Frequency & data · South Africa

Quantile regression technique flood estimation

The quantile regression technique (QRT) estimates each flood quantile directly from three catchment descriptors: area, the 1:10-year 24-hour design rainfall and the distance to the coast. Each of 41 South African clusters has its own coefficients for every return period from 1:2 to 1:200 years. Studio's QRT calculator finds the cluster and coast distance from the catchment centroid and reports every coefficient it used.

The regression

For each return period TT the QRT estimates the peak from three descriptors:

ln⁡QT=CONSTT+aTln⁡A+bTln⁡DR10%+cTln⁡Dc\ln Q_T = CONST_T + a_T \ln A + b_T \ln DR_{10\%} + c_T \ln D_c
SymbolMeaningUnits
QTQ_TPeak discharge for return period TTm³/s
AACatchment areakm²
DR10%DR_{10\%}1:10-year 24-hour point design rainfall at the centroidmm
DcD_cPlanar distance from the centroid to the nearest coastline, at least 0.01decimal degrees
CONSTT,aT,bT,cTCONST_T, a_T, b_T, c_TCoefficients of the cluster for return period TT–

“Quantile regression” here means one regression per quantile. The method does not fit a frequency distribution at each gauge and interpolate its parameters; the at-site quantiles are the response variable, and each return period gets its own equation. That lets the shape of the flood frequency curve change with catchment size and rainfall, which a single growth curve cannot do.

South Africa is divided into 41 homogeneous clusters, numbered 1 to 42 with no cluster 7. The cluster is found by a point-in-polygon lookup at the catchment centroid. In the NaFSAT 2.0 layers the QRT clusters use the same polygons as Index Flood and PRM, with the QRT’s own coefficients. Several clusters share a coefficient set: the 41 clusters carry 31 distinct sets, and clusters 30 and 31 are one example.

The QRT was developed in Calitz’s (2020) doctoral thesis, which tabulates the coefficients for every cluster and return period; the homogeneous clusters are those of Calitz and Smithers (2020). It post-dates the SANRAL Drainage Manual (6th ed., 2013), which does not include it. Cite Calitz (2020) and NaFSAT 2.0 when you report a QRT result. The NaFSAT tool labels its QRT output “Smithers and Calitz, 2020”; that is the same method.

When to use the QRT

The QRT is built for rural catchments without a usable gauge record, broadly 10 to 10 000 km², that sit inside one cluster. Its value is independence. The Rational, SDF and unit hydrograph methods all start from design rainfall and a runoff coefficient or loss model; the QRT starts from observed floods and never touches a coefficient table. When a QRT estimate lands close to the SDF or unit hydrograph result for the same catchment, the two errors are unlikely to be the same error, and that agreement supports the design value.

The limitations below set out where it should be used with care or not at all. When you report a QRT result, give the cluster, the three input values and the coefficient row for each return period. With those a reviewer can reproduce every quantile without the software.

Inputs and where they come from

  • Area: from a watershed delineation, an uploaded catchment polygon or your own GIS. The calculator warns above 50 000 km²; the NaFSAT guidance treats about 10 to 10 000 km² as the reliable range.
  • DR10%: from the design rainfall tool at 24 hours and 1:10 years, at the centroid. The calculator warns above 500 mm.
  • Distance to the coast: computed from the centroid in decimal degrees. Within 0.05 degrees of the coast the calculator warns that the regression is less reliable.
  • Cluster: found automatically at the centroid.

Worked example

A 26.45 km² catchment east of Pretoria has its centroid in QRT cluster 31, 4.196 decimal degrees from the coast. Its 1:10-year 24-hour point design rainfall, DR10%, is 97.8 mm. Cluster 31’s coefficients and the resulting peaks are:

Return periodCONSTTCONST_TaTa_TbTb_TcTc_TQTQ_T (m³/s)
1:22.030.43−0.260.009.46
1:52.280.48−0.18−0.1616.41
1:102.330.52−0.15−0.2320.41
1:202.290.55−0.12−0.3022.45
1:502.150.61−0.09−0.3923.95
1:1001.990.65−0.06−0.4624.15
1:2001.800.69−0.03−0.5423.29

At 1:10, with ln⁡A\ln A = 3.2752, ln⁡DR10%\ln DR_{10\%} = 4.5829 and ln⁡Dc\ln D_c = 1.4341:

ln⁡Q10=2.33+0.52(3.2752)−0.15(4.5829)−0.23(1.4341)=3.0159\ln Q_{10} = 2.33 + 0.52(3.2752) - 0.15(4.5829) - 0.23(1.4341) = 3.0159 Q10=e3.0159=20.4 m3/sQ_{10} = e^{3.0159} = 20.4 \text{ m}^3/\text{s}

Reading the result

At 1:10 the QRT agrees with the Index Flood method on the same catchment (19.3 m³/s). Above 1:20 the QRT curve flattens and the 1:200-year value comes out 3.6 % below the 1:100-year one. The reason is in the last column: ln⁡Dc\ln D_c is positive this far inland, and cTc_T grows more negative with return period, so the distance term subtracts more from the upper quantiles. Index Flood gives 47.6 m³/s at 1:100 and the probabilistic rational method 97.6 m³/s.

A falling or flat upper tail is a signal to stop leaning on the QRT for that catchment’s rare floods. Report its 1:2 to 1:20-year values as a check, and take the 1:50 to 1:200-year design values from the other methods. This catchment is also about two-thirds urban, which the QRT cannot see.

The coefficients also show how strongly each descriptor matters. At 1:100 in this cluster, doubling the area raises the peak by 20.652^{0.65}, a factor of 1.57, while a 10 % change in DR10% moves it by less than 1 %.

Limitations

  • The QRT applies to South Africa only, and the Studio calculator runs for South African sites.
  • It is calibrated on rural gauged catchments. Urban cover, large reservoirs and floodplain storage are not represented.
  • Reliability drops outside about 10 to 10 000 km² and close to the coast.
  • Cluster boundaries are sharp polygons through gradual change. Near a boundary, run the neighbouring cluster and report the difference.
  • At inland sites the upper quantiles can flatten or fall, as in the example. Check that QTQ_T rises with TT before using the tail.
  • The coefficients rest on the historical flood record and assume it is stationary.
  • The QRT gives peaks only. A routing or storage study needs a hydrograph from another method, such as the unit hydrograph generator.
  • Where a reliable record exists at or near the site, at-site flood frequency analysis should govern; the stream gauge finder shows what records exist.

What the full Studio tool adds

The calculator draws the cluster map, takes the centroid from a click, coordinates, an uploaded polygon or a saved watershed, and computes the coast distance and cluster for you. The result lists the cluster and, for each return period, the constant and three coefficients it substituted, so a reviewer can reproduce every quantile by hand. Calculations save to your projects and export to CSV, Word and Excel.

The QRT can also be ticked on a design flood estimation request. There it runs on the delineated catchment with DR10% from the request’s design rainfall, next to Index Flood, PRM and the deterministic methods, and you can override the cluster, DR10% and coast distance where you have reason to.

FAQ

Common questions

What is the quantile regression technique in flood hydrology?
It is a regional flood frequency method. At gauged sites in a homogeneous region, the flood for each return period is estimated from the record, and a separate log-linear regression is fitted for each quantile against catchment descriptors. The fitted equations are then applied to ungauged catchments in the same region. The South African version, in NaFSAT 2.0, uses area, DR10% and distance to the coast.
What is DR10%?
DR10% is the 1:10-year, 24-hour point design rainfall depth at the catchment centroid, in millimetres. In the QRT it is the rainfall predictor for every return period, not only the 1:10-year flood. Take it from the 24-hour, 1:10-year depth of a design rainfall analysis, not from the depth at the time of concentration, and check it against the national design rainfall grid.
Why is the distance to the coast in decimal degrees?
The regressions were calibrated with the planar distance from the centroid to the nearest coastline measured in decimal degrees, and the NaFSAT tool computes it the same way. Converting it to kilometres would put a different number into the logarithm and change every quantile. The calculator clamps the distance to at least 0.01 degrees so the logarithm stays finite at the coast.
How does the QRT compare with the Index Flood method?
Both are NaFSAT 2.0 regional methods calibrated on gauged floods. Index Flood estimates a mean annual flood from area, MAP and coast distance and multiplies it by a cluster growth curve. QRT fits a separate regression for each return period. They often agree near the 1:10-year flood and drift apart in the upper tail, so run both and report the difference.
Why is my QRT 1:200-year flood lower than the 1:100-year flood?
Each return period has its own regression, and nothing ties them together, so Q_T need not rise with T. Far inland the distance to the coast is greater than one degree, so its logarithm is positive, and in several clusters the coast coefficient becomes more negative as the return period rises. The distance term then pulls the upper quantiles down until the curve flattens or falls. This is how the published regressions behave: Calitz (2020) uses DR10% as the rainfall predictor at every return period, the NaFSAT 2.0 tool gives the same numbers, and Calitz notes that quantile regression can give an inconsistent growth curve. Do not adopt the tail; lean on Index Flood, PRM and the deterministic methods.
Can the QRT be used for urban catchments?
Not on its own. The QRT was calibrated on predominantly rural gauged catchments and has no term for impervious cover or piped drainage. For an urbanised catchment use the Rational or SCS-SA methods, which respond to land cover, and treat the QRT result as a rural baseline.
Full documentation: QRT Calculator
Every input, option and equation in the Studio tool, with its limitations.
Run it from your AI agent
This engine is open to Claude, ChatGPT, Cursor and other MCP clients as run_qrt. The agent passes the inputs, the engine does the calculation, and the record is saved to your project for you to review.

References

  1. Calitz, J.P. (2020) Development and assessment of regionalised approaches to design flood estimation in South Africa. PhD thesis, University of KwaZulu-Natal, Pietermaritzburg.
  2. Calitz, J.P. and Smithers, J.C. (2020) Development and Assessment of Regionalised Approaches to Design Flood Estimation in South Africa. WRC Report No. 2748/1/20, Water Research Commission, Pretoria.
  3. Maharaj, U. (2026) NaFSAT Beta Version 2.0 User Guide. GroundTruth, report GTE786/20022026/02, for the National Flood Studies Programme.
  4. Smithers, J.C. and Schulze, R.E. (2003) Design Rainfall and Flood Estimation in South Africa. WRC Report No. 1060/1/03, Water Research Commission, Pretoria.
  5. Smithers, J.C. (2012) Methods for design flood estimation in South Africa. Water SA, 38(4), 633–646.
  6. SANRAL (2013) Drainage Manual, 6th ed. South African National Roads Agency SOC Ltd, Pretoria.

Last reviewed by the Delta Hydro engineering team.

Keep the whole study in one place

Delta Hydro Studio puts this tool alongside catchment, rainfall, runoff and hydraulics tools, with projects, team sharing and Word and Excel reports. Run them by hand, or let your AI agent run them for you to review.