The Gold Standard of Weather Radar Rainfall Gauge Adjustment | Comparing Station Ratios, Pooled Totals and Least-Squares Factors

Let Rᵢ be a radar-estimated rainfall accumulation and Gᵢ a quality-controlled gauge accumulation paired to it. Both use millimetres over the same stated interval. The station index identifies the pair, not an automatically independent experiment.

A radar cell and a gauge have different spatial footprints. A gauge records a local accumulation, while radar-derived rainfall represents an estimate produced from atmospheric measurements and processing. Their disagreement need not be entirely a single multiplicative calibration error.

For the first calculations, stipulate exact positive R values and nonnegative G values. The data are accepted as compatible paper pairs so we can isolate the estimator differences before addressing missing observations, zeros and measurement uncertainty.

Write the time interval and units beside the table. A ratio can be dimensionless and still be physically meaningless if its numerator and denominator cover different storms or observation windows. Correct algebra begins after the records have been matched appropriately.

Contents · Next chapter

The simplest model multiplies every relevant radar value by one common factor k, producing R* = kR. It assumes that one proportional adjustment is a useful description over the stated scope.

This is a restricted model. It does not include a separate additive offset, a location-dependent correction or a change with rainfall regime. Those alternatives may be relevant in an actual study, but they are not hidden inside the common multiplier.

The factor must be derived through a stated rule. Saying that it is the average bias is insufficient because several averages and loss functions are possible. Their numerical values can differ even on the same perfectly copied table.

NOAA’s MRMS information describes a broader multi-radar, multi-sensor estimation system and links its evolving product documentation. Our paper model is much simpler. Its purpose is to show what the chosen mathematical operation preserves, not to label one elementary factor as the complete operational algorithm of that system.

Contents · Previous chapter · Next chapter

Use two fictional matched pairs. At station A, R = one millimetre and G = four. At station B, R = nine millimetres and G = nine. The individual gauge-to-radar ratios are therefore four and one.

The radar total is ten millimetres across this list of point-pair records, while the gauge total is thirteen. These sums are bookkeeping totals across the supplied pairs. They are not automatically an area-integrated rainfall volume over a region.

The unequal ratios are deliberate. If every pair had the same ratio, all three estimators introduced here would agree, hiding the weighting issue. A carefully chosen counterexample makes the difference visible without relying on an actual rainfall dataset.

Contents · Previous chapter · Next chapter

Define kᵤ as the unweighted mean of qᵢ = Gᵢ/Rᵢ over the n accepted positive-radar pairs. In the original table, kᵤ = (4 + 1)/2 = 2.5.

Each pair receives the same weight in ratio space. The one-millimetre radar pair influences the factor just as strongly as the nine-millimetre pair. That is a property of the estimator, not an arithmetic mistake.

Applying kᵤ to the original radar values gives adjusted depths 2.5 and 22.5 millimetres. Their sum is twenty-five, which does not match the gauge sum thirteen.

The estimator was not designed to enforce that sum match. It describes the mean of the supplied station ratios. Whether that target is useful depends on the intended question and the quality of those ratios. Do not reject or endorse it merely because its value is the largest; examine what it weights and which consequence the application actually needs.

Contents · Previous chapter · Next chapter

Define kₜ = sum G divided by sum R, assuming the radar sum is positive. For the original pairs, kₜ = 13/10 = 1.3.

Applying that factor gives adjusted radar values 1.3 and 11.7 millimetres. Their sum is thirteen, exactly matching the gauge sum used to derive the factor. Individual pairs still disagree: the adjusted first value is below four, while the second is above nine.

This is a conservation property of the fitted pair-list sum. It does not establish that every station is corrected accurately or that the same factor transfers to an unobserved grid cell.

The subject root introduces an elementary ratio-of-totals example. This leaf develops what happens when station ratios differ and contrasts that estimator with other objectives. Keep the underlying records and the chosen aggregation rule visible so the result can be reproduced rather than guessed from the phrase mean-field correction.

Contents · Previous chapter · Next chapter

For every positive Rᵢ, write Gᵢ = Rᵢqᵢ. Substituting into the pooled factor gives kₜ = sum(Rᵢqᵢ)/sum Rᵢ.

Thus the ratio of totals is a weighted mean of the individual ratios, with weights proportional to radar accumulation. In our example, the weights are one tenth and nine tenths. The result is 0.1 × 4 + 0.9 × 1 = 1.3.

The unweighted ratio mean instead uses one half for each station. The difference between 2.5 and 1.3 is therefore an explicit difference in weights. It is not evidence that the stations were copied incorrectly.

These radar-amount weights should not be called geographical area weights unless an additional spatial account establishes that role. They arise algebraically from the chosen pair-list total. A station with more radar rainfall receives more influence under this rule, regardless of how much land its location can defensibly represent.

Contents · Previous chapter · Next chapter

A third rule chooses k to minimise the sum of squared depth residuals, S(k) = sum(kRᵢ − Gᵢ)². This measures errors in the original accumulation units rather than errors in station ratios.

Differentiating gives S'(k) = 2 sum[Rᵢ(kRᵢ − Gᵢ)]. Setting the derivative to zero yields kₗ = sum(RᵢGᵢ)/sum(Rᵢ²), provided the denominator is positive.

For the fictional pairs, the numerator is 1 × 4 + 9 × 9 = 85. The denominator is 1² + 9² = 82. The least-squares factor is 85/82, approximately 1.0366.

The second derivative is twice the positive denominator, so this is the unique unconstrained minimum for the stated objective. The derivation does not establish that squared depth error is the correct operational decision loss. It identifies what that particular mathematical fitting rule selects on the supplied observations.

Contents · Previous chapter · Next chapter

When all Rᵢ are positive, the numerator sum(RᵢGᵢ) can be written sum(Rᵢ²qᵢ). Therefore kₗ is a weighted mean of the station ratios with weights proportional to squared radar accumulation.

In the original table, those weights are one eighty-second and eighty-one eighty-seconds. The larger-radar pair receives even more relative influence than it did under the ratio of totals.

This explains why kₗ is closer to one. The second pair has ratio one and a much larger squared-radar weight. No additional meteorological evidence appeared between calculations; the objective changed how the same observations contributed.

A useful comparison table should identify all three rules by their weights or objectives. Calling each an average factor conceals the most important distinction. A learner who can derive these weights can recognise the same issue in another positive-valued calibration problem without assuming that every ratio summary serves the same target.

Contents · Previous chapter · Next chapter

For kₜ = 1.3, radar-minus-gauge residuals after correction are minus 2.7 and plus 2.7 millimetres. Their sum is zero, while the squared-error sum is 14.58 square millimetres.

For kᵤ = 2.5, the residuals are minus 1.5 and plus 13.5. Their sum is twelve and their squared-error sum is 184.5. For kₗ = 85/82, the squared-error sum is approximately 8.89, the smallest value under its stated fitting objective.

The least-squares result should win that particular in-sample squared-error comparison because it was chosen to minimise it. That is a mathematical property, not independent evidence of better rainfall estimates on new storms.

The pooled factor instead enforces the fitted pair-list sum. Different objectives can produce different apparent winners. State the objective before comparing the numbers so the evaluation does not quietly reward an estimator for the same condition used to construct it.

Contents · Previous chapter · Next chapter

If every accepted pair has Gᵢ = k₀Rᵢ for the same k₀ and positive Rᵢ, all station ratios equal k₀. Any of the three positive-weighted averages then gives k₀.

This explains why a proportional teaching example such as radar [10,20] and gauge [12,24] gives agreement at 1.2. It is a useful first calculation, but it cannot reveal how estimator choices behave when ratios vary.

Another agreement condition for the two ratio averages is equal positive R values, which make the pooled weights equal. Least squares also has equal weights in that case. These are mathematical conditions, not evidence that actual rainfall pairs must satisfy them.

Use agreement as a check with an identified scope. If a software implementation gives different results on an exact common-ratio table under these definitions, investigate the calculation or data filtering. If it gives different results on unequal-ratio data, that may be exactly what the stated estimators require.

Contents · Previous chapter · Next chapter

When R is positive and G is exactly zero in a paper record, q = G/R is zero. The pair is mathematically valid for all three estimators and can influence the factor downward.

A missing gauge observation is not the same thing. Replacing missing data with zero invents a rainfall measurement and can change the fitted correction. A quality flag should preserve the reason a record is unavailable or excluded.

Likewise, a gauge zero may need scrutiny in actual practice if timing, instrument condition or precipitation type raises a concern. The correct response follows the measurement evidence rather than an automatic assumption that every zero is either perfect truth or an error.

For a classroom exercise, add a third pair with positive R and G = zero. Recalculate the factors and explain which weights changed. Then replace that row with a genuinely missing G and show why it cannot remain in the same numerical formula without an explicit missing-data model.

Contents · Previous chapter · Next chapter

If Rᵢ = zero and Gᵢ is positive, the individual ratio Gᵢ/Rᵢ is undefined. The mean-of-ratios rule cannot include that row as an ordinary finite ratio. Assigning a huge arbitrary ratio is a different estimator, not a repair justified by algebra.

The ratio of totals can still be calculated if other rows make the radar sum positive. The zero-radar row contributes gauge rainfall to its numerator but nothing to its denominator.

In through-origin least squares, the same row contributes the constant residual square Gᵢ², independent of k. It therefore does not affect the derivative or fitted multiplier. No finite common multiplier can turn a zero radar estimate into a positive depth.

These differences are informative. The estimators respond differently because their mathematical targets differ, and the proportional model has a real limitation at zero. A report should preserve that limitation rather than hide it behind a numerical replacement for an undefined ratio.

Contents · Previous chapter · Next chapter

Take two fictional pairs: radar [0,10] millimetres and gauges [5,10]. The pooled factor is 15/10 = 1.5, giving adjusted values [0,15]. Their pair-list sum matches fifteen, but neither adjusted row equals its gauge.

The least-squares factor is (0 × 5 + 10 × 10)/(0² + 10²) = one. It gives [0,10], with an unavoidable five-millimetre first-row residual. The mean of individual ratios is not defined on both rows.

This is not a contest in which the factor with the most desirable number wins. The example exposes both the model’s zero-input limitation and the distinct constraints of the estimators.

Before changing the model, ask what the zero represents. A valid no-echo estimate, a missing radar value and a screened invalid value are different records. The paper exercise supplies an exact zero; an actual analysis must preserve the observation and quality definitions before deciding how the pair belongs in a fit.

Contents · Previous chapter · Next chapter

Suppose one pair has R = 0.01 millimetres and G = one millimetre. Its individual ratio is one hundred. A small absolute radar uncertainty near zero can change that ratio dramatically.

If R instead equals 0.02 with G unchanged, the ratio becomes fifty. The gauge difference is unchanged while the ratio halves. A mean of station ratios can therefore be strongly influenced by low-radar pairs under this setup.

This does not prescribe a universal exclusion threshold. A threshold, filter or regularisation rule changes the estimator and needs a stated rationale appropriate to the measurement and intended use.

For learning, keep the exact arithmetic and the decision separate. Calculate how the ratio changes across a supplied positive interval, then state what additional information would justify including, excluding or modelling that pair differently. The goal is to recognise sensitivity, not to invent a convenient screening rule after seeing which observations make the correction look better.

Contents · Previous chapter · Next chapter

A lower numerical weight is not automatically a quality judgement. In kₜ, weights arise from radar amounts. In kₗ, they arise from the squared-depth loss. A pair can receive a large mathematical weight while still having poor measurement quality.

Quality control asks whether the record is suitable for the comparison, based on instrument, timing, siting and other evidence. Weighting asks how accepted records contribute to a specified calculation. A complete method may involve both, but they should remain identifiable.

NASA’s publication record for Steiner and colleagues’ 1999 study identifies bias adjustment and gauge-data quality control as a research subject. We link that primary publication route as context, not as a claim that our three elementary estimators reproduce the study’s methods or conclusions.

A useful paper exercise assigns each row a supplied quality status before revealing its rainfall values. Apply the declared acceptance rule first, then calculate. This makes it harder to justify exclusions merely because they push the fitted factor away from a preferred answer.

Contents · Previous chapter · Next chapter

Duplicating a pair gives it additional influence under all three list-based estimators. The change is not new independent physical information if the second row is simply a copy of the same observation.

Duplicate the original low-radar pair [1,4], making radar [1,1,9] and gauge [4,4,9]. The unweighted ratio mean becomes three. The pooled factor becomes 17/11, approximately 1.5455. The least-squares factor becomes 89/83, approximately 1.0723.

Each changed because the represented list weights changed. None of those changes proves that the rainfall relationship became better known. Data provenance is therefore part of the estimator’s meaning.

For actual observations, closely related reports can also share information without being exact duplicates. Consecutive accumulations, neighbouring estimates or repeated processing of one gauge need suitable dependence and representation accounts. The key habit is to distinguish an additional row from an additional independent piece of evidence before interpreting a fitted factor’s apparent stability.

Contents · Previous chapter · Next chapter

A station-hour pair and a station-day pair are different aggregation units. If a proportional relation varies through time, pooling before taking a ratio changes the weights relative to averaging factors from separate intervals.

For a fictional station, suppose two accepted intervals have radar depths one and nine, and gauge depths four and nine. Their separately calculated ratios are four and one. The average interval ratio is 2.5, while the ratio of their combined depths is 1.3, just as in our two-station example.

The algebra does not care whether the index labels stations or times. The scientific meaning does. Overlapping windows should not be treated as independent disjoint accumulations, and missing intervals must not silently become zero rain.

Write the aggregation unit in the estimator description. A regional factor over a month and an average of hourly factors can answer different questions even when both use millimetres and are described informally as an average radar bias.

Contents · Previous chapter · Next chapter

A gauge network is a set of locations, not automatically an equal-area partition. Summing station depths does not directly calculate the total rainfall volume over a catchment.

If a hydrological task supplies justified spatial representation weights aᵢ, an area-related comparison may involve sum(aᵢGᵢ) and sum(aᵢRᵢ). That is a new explicitly weighted account. The weights must come from a relevant spatial model or evidence, not merely the order of rows in the table.

The Weather Radar Rainfall Estimation guide supplies the wider radar-to-rainfall route. Return to the gauge-adjustment root for footprints, interpolation and held-out validation.

The transferable lesson is to ask what each observation represents before adding it. Radar-amount weights that appear inside kₜ are not substitutes for spatial weights needed by a catchment question. A factor can match a pair-list total exactly while leaving the basin-volume question unresolved.

Contents · Previous chapter · Next chapter

A weighted depth-error objective can be written S(k) = sum[wᵢ(kRᵢ − Gᵢ)²] with supplied nonnegative weights. If sum(wᵢRᵢ²) is positive, differentiation gives k = sum(wᵢRᵢGᵢ)/sum(wᵢRᵢ²).

This is another estimator. Its weights might be motivated by an appropriate uncertainty model or a declared comparison target, but that interpretation must be justified rather than inferred from the letter w.

A zero weight removes a pair’s influence on the fit without necessarily deleting the original record. A large weight makes a pair more influential, not automatically more accurate. Negative weights would no longer describe the same nonnegative squared-loss weighting model.

For a classroom exercise, give the original two pairs weights nine and one. Work the numerator and denominator before comparing the result with the unweighted least-squares factor. The exercise demonstrates that a weighting decision is part of the model specification, not a cosmetic option selected after the preferred factor has been calculated.

Contents · Previous chapter · Next chapter

Suppose someone fits the reverse relationship, radar as a multiple of gauge, by minimising squared radar residuals. The reverse least-squares factor is h = sum(RᵢGᵢ)/sum(Gᵢ²). Taking its reciprocal generally does not recover the forward least-squares factor because the fitted residual direction has changed.

For the original pairs, h = 85/97. Its reciprocal is 97/85, approximately 1.1412, whereas the forward factor is 85/82, approximately 1.0366. Both calculations are dimensionally consistent. Their difference comes from the objective, not a disagreement about the table.

The pooled factor behaves differently under this reversal: interchanging the two positive totals gives its reciprocal exactly. That property belongs to that rule and should not be assumed for every calibration method. An inverse physical relationship and a reversed statistical fit are different operations.

For learning, derive both least-squares slopes before interpreting either. A report saying only that a factor was inverted leaves an important ambiguity: was an already fitted number reciprocated, or was a new regression fitted with a different response variable? Preserve that choice when checking reproducibility.

Contents · Previous chapter · Next chapter

Convert every matched depth from millimetres to centimetres by dividing both R and G by ten. Individual ratios remain unchanged. The ratio of totals remains unchanged too.

For least squares, both numerator sum(RG) and denominator sum(R²) scale by one hundredth, so their ratio is unchanged. The squared-error value changes units and numerical scale, but the fitted multiplier does not.

This is a useful implementation check. If a common unit conversion applied to both columns changes the factor under these exact definitions, investigate the calculation, filtering or numerical treatment.

Converting only one column is different: it changes the ratios because the numerator and denominator no longer use matching units. A dimensionless-looking output does not protect a spreadsheet from a unit mismatch. Keep units in the input headings and test one supplied pair independently before interpreting a large table’s summary as a rainfall correction.

Contents · Previous chapter · Next chapter

The proportional model passes through zero: k times zero is zero. An additive model R* = R + b has a different structure. A model containing both k and b introduces two parameters and a different fitting question.

In the zero-radar example, an additive term could produce a positive adjusted value at the first row. That flexibility does not automatically make it a better physical model or a better predictor on new data.

Model choice should follow the intended error account and validation, not the desire to force every training pair to match. More parameters can improve in-sample fit while making transfer less stable or creating other problems, such as negative adjusted values outside the fitted range.

The subject root develops a basic additive example separately from its multiplicative example. Keep that distinction when comparing factors. A disagreement between kᵤ, kₜ and kₗ occurs within one proportional family; it should not be confused with the broader choice among different correction-model structures.

Contents · Previous chapter · Next chapter

A fitted multiplier summarises a relationship under a chosen rule. It does not uniquely identify why radar and gauge values differed. The discrepancy can involve radar retrieval, gauge observation, spatial representativeness, timing or several influences together.

Suppose a factor above one improves agreement on the fitting pairs. It would be too strong to conclude solely from that result that the radar instrument was uniformly undercalibrated by the same percentage everywhere.

A causal interpretation needs observations and models addressing the relevant mechanism. The simple arithmetic can remain useful while the cause remains unresolved. This distinction is particularly important when an adjusted map looks smoother or more convincing than the original.

Write the fitted relationship first, then the proposed cause as a separate statement with its supporting evidence. That makes the account easier to revise when a new station, storm regime or quality review shows that a common proportional explanation does not transfer across the intended domain.

Contents · Previous chapter · Next chapter

An estimator can satisfy its fitting objective by construction. kₜ matches the fitted pair-list sum. kₗ minimises the fitted squared-error sum. Neither property is independent evidence that the correction improves new rainfall estimates.

Reserve appropriate observations or events for evaluation, with compatible units, windows and quality definitions. The relevant independence and grouping depend on the scientific design; neighbouring pixels or successive scans should not automatically be counted as unrelated trials.

For a paper demonstration, supply a new radar estimate R = four and a withheld gauge G = five millimetres. The original factors predict ten, 5.2 and about 4.1463 respectively. Their absolute errors are five, 0.2 and about 0.8537.

The pooled factor happens to perform best on this one invented check. One case cannot establish a general ranking. Its role is to show that the in-sample squared-error winner need not win every new observation, and that evaluation requires evidence beyond the fitted objective.

Contents · Previous chapter · Next chapter

Total bias, absolute station errors, squared errors, spatial placement and extreme-event behaviour describe different aspects of a precipitation product. A correction can improve one while worsening another.

The withheld example makes this concrete: the factor chosen to minimise original squared errors did not have the smallest absolute error at the new pair. A different test collection could produce another ordering.

Choose metrics according to the stated use and retain their definitions. A hydrological application may need its own downstream evaluation rather than assuming that a smaller pointwise rainfall error guarantees better flood outcomes.

This guide supplies no live warning or flood forecast. It teaches how an estimator and an evaluation objective interact. A useful conclusion names the cases, metric and scope: the method improved this defined score on these independent comparisons. It does not turn that limited result into a universal claim that the adjusted rainfall field is accurate everywhere or under every precipitation regime.

Contents · Previous chapter · Next chapter

An application may restrict k to an allowed interval. That changes the optimisation problem. For the simple convex least-squares objective, a constrained minimum on a closed interval is the unconstrained kₗ when it lies inside, otherwise the appropriate endpoint nearest to that minimiser.

This mathematical result does not provide an operational interval. Any actual lower or upper limit needs a justified method and documentation. Clipping after calculation is not equivalent to having chosen an unconstrained estimator if the clip is active.

For a paper example, constrain k between 1.1 and two. The original least-squares minimiser 85/82 lies below 1.1, so the constrained minimum occurs at 1.1. It no longer has the same first-order derivative condition as an interior solution.

Record both the fitted unconstrained value and the declared constraint where relevant. Otherwise another reader may be unable to reproduce why the reported factor differs from the formula, even though the software followed its configured rule correctly.

Contents · Previous chapter · Next chapter

Try radar depths [2,8] and gauge depths [6,8] millimetres. The station ratios are three and one. Their unweighted mean is two, while the ratio of totals is 14/10 = 1.4.

The least-squares numerator is 2 × 6 + 8 × 8 = 76 and denominator 2² + 8² = 68. The resulting factor is 19/17, approximately 1.1176.

Before calculating residuals, explain why the least-squares factor is closer to one than the unweighted ratio mean. The eight-millimetre radar pair has the larger squared-radar weight and ratio one.

Now apply the common change of units to centimetres and verify that all three factors remain unchanged. Finally, say which factor matches the pair-list sum and which minimises the stated squared-error objective. A good transfer answer connects each number with its estimator rather than describing one as the correct average without defining what average means.

Contents · Previous chapter · Next chapter

For repair, work with one positive matched pair and explain the units and interval before taking its ratio. Then compare two unequal-ratio pairs so that weighting becomes visible.

For stabilisation, derive all three factors, reproduce their adjusted values and check the property each is designed to satisfy. Change counts, rainfall magnitudes and units one condition at a time.

For extension, introduce a zero radar value, an uncertain low denominator, duplicate records or explicit external weights. State which formula remains defined and which assumption changed. Then compare against a genuinely withheld paper record without generalising from one successful case.

The Gold Standard of Education: Learning That Transfers into Life connects these tasks with independent judgement. The learner’s progress is visible when they can choose and explain an estimator for a new question, not simply repeat the multiplier calculated in the demonstration.

Contents · Previous chapter · Next chapter

Is a ratio of totals the same as a mean of ratios? It is a radar-amount-weighted mean when the individual radar values are positive, not generally the unweighted mean. Is least squares the same calculation? It minimises a declared depth-error objective and uses another weighting pattern.

Does matching the fitted total prove every station is accurate? No. Does the smallest fitted squared error prove the best new prediction? Not without independent evidence. Can a positive gauge value be recovered from zero radar using a finite multiplier alone? No.

Are the pair-list sums an automatic catchment rainfall volume? They need a spatial representation account before being given that meaning. Does a factor identify the physical cause of disagreement? It summarises a fitted relationship, while causal explanation requires further evidence.

The useful first question is which rule produced the factor and why that rule serves the stated purpose. Once the objective is clear, the numerical differences become understandable rather than contradictory.

Contents · Previous chapter · Next chapter

Return to Weather Radar Rainfall Gauge Adjustment for footprints, timing, quality control and independent verification. Use Weather Radar Rainfall Estimation for the preceding radar-to-rainfall question.

NOAA's MRMS QPE information supplies product and multi-sensor context. NASA's publication record for Steiner and colleagues' 1999 bias-adjustment and gauge-quality-control study provides a primary research route. Both were checked on 24 October 2026. Our estimator derivations and numerical pairs are original paper models, not operational settings or results taken from those pages.

Continue through How X Works with a reusable habit: name the observations, define the estimator and verify the consequence it actually promises. That sequence keeps a convenient correction factor from becoming an unexplained claim about rainfall truth.

Contents · Previous chapter · Return to Weather Radar Rainfall Gauge Adjustment