Modelled energy-balance fluxes at the Kc1 tower, scored against eddy-covariance observations across SUEWS releases.
Reference observations: Ward et al. (2016). Each release is run on the same forcing and observations over the same period, with a config valid in its own schema; only the SUEWS version changes. Every release’s statistics reproduce a byte-identical fingerprint on re-run.
Energy-balance error is unchanged across all six benchmarked releases (2025.7 → 2026.6); the largest shift in any flux is 0.03 W m−2 (QE). Sensible heat QH, the residual, carries the largest absolute error throughout.
Full-period mean absolute error (W m⁻²) by release — the at-a-glance regression view. Each row is a flux; cell shade tracks error size. A regression would appear as a row darkening left-to-right; instead every row is a flat band — no drift across releases.
Mean absolute error (W m⁻²), full period (2011–2013). Cell shade increases with error; uniform shading across a row means the flux’s error is unchanged across releases. QH (the residual) carries the largest error throughout.
Pick any two releases to see how each error metric moved between them. Green means the error shrank, amber means it grew; muted means the change is within the regression tolerance (0.05 W m⁻² / 0.05 K).
Values are mean absolute error against the Ward et al. (2016) observations; Δ is To − From, so a negative Δ (green) is an improvement and a positive Δ (amber) a regression. Energy-balance fluxes in W m⁻², RSL air temperature in K. Most release pairs are identical — the only non-trivial shift is the near-surface RSL profile at 2026.1.28 → 2026.4.3.
Everything above holds the releases up against each other. From here down the lens narrows to one release — its full energy balance, the seasonal breakdown, and the near-surface air-temperature profile. Pick the release to dissect:
Bar length and value are the mean absolute error (MAE) against the Ward et al. (2016) observations, full period (2011–2013), with the bar’s shade deepening as the error grows. MBE is the mean bias (model − observation); n is the count of valid hourly pairs. All fluxes in W m⁻².
Mean absolute error (W m⁻²) by flux and season. Cell shade increases with error (deeper = larger MAE). DJF / MAM / JJA / SON are winter / spring / summer / autumn; QH dominates the error budget, peaking in summer.
Modelled air temperature interpolated from the RSL vertical profile to the tower heights, scored against observations over two 2013 windows (Jan–Mar, Apr–Jun); bar length and shade both grow with the error. MAE/MBE in K; n is valid hourly pairs. Across releases the profile error is steady (~0.3–0.4 K) through 2026.1 and rises slightly at 2026.4 (6.5 m: 0.42 → 0.54 K). That shift is isolated to the diagnosed near-surface profile and traces to a 2026.4 change in the model backend, not the surface physics — the energy-balance fluxes are unchanged.
The evaluation observations and forcing are access-restricted; only the derived statistics on this page are published. Reproduction data is archived in the Zenodo reproducibility record, with the full reproduction recipe in REPRODUCE.md. See the SUEWS documentation for the model description and benchmarking methodology.