Improving latent heat flux prediction via enhanced cross-variable interaction and temporal dependence learning