Radar estimates rainfall over a wide area, while rain gauges record local accumulation. Their combination can improve spatial precipitation estimates, but only when both datasets have been properly matched and quality-controlled.
The gold standard is a transparent adjustment method tested against independent gauges, with explicit assumptions about time, rainfall type, spatial transfer and uncertainty.
eduKateSG · Weather Radar Rainfall Gauge Adjustment
Find your next learning step
Calculate bias corrections, then see how to challenge their apparent improvement with withheld observations.
Full chapter index · Practice and transfer · Education root · How X Works
Full chapter index
Define matched rainfall · Chapters 1–2
Calculate corrections · Chapters 3–4
Investigate limitations · Chapters 5–7
Verify improvements · Chapters 8–10
Follow the evidence · Chapters 11–12
CHAPTER 1 OF 12 · Define matched rainfall
1. Radar rain estimates and rain gauges measure different footprints
Weather radar samples electromagnetic scattering from atmospheric particles. An operational quantitative precipitation estimation (QPE) algorithm infers rainfall on a grid using radar calibration, precipitation classification and conversion models; that grid is not an array of direct ground rain measurements.
A gauge accumulates water at one fixed point. Its recorded depth depends on siting, instrument design, wind, maintenance and the measurement interval. A single gauge cannot prove that every location in its radar cell experienced the same rain.
Gauge adjustment attempts to reduce selected disagreements between these different instruments using a quality-controlled network of ground observations. It cannot recreate every unobserved convective peak, correct every sensor error or guarantee more useful flood forecasts.
NOAA’s MRMS system documents radar-only and multi-sensor QPE products and local gauge-bias correction. The underlying methods can change between software releases, making product identity and date integral to any scientific comparison.
Write down radar variable, grid resolution, precipitation type, station location, units and exact accumulation period before deriving any correction. A mispaired gauge and radar window can produce a perfectly calculated but physically meaningless bias.
CHAPTER 2 OF 12 · Define matched rainfall
2. Bias has a sign and an observation interval
Define error e = R − G, where R is radar-estimated accumulation and G is a properly paired gauge accumulation over the same interval. Under this sign convention, a negative value is radar underestimation relative to the gauge.
Bias means an average error over an identified sample, whereas absolute error retains the magnitude of individual differences and RMSE gives additional weight to larger misses. A bias close to zero can conceal a combination of serious overestimates and underestimates.
Other publications may define error as G − R. Their bias correction then uses a different subtraction sign, although the underlying paired observations are identical. Always state the convention before showing a table.
Observations need quality flags. An hourly radar accumulation cannot be compared without qualification with a gauge’s daily total, and a missing gauge value must not be set equal to a valid zero-rain observation.
Radar and gauge records also differ in representativeness: a rain cell passing within a kilometre of a gauge can affect a grid-area estimate without striking that point. The apparent disagreement is not always solely a radar calibration fault.
CHAPTER 3 OF 12 · Calculate corrections
3. Worked example: correct one radar estimate additively
Imagine a hypothetical one-hour radar estimate R = 12 mm at a grid cell and a properly time-matched gauge depth G = 15 mm. The chosen radar-minus-gauge error is e = 12 − 15 = −3 mm.
Under a simple additive correction, R* = R − b. Taking local bias estimate b = −3 mm gives R* = 12 − (−3) = 15 mm, exactly matching the gauge in this particular example.
The match gives no independent proof of skill because the same gauge was used to create the correction. A new independent station, event or time must be reserved to test whether the method improves unseen cases.
Suppose an independent gauge at another nearby cell records 13 mm, its radar-only estimate is 11 mm, and a justified interpolated local correction adds 3 mm. Corrected rainfall becomes 14 mm: the original −2 mm error changes to +1 mm, a smaller absolute error for that one check.
This arithmetic is intentionally simple. A real local bias surface changes with gauge positions, quality, weights, rainfall type and method-specific constraints. It should not be constructed from one station and treated as universally accurate.
CHAPTER 4 OF 12 · Calculate corrections
4. Worked example: mean-field multiplicative adjustment
Consider a fictional region and one valid hour with two accepted radar–gauge pairs. Radar accumulations are [10, 20] mm and gauges measure [12, 24] mm. Assume the measurements have been paired correctly.
One elementary regional factor is F = (ΣG)/(ΣR) = (12+24)/(10+20) = 36/30 = 1.20. A third grid cell with 15 mm radar rainfall would then be adjusted to 15×1.20 = 18 mm if the model were applicable there.
That 18 mm is not observed rainfall. It is a modelled value using a regional relation estimated from two other stations, with uncertainty in both the reference and spatial transfer.
A multiplicative ratio is unstable if summed radar rainfall is near zero. It may also misrepresent rain gradients, snow, hail-bearing precipitation or systematic offsets not proportional to intensity.
Other mean-field implementations have their own weighting and filtering. The simple ratio here demonstrates dimensional logic and the assumption of a common proportional error, not NOAA’s full operational algorithm.
Compare with an additive model when appropriate and score both on independent cases. A more complicated method is not automatically more reliable in a sparse or inaccurate network.
Unequal station ratios reveal differences between common multiplier rules. Comparing Station Ratios, Pooled Totals and Least-Squares Factors compares station-ratio means, pooled totals and least-squares factors with zero-input and independent-check examples.
CHAPTER 5 OF 12 · Investigate limitations
5. Local interpolation adds assumptions between gauges
NOAA’s MRMS gauge-bias-correction documentation describes quality-controlled gauge observations and an interpolated, spatially varying error field. Distance, gauge-network density and cross-validation are important parts of the method.
A station’s measured difference may be informative nearby but less relevant on the opposite side of a mountain range or a sharp convective rain gradient. A correction model must specify its weighting, radius of influence and limits of extrapolation.
In an illustrative three-station setup, imagine local radar-minus-gauge biases of −2, +1 and −4 mm. A new grid cell could receive a weighted average based on location and prescribed rules; simply taking −5/3 mm without regard to spatial context is one possible but not universally justified assumption.
An uneven station network can generate localized correction patches that appear like true rainfall structures. Validate against withheld stations and inspect the rainfall field to make sure spatial features are plausible.
Where no reliable gauges exist, retain the fact that correction evidence is weak. A completed-looking map does not justify equal confidence everywhere.
Save the raw grid, accepted gauges, rejected gauges, weights, bias field, correction procedure and final result so the operation can be independently reproduced.
CHAPTER 6 OF 12 · Investigate limitations
6. Gauge screening protects against bad corrections
Gauge measurements have their own failure modes: instrument blockage, tipping-bucket nonlinearity, wind effects, frozen precipitation, incorrect timestamps, incomplete reports and maintenance problems.
NOAA’s gauge-bias product description explains screening implausible or late records before allowing their differences to influence nearby radar pixels. The algorithm can also reject some anomalous gauges after spatial cross-checking.
Quality screening must be physically informed. A genuine very localized downpour may differ from its neighbours, so deleting every statistically unusual gauge could erase the very extremes that matter most.
A gauge that reads too high can force adjusted radar values upward across its interpolation neighborhood. A missed gauge report incorrectly entered as zero can do the opposite.
Retain reasons for flags and exclusion rather than silently modifying the raw observations. Verification stations should, wherever feasible, be separate from those used in correction.
Gauge quality does not simply improve by averaging more bad stations: correlated calibration or siting errors can affect an entire network.
CHAPTER 7 OF 12 · Investigate limitations
7. Timing, rate and depth are separate considerations
Radar volumes are acquired over a finite scan time, while gauges accumulate rain between clock-defined observation endpoints. Storm motion can make a five- or ten-minute mismatch important, especially in short intense rain.
Suppose a one-hour radar estimate ends at 12:00 while the corresponding gauge report describes 11:10–12:10. Both span one hour, but their time intervals are not identical; a thunderstorm near noon may produce a large false difference.
An intensity in mm/h is not an accumulated depth until multiplied by its relevant time in hours. Two successive 15-minute rates of 8 and 4 mm/h imply depths 2 and 1 mm and a 30-minute total of 3 mm under the stated piecewise-constant model.
Overlapping rolling accumulations must not be summed as though separate. Likewise, gauge-adjusted hourly products and a longer aggregate derived from them are not independent rainfall measurements.
Quality control should verify start and end times, timezone, radar scan coverage, gauge reporting delay, missing intervals and whether the radar value is a rate, depth or rolling product.
A valid mathematical comparison always preserves the interval in which the physical rain was observed.
CHAPTER 8 OF 12 · Verify improvements
8. Not every radar error is solved with gauge adjustment
Radar QPE can be affected by reflectivity calibration, Z–R conversion, drop-size distribution, beam blockage, attenuation, melting-layer bright-band enhancement and low-level evaporation.
A ground gauge helps quantify some total discrepancy but does not necessarily explain its origin. A corrected map can have smaller point errors while still confusing where a storm’s intense core was located.
NOAA’s MRMS information describes combining radar, atmospheric, surface and other information to improve precipitation estimation. Physics corrections before gauge adjustment are not interchangeable with the statistical adjustment afterward.
Rain gauges can undercatch blowing rain or snow. Terrain makes beam height and local precipitation particularly difficult, and an isolated point measurement may not represent a coarse radar area.
An adjustment tested only in stratiform rain might not perform equally under convective hailstorms. Partition validation by relevant precipitation type, season, intensity and radar range when sufficient independent cases exist.
Do not turn a low radar–gauge bias into a blanket assertion that the entire precipitation product is free from physical or spatial uncertainty.
CHAPTER 9 OF 12 · Verify improvements
9. Worked example: independent RMSE versus mean bias
Imagine three withheld gauge observations [10, 10, 20] mm with radar-only estimates [8, 14, 18] mm. Radar-minus-gauge errors are [−2, +4, −2] mm.
The mean error is (−2+4−2)/3 = 0 mm, although each individual radar estimate is wrong. RMSE = √[(4+16+4)/3] = √8 ≈ 2.83 mm.
Suppose a separately developed correction yields estimates [9, 11, 21] mm on those same withheld gauges. Corrected errors are [−1,+1,+1] mm, mean error is +1/3 mm and RMSE is √[(1+1+1)/3] = 1 mm.
The corrected product has lower RMSE in this miniature teaching case even though its mean bias moved away from zero. These three samples cannot support a credible operational performance claim or confidence interval.
Real verification needs many independent storms or appropriate blocked comparisons. Neighbouring cells and consecutive radar scans are correlated, so counting every pixel as a new independent experiment exaggerates precision.
Document whether the gauges used to score the result were withheld from interpolation and model tuning.
CHAPTER 10 OF 12 · Verify improvements
10. Decision quality, FAQs and changed catchment transfer
A corrected rainfall field can help hydrological modelling, water storage studies or operational forecasting, but their error priorities differ. Total monthly rain, five-minute peak intensity and storm positioning are not interchangeable success criteria.
A model that improves overall gauge agreement may still create artificial local patches or miss short peaks. Downstream flood-model performance should be verified separately from the rainfall estimate.
Practice A: With radar 12 mm and gauge 15 mm, compute bias −3 mm and adjusted depth 15 mm under the stated sign convention; explain why this is in-sample agreement.
Practice B: From [10,20] radar and [12,24] gauge totals, find a factor 1.20 and explain why a 15 mm estimate becomes 18 mm without proving the latter is observed truth.
Practice C: Check that errors [−2,+4,−2] have zero bias yet RMSE about 2.83 mm.
FAQ — Is every rain gauge perfect ground truth? No. It samples a point and requires quality control.
FAQ — Does a lower rainfall RMSE prove better flood alerts? No. Timing, spatial distribution, model response and uncertainty also matter.
Transfer from a dense lowland gauge network to a mountainous catchment with few stations. Explain how the meaning of local bias and its uncertainty would change.
CHAPTER 11 OF 12 · Follow the evidence
11. Connected radar, meteorology and hydrology routes
The radar gauge-adjustment problem begins with physical radar QPE, passes through observations and statistical correction, and ends with a defined user decision. Each part has different failure modes.
Read the related Gold Standard explanations on radar attenuation, bright-band correction, hydrometeor classification, hydrology and forecast verification. They offer complementary evidence rather than generic substitutes for a valid gauge-adjustment study.
Verified related eduKateSG pages:
- The Gold Standard of Weather Radar Rainfall Estimation — radar-to-depth physics
- The Gold Standard of Weather Radar Attenuation Correction — propagation signal errors
- The Gold Standard of Weather Radar Bright-Band Correction — melting-layer effects
- The Gold Standard of Weather Radar Dual-Polarization Hydrometeor Classification — precipitation-type inference
- The Gold Standard of Hydrology — basin response and hydrological use
- The Gold Standard of Weather Forecast Verification — fair comparisons and metric definitions
- The Gold Standard of Statistics — uncertainty and independence
CHAPTER 12 OF 12 · Follow the evidence
12. Official references and next learning step
NOAA’s MRMS documentation describes local gauge correction, station quality control, interpolation and cross-validation. Peer-reviewed research explains alternative additive and multiplicative adjustments and their sensitivity to precipitation and station coverage.
All rainfall, bias and performance values here were invented for instruction. The article does not provide live rain conditions, flood forecasts or a certified algorithm for a named radar network.
An independent assignment should compare radar-only and adjusted QPE against gauges withheld from adjustment, with common accumulation windows, quality flags, multiple storm regimes and confidence intervals.
Primary sources supporting the measurement and correction principles:
- NOAA — Radar QPE With Gauge Bias Correction — quality control and local error interpolation
- NOAA — MRMS QPE — operational product and version context
- AMS — Local Gauge Correction in MRMS — network density and correction performance
- Hydrological Sciences Journal — Bias Correction Methods — alternative models and assumptions
- NASA GPM — Rain Gauge Quality and Radar Bias — the importance of ground measurement quality
The immutable Clementi Mathematics reference illustrates lesson structure; it is not an operational radar-processing protocol.