AtmosGrid — Planetary IntelligenceCOMMERCIAL PRODUCT
Forecast verification error attribution and radar-ground gauge comparison grid
← All basin intelligence modules

NUMERICAL PREDICTION DIAGNOSTICS

Rainfall Forecast Verification API — 48-hour model error and bias diagnosis.

A continuous verification record of what each weather forecast predicted, what later occurred at the catchment level, and where models repeatedly miss heavy rain. Assess and improve existing 48-hour numerical models.

AVAILABLE · GET & POST ENDPOINTSee how it works ↓
Illustrative agriculture imagery

THE DECISION IT SUPPORTS

Where are our numerical weather forecasts repeatedly under-predicting or over-predicting rain?

AtmosGrid continuously matches forecast lead runs (e.g. 48-hour horizons) against observed radar and gauge accumulations, computing Mean Absolute Error (MAE), Probability of Detection (POD), and False Alarm Ratio (FAR). Teams identify orographic blind spots and model biases.

Built for

  • River sensor network operators
  • Catchment flood warning authorities
  • Hydrological digital twin platforms
  • Municipal flood defense & siren controllers

WHY TEAMS USE IT

Turn broad data into a clearer next step.

Each result is designed to narrow a decision, show the inputs behind it, and surface what must still be checked locally.

01

Diagnose systematic model bias

Differentiate orographic under-prediction on windward ridges from leeward rain-shadow drying errors.

02

Improve 48-hour model calibration

Supply empirical bias-correction coefficients to internal digital twin forecasts and hydrological models.

03

Track skill on flood thresholds

Evaluate model accuracy specifically on rain events >25mm/6h rather than low-impact drizzle.

04

Transparent proof of accuracy

Build audit trails proving where forecast model updates have genuinely improved accuracy.

FROM REGION TO FIELD

See where each signal fits.

Images illustrate the operational setting. They are not live observations; use source timestamps and local measurements for decisions.

Forecast runs matched against ground truthVERIFICATION DASHBOARD

Forecast runs matched against ground truth

Continuous matching of 48-hour model predictions against subsequent gauge and radar observations.

Pinpoint repeated convective miss zonesBIAS HOTSPOT DETECTION

Pinpoint repeated convective miss zones

Maps where steep orographic terrain causes numerical models to systematically under-count rain.

POD, FAR and Critical Success IndexSKILL SCORE METRICS

POD, FAR and Critical Success Index

Calculates Probability of Detection (POD) and False Alarm Ratio (FAR) specifically for high-runoff storm thresholds.

THE INFORMATION STACK

What goes in—and what it helps explain.

Availability is stated per input. AtmosGrid never converts a missing parcel observation into a made-up field value.

INFORMATIONSOURCEHOW IT IS USEDSTATUS
0148-hour forecast runs (NWP)Operational weather modelHistorical forecast time-series at 3h/6h intervals.REQUIRED
02Observed gauge accumulationsGround rain gauges / radarTrue ground-truth rainfall for verification.REQUIRED
03Catchment boundary GeoJSONGIS databaseSpatial domain for area-weighted verification.REQUIRED
04Terrain elevation model (DEM)Copernicus 30m DEMIsolates orographic elevation-dependent biases.OPTIONAL

A PRACTICAL WORKFLOW

From input to reviewable decision.

The workflow keeps model context, field evidence and human approval in their proper places.

  1. 01

    Archive forecast runs

    Capture and archive 48-hour numerical forecast cycles at regular 6-hour cycle intervals.

  2. 02

    Co-locate with ground observations

    Match each forecast cell against ground rain gauges and radar accumulations across the catchment.

  3. 03

    Compute verification metrics

    Calculate continuous errors (MAE, RMSE, Bias) and categorical skill scores (POD, FAR, CSI).

  4. 04

    Pinpoint geographic bias zones

    Flag specific tributaries or ridges with persistent under-prediction to tune downscaling algorithms.

HOW INFORMATION BECOMES USEFUL

Examples of decisions it can inform.

These are screening patterns, not automatic instructions. Operational thresholds and local advisories remain authoritative.

SCENARIO 01

Persistent under-prediction of rain on windward mountain ridge

Signal
Systematic bias ratio is 0.72 (-14mm average error on events >30mm) in South Branch Ridge.
Useful action
Apply localized orographic elevation lapse-rate correction factor (1.35x multiplier) for southerly storm tracks.
Check before acting
Ensure correction factor is applied only when wind direction and humidity support ridge uplift.
SCENARIO 02

High false alarm ratio during summer convective showers

Signal
FAR is 0.42 for 6h lead forecasts; model triggers warnings for cells that evaporate before hitting the ground.
Useful action
Incorporate sub-hourly dew point depression and radar reflectivity to filter ungrounded virga rain.
Check before acting
Do not suppress alarms when low-level humidity exceeds 85%.
SCENARIO 03

Evaluating performance before digital twin ingestion

Signal
Forecast verification audit reveals 48h lead model achieves 0.78 Critical Success Index in main basin.
Useful action
Approve model stream for automated downstream river level simulation.
Check before acting
Maintain manual operator confirmation for alert levels exceeding Severe Stage 4.

FOR PRODUCT TEAMS

A transparent API contract.

The response exposes gaps and limitations so an application can explain confidence instead of hiding it.

POST/api/v1/hydrology/basin/forecast-verification

Example request

{
  "catchment_id": "basin-wye-severn-upper",
  "forecast_lead_hours": 48,
  "start_date_utc": "2026-09-01T00:00:00Z",
  "end_date_utc": "2026-09-30T23:59:59Z"
}
JSONdecision response

Example response

{
  "catchment_id": "basin-wye-severn-upper",
  "model_evaluated": "Operational 48h Mesoscale NWP vs Ground Telemetry",
  "evaluation_window_days": 30,
  "total_forecast_runs_evaluated": 60,
  "metrics": {
    "continuous": {
      "mean_absolute_error_mm": 4.82,
      "root_mean_square_error_mm": 6.94,
      "overall_bias_ratio": 0.86,
      "correlation_coefficient_r": 0.79
    },
    "convective_heavy_rain_scores": {
      "threshold_mm_6h": 25.0,
      "probability_of_detection_pod": 0.74,
      "false_alarm_ratio_far": 0.21,
      "critical_success_index_csi": 0.62
    }
  },
  "systematic_error_diagnosis": [
    {
      "region_name": "South Branch Ridge (Orographic Zone)",
      "bias_pattern": "under_prediction",
      "bias_magnitude_mm": -12.4,
      "primary_cause": "Sub-grid orographic terrain uplift under-resolved by standard numerical grid."
    }
  ]
}

CAPABILITY BOUNDARIES

Useful now. Honest about what comes next.

Supported now

  • Continuous error tracking (MAE, RMSE, Bias Ratio, Correlation r)
  • Categorical flood skill scores (POD, FAR, CSI, True Skill Statistic)
  • Geographic bias hotspot diagnosis per sub-tributary
  • Event-by-event classification (accurate, missed heavy rain, false alarm)

Additional data required

  • At least 30 days of archived forecast runs alongside ground observations
  • Ground rain gauge telemetry to establish authoritative ground-truth

BUILD RESPONSIBLY

Use the result to focus attention—not replace field judgment.

Pair AtmosGrid’s regional weather context with local measurements, crop stage, variety guidance and qualified agricultural advice.

Read the API documentation ↗Open the India viewer ↗Explore all basin intelligence modules ↗