Currently bilinear regridding gives unexpected results for source data which is of boolean or unsigned integer type due to the way that type promotion is currently handled here

An easy fix is the following:

dtype = np.promote_types(dtype, 'float16')