Calibration Examples: Controlled Basins to Regional Fleets

Florida controlled basin (USGS 02297600)

Calibration station: 02297600 (gage channel 2).

Use the figure controls to compare calibration and verification performance at the same station and to see the Morris screening outcome that selected the calibration subset.

Interactive figures: switch between calibration, verification, and sensitivity outputs. Hydrograph plots include the run-stage metrics on the figure (as in the publication artifacts).

CalibrationVerificationSensitivityDailyMonthlyInit pool bestGlobal best

Calibration global best (daily)

• Calibration: compares the pre-optimization initialization pool and the calibrated global best for the scored calibration window.
• Verification: shows the independent holdout window results (no overlap with the scored calibration period).
• Sensitivity: Morris tornado plot used to rank candidate parameters and pick the trimmed calibration subset.

Sensitivity analysis (Morris)

Morris screening runs a fixed evaluation budget and reports μ* (magnitude of the elementary effect) and σ (interaction / nonlinearity signal). The table below lists the top drivers by μ*.

Sensitive parameters (top 8 by μ*): surq_lag, perco, cn3_swf, alpha_bf, dep_wt, spec_yld, flo_min, dp_es.

Parameters selected for calibration in this example: 8 (the Morris-trimmed subset above). The complete candidate parameter template and bounds are documented on /calibration-methods.

Calibration and verification results

Metrics are reported for daily and monthly time steps (NSE, KGE, and PBIAS). Values below are pulled from the exported run summary tables used in the publication workflow evaluation.

Illinois controlled basin (USGS 05536265)

Calibration station: 05536265 (gage channel 25).

Use the figure controls to compare calibration and verification performance at the same station and to see the Morris screening outcome that selected the calibration subset.

Interactive figures: switch between calibration, verification, and sensitivity outputs. Hydrograph plots include the run-stage metrics on the figure (as in the publication artifacts).

CalibrationVerificationSensitivityDailyMonthlyInit pool bestGlobal best

Calibration global best (daily)

• Calibration: compares the pre-optimization initialization pool and the calibrated global best for the scored calibration window.
• Verification: shows the independent holdout window results (no overlap with the scored calibration period).
• Sensitivity: Morris tornado plot used to rank candidate parameters and pick the trimmed calibration subset.

Sensitivity analysis (Morris)

Morris screening runs a fixed evaluation budget and reports μ* (magnitude of the elementary effect) and σ (interaction / nonlinearity signal). The table below lists the top drivers by μ*.

Sensitive parameters (top 8 by μ*): melt_min, k, cn3_swf, perco, melt_max, urban_cn_c, mann, surq_lag.

Parameters selected for calibration in this example: 8 (the Morris-trimmed subset above). The complete candidate parameter template and bounds are documented on /calibration-methods.

Calibration and verification results

Metrics are reported for daily and monthly time steps (NSE, KGE, and PBIAS). Values below are pulled from the exported run summary tables used in the publication workflow evaluation.

The controlled basins above are single-gauge problems. The west-central Florida fleet — Tampa Bay, Peace, Hillsborough, and Myakka, four platform-generated coupled SWAT+/MODFLOW 6 basin models — is where the calibration methodology scales up: 29k–77k-HRU models, rosters of dozens of USGS gauges, regionalized parameters, and the information-cascade workflow (pooled priors → zero-shot seeding → bounded terminal MMPSO) documented on /calibration-methods. These are living examples: the campaign is running now, and this section is updated as stages complete. The fleet’s home is the Florida nonpoint-source living page, with interactive 3D viewers for each basin (Tampa Bay, Peace, Hillsborough, Myakka).

Tampa Bay: what a pooled regional prior does zero-shot

A parameter distribution pooled from previously calibrated Peace-area models was applied to the 77.6k-HRU Tampa Bay catalog model — a different basin, a different estuary — in one simulation, with no optimizer and no tuning. The scatter compares each gauge’s monthly NSE under the default (uncalibrated) parameters against the pooled prior, over the 2016–2024 scored window.

Default (uncalibrated) vs pooled-prior monthly NSE per gauge, clipped at −1 for display. Blue points are well-assigned gauges (drainage-area ratio 0.7–1.4); grey squares are assignment artifacts excluded from the honest subset. Shaded band = Moriasi Satisfactory and better. Source artifact: tampa_catalog_pooled_vs_default.png/.csv in the calibration-campaign research log.

All values are default (uncalibrated) → pooled prior, from one 10-year simulation scored with the same evaluation pipeline as the baseline (2016–2024 window, monthly Moriasi ratings). The honest trade-off: a few small urban canals that the default happened to fit went down under the regional prior — it trades a handful of idiosyncratic gauges for large basin-wide gains. The terminal calibration stage that follows the Peace playbook has since completed on this model — its per-gauge results are below.

Peace River: the first production regionalized MMPSO calibration

Complete

The Peace River model is the fleet’s calibration flagship and the demonstration target of the information-cascade methodology (manuscript in preparation). Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 8. The terminal step — a regionalized MMPSO over only the sensitive parameters, seeded from the pooled prior — closes the cascade: prior (a pooled distribution that transfers zero-shot) → diagnosis (which parameters to move, and a calibratability screen) → a short, seeded, regional terminal search that satisfices the fixed criteria on the reaches that carry the basin.

The search converged fast: the global-best objective (−19.25 in penalized units) was reached at iteration 8 and held unchanged for the remaining 15 iterations until the AWS spot instance was reclaimed — convergence, not a truncated run. A good pooled prior plus a fixed search space places the satisficing target within ~8 iterations, at roughly 100 model evaluations against the several hundred a from-scratch campaign would spend.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

The mainstem and Charlie Creek core — the reaches that integrate the basin’s water balance — is Good to Very Good (monthly NSE 0.86–0.92, daily 0.63–0.79). Four lower-basin and tidal gauges remain weak: the per-gauge floor is why they neither dominated the search nor were quietly dropped. Their failures are hydrologic-signal and regulation artifacts of the coastal setting (e.g. gauge 947, whose daily NSE is degenerate under tidal aliasing yet whose monthly NSE recovers to 0.53), not parametric deficits the free set could repair. They are reported at face value rather than excluded post hoc to inflate a count.

Regionalization is visible in the solution: percolation (perco) ranges from 1.98 in Horse Creek to 2.83 in Shell Creek across tributary groups (Prairie 2.79, Charlie 2.70, Peace mainstem 2.52, other 2.63) — spatially-varying calibrated parameters a single global scalar could not produce, reconciling the opposite per-tributary biases.

Per-gauge NSE above are from the best-so-far synchronized forward-run snapshot (the readable, directly-interpretable units); the assembled global-best solution is marginally better but stored in penalized units. Why the objective is floored: an earlier attempt collapsed because one pathological near-zero-flow gauge dominated a naive summed-NSE objective — the honest finding that produced the per-gauge floor and the roster discipline documented on /calibration-methods. Cost context: this class of run — a full regional calibration of a 57k-HRU model — has a measured compute cost of about $14 on AWS (see the AWS calibration page).

Cross-basin replication: the same cascade on Tampa Bay and Myakka

The cascade is designed to be basin-agnostic, so the demonstration should not rest on one basin. We ran the identical workflow — pooled prior → optimizer-free diagnosis → a short, seeded, regional terminal MMPSO — to completion on two more fleet models. Both converged from the same pooled-prior seed, giving a three-basin demonstration rather than a single case.

Tampa Bay (77.6k HRUs): terminal calibration

Complete

The zero-shot pooled prior above lifted the 58-gauge roster in one simulation; the terminal step then refines it. Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 6. This is the cross-basin case: the pooled prior carried no Tampa Bay information, yet a six-iteration regional search on a 77.6k-HRU model recovers Satisfactory-to-Very-Good skill on the reaches that carry the basin.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

The well-gauged core is Satisfactory-to-Very-Good on the reaches that carry the basin (core daily NSE 0.52–0.70, monthly 0.62–0.85, with several interior gauges reaching monthly NSE 0.80–0.85). A cluster of seven small urban canals stays weak — their monthly PBIAS pins at roughly −100% (near-zero measured baseflow the coarse 500 m routing cannot reproduce) — and the per-gauge floor is why they neither dominated the search nor were quietly dropped. The same cascade, the same honesty rail: seeded from the pooled prior, converged in six iterations on a 77.6k-HRU model, weak tail reported at face value.

Myakka River (29k HRUs): terminal calibration

Complete

Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 5. Myakka ran against the coverage-vetted 11-gauge roster — the corrected roster after a −99 sentinel value was found and scrubbed, the same roster-discipline point documented on /calibration-methods.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

Myakka is the flattest, most diffuse basin in the fleet, and its daily skill is inherently lower — but the whole coverage-vetted roster stays positive at both time scales. The core mainstem/wet-prairie gauges reach monthly NSE 0.56–0.65 (gauges 3592, 148, 191, 297), with daily NSE a modest 0.17–0.40 reflecting the flatwoods hydrology. No pathological floored tail: a fully-positive but honestly modest result from the same pooled-prior seed and the same short terminal search.

Across three basins — Peace (57,998 HRUs), Tampa Bay (77,636), and Myakka (29,071) — the same three-step cascade converged from the same pooled-prior seed in five to eight iterations, each time recovering defensible skill on the core gauges while retaining any weak reaches honestly under the per-gauge floor. That the workflow holds across a phosphorus-driven mining basin, a nitrogen-limited estuary drainage, and a flat wet-prairie river is the point: the method is basin-agnostic, now with three basins of evidence.