The data comes from different sources. At some point they are combined as bands of a single image, then exported as raster data, then converted to the saved numpy arrays with:

src = rasterio.open(filename)

arr = src.read()

with open(save_name, 'wb') as f:

np.save(f, arr)

I would guess that uint32 is the default somewhere in that pipeline. Most of the variables (image data, rainfall, wind...) are all positive. Temperature can go negative, but in Kenya, even the minimum monthly average temp is ~9 degrees I think. Have you hit issues with the choice of datatype?