at 640x512@30hz and radiometric you are already way at the top.
There are higher resolution uncooled detectors, only a few. The very highest resolution uncooled sensor is a
TWV1912 made by Fairchild/BAE - and to my knowledge there is one available camera build around it by
Sierra Olympic, the Vayu HD. While I was able to hold myself back from asking them about the price, I am expecting it to be in the 30k range.
But implementation will be much more difficult, eventhough they market it for airborne use, it does not produce the radiometric flir images that some photogrammetry apps now support. I would ask for a demo, if you are in such a position where they answer your and workout the software workflow with help from the developers of whatever software you are currently using, if they offer specelized solution.
Any kind of airborne detector that is cooled, can have even higher resolutions, these are normally build for police helicopters, or military aircraft and focus on target tracking, range indication and high zoom levels and less radiometric data for mosaics.
Is superresolution the thing you talk about? It normally works on a camera that is in about the same place and takes 4 images in short succession and using the hand wiggle or wind to move the sensor by a fraction of a pixel resulting in a single high resolution image when combined. In photogrammetry you have a point cloud, where every point is based on data gathered, to get better resolution, just gather more points by flying over the same section from a different angle and direction. In theory there are many ways to interpolate more data points based on the seemingly random intervals you have with your reconstruction when making up a wavelet direction. But I have yet to read about it in detail. I am not sure how advanced software is in the field, but using a visible light camera to create the pointcloud gives you more data points due to the high resolution, if possible a solid model could be reconstructed and the thermal pixels projected onto them. Giving you a "3D MSX" dataset.