Thanks for the link, comments and imagico, very useful!

There certainly will be more issues and you’re right about the ‘very different things’. However, all 3 datasets still indicate places with the maximum likelihood of water. With a proper understanding - a nice data fusion should be possible. In any case, I want to analyse the differences pairwise between different dataset after splitting them into areas where different approach can be used to do the analysis. For example:

1. If you can observe water from LANDSAT - probably that will be the most precise estimate (horizontally +/- 15-30m). Then comparing OSM with LANDSAT water centerline probably makes sense. It will be also nice to check how good SRTM looks like in these areas and I certainly won’t use SRTM to define centerlines for these rivers, actually I plan to burn SRTM for the areas where it does not match.

… OpenStreetMap water based on 1m imagery is still better, but we don’t know that. From what I saw, water centerline for wide rivers is not always a centerline in OSM. I’m not sure if there is an agreement on that, most probably if differs from place to place. For modeling applications it would be actually good to have a geometric centerline and then thalweg.

1. For small rivers located in hilly areas SRTM might be the best guess, but again, if OSM is edited manually - theoretically it should be much better, especially when there is a riparian vegetation present (SRTM will be noisy) :).

Based on flow accumulation area it will be possible to estimate approximately how big the river ~actually is. Maybe based on the landuse can estimate the SRTM noise.

1. For urban areas it will be a mess, probably LANDSAT (for large waterbodies) and then OSM is the best guess.

About applicability to other areas, I expect the following dataset to come out somewhere soon: Peckel, 2014, so my focus is not to “detect water in the best way using LANDSAT” but rather to focus on the methodology to compare different vector/raster river datasets given watermask based on LANDSAT.