You could also just get a recording from someone who went out to a beautiful-sounding place and recorded several hours, these have existed since before digital sampling techniques, much less machine learning. Looping short samples is pretty much the worst way possible to generate a soundscape.