Every output is reproducible.
Calculations include the inputs, data sources, and step-by-step values. Re-running the same parcel against the same dataset must produce the same answer.
The methods underneath
Hydrology that holds
in regulatory review.
StormwaterIQ is a computational platform, not a forms platform. Every number it produces is traceable to a public dataset, a published method, and a versioned calculation. This page documents what that means in practice.
A pour point and a 1/3 arc-second DEM. D8 flow direction traces the contributing area; the catchment polygon resolves around it.
01 — Terrain & watershed
From a pour point, an entire contributing area.
We start with the USGS 3DEP DEM at 1/3 arc-second. A D8 algorithm (PySheds) computes flow direction at each cell, then accumulates upstream contribution from a chosen pour point. The output is a polygon — the watershed — plus its area, mean slope, and primary flow paths.
- Algorithm
- D8 flow direction; D-infinity available for low-relief terrain
- Source DEM
- USGS 3DEP 1/3 arc-second (≈ 10 m). 1 m where lidar exists.
- Performance
- Watershed up to 10 sq mi delineated in ≤ 10 s, p95
- Validation
- Elevation queries within ±0.5 m of published USGS benchmarks
A 25-year design storm hyetograph drives the watershed's unit-hydrograph response. Tc (time of concentration) and Tp (time-to-peak) fall out of the geometry.
Peak runoff for a parcel — Q = C · i · A. C is auto-populated from NLCD land cover, i from NOAA Atlas 14 IDF curves, A from the parcel polygon.
Q=C·i·A
C · runoff coefficient
0.00
NLCD developed-medium intensityi · 25-yr intensity
0.00 in/hr
NOAA Atlas 14, Tc = 18 minA · drainage area
0.00 ac
Parcel polygon (GeoJSON)Peak flow Q
0.00 cfs
Returned in < 4 s, p9502 — Runoff calculation
The simple equation, fully sourced.
The Rational Method has been the engineer's default for a century — Q = C · i · A. What changes here is the provenance of every term. C is intersected from NLCD land cover. i is read directly from NOAA Atlas 14 IDF curves at the parcel's exact lat/lng. A is the parcel polygon you supplied or drew.
- Methods supported
- Rational; SCS / NRCS curve number; pre/post comparison; detention sizing
- Tc methods
- Kirpich, Izzard, TR-55 — selected by site type
- Validation target
- Within ±2% of hand calculation across the verified test corpus
- Performance
- Full Rational including Atlas 14 lookup ≤ 4 s, p95
03 — Computer vision
Faster than the photo upload. Always advisory.
A compressed Core ML / TFLite model classifies the BMP type and detects deficiencies on the inspector's device — under 1.5 s per photo, no network required. On sync, the original full-resolution image is re-analyzed by a server model. Every finding is advisory: an inspector confirms or overrides before it becomes a compliance record.
- Backbone
- YOLOv8 fine-tuned on labeled field photography
- BMP classes
- 20+ — silt fence, rock check dam, sediment basin, wattle, inlet protection, bioretention…
- Deficiency targets
- Precision ≥ 0.80, recall ≥ 0.75 on silt-fence tear (highest-priority class)
- Active learning
- Inspector overrides feed a SageMaker retraining pipeline quarterly
On-device classification across 20+ BMP types. Deficiencies — fence tears, undercutting, turbid discharge — surface for inspector confirmation, never auto-recorded.
04 — Permit verification
NPDES coverage, checked nightly.
Scrapers and API connectors maintain a current view across all 50 state NPDES registries. Permit lookups by ID, project address, or parcel APN return status, expiration, issuing authority, and MCM requirements in under two seconds. Alerts route at 90, 60, 30, and 7 days. Annual reports are populated from program data when due.
NPDES coverage is verified nightly across all 50 states. Alerts at 90 / 60 / 30 / 7 days route to configurable escalation contacts; the annual report is generated from program data when due.
05 — Operating assumptions
Four non-negotiables.
These are the constraints we let shape the platform — not aspirations and not marketing. Each one is enforced in code, not in process.
Computer vision is advisory.
On-device classification surfaces likely deficiencies and labels confidence on each. Findings become compliance records only after an inspector confirms or overrides them.
Public data sources are cited, not paraphrased.
Atlas 14, 3DEP, SSURGO, NLCD and ECHO are queried directly. Where caching is required for performance, the cache key includes the source's published version.
Performance targets are SLAs.
The published p95 numbers — 4 s hydrology, 10 s watershed, 20 min storm trigger latency — are commitments under load, not demo conditions.
06 — Sources & cadence
Receipts.
The public datasets and federal services that StormwaterIQ queries directly, with the refresh cadence each one is treated under.
USGS 3DEP
1/3 arc-second DEM. Source for elevation queries, slope rasters, watershed delineation, and flow-path tracing.
Static; updated as new lidar tiles publishNOAA Atlas 14
Precipitation Frequency Estimates. IDF curves used for rainfall intensity at any US lat/lng across return periods 2 – 100 yr.
Reference set; periodically refreshed by NOAANOAA NWS · MRMS
Multi-Radar Multi-Sensor mosaic. Real-time 1 km² rainfall accumulation drives storm-event triggers.
Sub-hourly; trigger evaluation every 15 minNRCS · SSURGO
Soil Survey Geographic Database. Hydrologic soil group (A/B/C/D) and infiltration parameters for CN weighting.
Static; reissued by NRCS on scheduleMRLC · NLCD
National Land Cover Database. Pixel-level land-use classification for runoff coefficient lookup.
Static; refreshed every two to three yearsEPA ECHO · NPDES
Facility search and NPDES status. Used to verify permit coverage by ID, address, or parcel APN.
Nightly sync across all 50 states