Land Surface Model
A land-surface model computes the state of the ground and vegetation and returns the fluxes the atmosphere needs: sensible heat, latent heat, upward longwave, and momentum. It carries soil temperature and moisture in several layers, some representation of vegetation, and usually snow.
Its central job is the partitioning of available energy between sensible and latent heat — the Bowen ratio. Everything else the LSM does is in service of getting that split right, and the split has a larger effect on near-surface climate than most choices available elsewhere in a model. Energy into evaporation produces a cool moist boundary layer that stays shallow; energy into sensible heat produces a warm dry one that grows deep. The same net radiation gives different weather depending on which way the LSM sends it.
Why soil moisture propagates so far
Soil moisture controls the partitioning, and it has memory — weeks to months, far longer than anything in the atmosphere. That combination makes it the main source of land-driven predictability at subseasonal range, and the main source of slowly accumulating error.
The feedback runs in a loop that can go either way. Dry soil sends energy into sensible heat, warming and deepening the boundary layer, reducing relative humidity, suppressing convection, so less rain falls and the soil gets drier. This is the mechanism behind drought amplification and behind the tendency of regional models to intensify heatwaves. Whether the loop closes positively depends on the moisture regime, and models disagree about where the transition sits.
The practical consequence is initialisation. Soil moisture is not well observed, so it is usually taken from a reanalysis or from a spin-up run. A regional simulation started with soil moisture inconsistent with its own land-surface scheme will drift for weeks while the soil adjusts, and that drift is easy to mistake for a model bias. In downscaling work, spinning up the land surface properly is one of the highest-return steps available, and skipping it has cost me more time than it would have taken.
What separates schemes
Complexity varies enormously — from a few soil layers and bulk vegetation to explicit carbon cycling, dynamic vegetation, groundwater, and urban tiles. Added complexity does not reliably improve simulated fluxes, because it adds parameters that are not well constrained by observations. A simpler scheme with well-chosen parameters frequently outperforms a more sophisticated one used at default settings.
For urban work the standard schemes are inadequate on their own, since they have no representation of buildings — hence urban canopy models and, at finer scales, resolving the urban surface directly.
See also: Noah LSM for the common WRF choice, surface-layer schemes for how these fluxes reach the atmosphere, evapotranspiration for the latent heat term, and soil moisture.