The model
A terrain engine, an observation stack, and the layer that makes both believable.
The RegenX Geospatial Computational Model is three layers deep. The first two are what most people mean by geospatial analysis.
The terrain engine
Where water goes, how fast it gets there, and what it passes on the way.
Elevation is the one layer that constrains everything else. We run our own hydrological routing over 30 m Copernicus elevation data: a priority-flood depression fill, D8 flow directions, flow accumulation, and an upstream search on the reversed flow graph. From that we derive catchments, channel geometry, impoundment volumes and arrival times.
Routing rather than proximity is not a technicality. In the 2026 Nepal–Tibet flood, the largest lakes near the channel sat north of the drainage divide and drained away from Nepal entirely. Ranking by distance includes them. Only flow-routed delineation excludes them.
- Catchment delineation — contributing area above any pour point, with adaptive snapping so the point lands on the real channel rather than a side stream.
- Channel confinement — valley width measured at a stated height above the thalweg, which is what governs whether a surge stays deep or attenuates.
- Impoundment volume — DEM flood-fill restricted to the dam cell's own upstream mask.
- Arrival time — routed travel time from source to the first inhabited structure.
- Exposure — structures inside a routed corridor, counted from open building footprints, never inferred from population rasters.
Nepal's lethal 2026 corridor was 180–480 m wide for 60 km, measured 50 m above the channel bed. Measured at the default 4 m, the same valleys read 10–75 m — the summer stream, not the flood. The comparison only works at the height the flood actually occupies.
The observation stack
Multi-year, cloud-masked, radiometrically correct time series — not a single pretty scene.
A composite from one date is a photograph. A decision needs a record. We build per-pixel time series from Sentinel-2 and Sentinel-1, masked scene-by-scene with the scene classification layer, and read directly from cloud-optimised archives so a five-year stack over a site is minutes of work rather than a data-download project.
The unglamorous part is correctness. Sentinel-2 scenes from processing baseline 04.00 onward carry a reflectance offset that must be subtracted; skip it and every band ratio after January 2022 is biased toward zero while everything before is untouched — a series straddling two radiometric scales, with a trend that is pure artefact. We found that in our own glacier lake series, corrected it, and now cross-check archives against one another as a standing test.
- Water-dominated averages. Never compute a band-index trend over an unmasked, water-dominated frame. At Al Marjan, 81% of an apparent index rise came from seawater pixels where both bands sit near zero and the ratio swings on noise.
- Dilution. If the feature of interest is 5.7% of the frame, an area-wide mean cannot see it. Zone the frame first; measure the zone.
- Seasonal sampling. A first year covering only the low season and a last year covering only the high season will manufacture a trend from nothing. Compare month-matched windows or not at all.
Honest validation
The layer that decides whether the first two produced a finding or a coincidence.
Our standing rule is simple: never ship an ordinary least-squares trend alone. A claim is significant only when a Mann–Kendall test gives p < 0.05 and the t-statistic exceeds 2 after the series has been deseasonalised against its own month-of-year anomalies and corrected for serial correlation.
That correction matters more than it sounds. An index series with lag-one autocorrelation of 0.77 has an effective sample size of 36, not 283. Trends that look decisive at n = 283 routinely evaporate at n = 36 — and a client who acts on the first number has bought an artefact.
For predictive models we go further and retrain honestly. Leave-one-year-out hindcasting retrains the model on every year except the held-out one, refits the calibration and threshold on those years too, then predicts the year it has never seen. It is the only number that answers the question a client is actually asking.
- Theil–Sen slope in place of OLS, so one anomalous year cannot lever a trend.
- Mann–Kendall for monotonic significance without a distributional assumption.
- AR(1) effective-n correction before any t-statistic is quoted.
- Leave-one-year-out retraining for every predictive model, reported beside the in-sample figure.
- Back-testing against documented events with no per-site tuning — the same pipeline, run on cases with an established cause.
- Cross-archive agreement — the same scene read from two independent providers, as the most direct correctness test available.
A back-test on three floods with established causes scored our own lake-size argument 2 of 3. The miss was fatal to the claim, so the claim was withdrawn — in the client document, in the deck, and in the briefing that followed.
The replacement argument, built on catchment geometry and channel confinement, was never dependent on lake size and stands unchanged. Read that case
Foundation
The archive is public. Reading it is not.
We build on the Earth-observation archives the space agencies publish, and we name every one of them. That is deliberate — it means a client can have our conclusions independently re-derived and audited rather than taken on trust. What we are engaged for is everything that sits on top.
| Source | Provides | Resolution | Record |
|---|---|---|---|
| Copernicus Sentinel-2 L2A | Optical surface reflectance; all band indices | 10–20 m | 2015 → |
| Copernicus Sentinel-1 GRD / RTC | Radar backscatter; surface change through cloud | 10–20 m | 2014 → |
| Copernicus DEM GLO-30 | Elevation for all terrain routing | 30 m | TanDEM-X 2011–15 |
| NASA HLS (Harmonised Landsat–Sentinel) | Denser optical revisit for urban change | 30 m | 2013 → |
| ERA5 reanalysis | Temperature, precipitation, degree-days | ~31 km | 31-year baseline |
| NASA POWER | Point climatology, solar resource | Point | 1984 → |
| ITS_LIVE | Glacier surface velocity | 120 m | 1985 → |
| RGI 7.0 · HMAGLOFDB | Glacier outlines; documented outburst events | Vector | Inventory |
| ESA WorldCover · Open Buildings · OSM | Land cover, building footprints, hydrography | 10 m / vector | Current |
Anyone can download a Sentinel-2 scene. Turning 746 of them into a defensible carbon figure is a different job — knowing which 284 to discard, knowing that a 2022 change to the processing chain silently biases every band ratio after it, knowing that an island signal is diluted twentyfold in an area-wide average, and being able to show which of the resulting claims survived testing.
The archive is the raw material. The engine, the corrections, the domain judgement and the validation are the work — and they are what an engagement buys. Where a commercial very-high-resolution tasking genuinely adds something the public record cannot give, we say so and scope it separately.
Engagement
How a site becomes a decision.
Boundary and question
You send a boundary — KML, shapefile or corner coordinates — and the decision you are trying to make. We come back with what the open record can and cannot resolve at that location, before anything is committed.
Baseline
Terrain, land cover, hydrology, built form and the multi-decade climate envelope assembled onto one frame, with every layer's source, date and resolution recorded.
Time series
Five to ten years of cloud-masked index series over the exact boundary, zoned so the signal of interest is not diluted by whatever surrounds it.
Testing
Every candidate finding runs the robustness suite. Findings that survive carry their statistic; findings that do not are listed as failures with the test that killed them.
Delivery
A determination — not a data dump. Typically a ranked list of findings, the figures behind them, the assumptions you need to confirm, and a rebuildable source so the whole analysis can be re-run.
Monitoring, where it earns its place
For live hazard or portfolio work, the pipeline runs on a schedule and the outputs update themselves — with observation freshness shown beside every reading, because an old reading is an absence of information, not an all-clear.
Next
See the method applied.
Six engagements, with the numbers that survived testing and the ones that did not.