Learning from Observations: The Case for a New Generation of Land Surface Models