlights are going to be blown anyway, overexpose the image a little bit to get the shadow pixels closer to FWC and then bring it back down in post (basically ETTR). You can also stack multiple exposures if you are still unhappy with the noise levels as one way to avoid using such heavy NR.
More good news is that I do not think anything is wrong with your drone at all, that sensor is behaving exactly as it should in that scenario.
Twice earlier you said that you can apply NR without destroying detail, which is the main point I was contesting, and it seems we agree now that it does.
