An efficient method for building a database of diatom populations for drowning site inference using a deep learning algorithm