Find your next learning step
Clutter tails change with range, scan mode, weather and geometry. This guide adds covariates to conditional-extremes CFAR without pooling incompatible regimes or chasing unstable threshold effects.
Open the chapter contents · Radar root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · Find the changing regime
Route 2 · Fit conditional covariates
Route 3 · Pool with discipline
Route 4 · Validate and challenge
Route 5 · Check and transfer
CHAPTER 1 OF 12 · FIND THE CHANGING REGIME
1. One CFAR model may hide several clutter regimes
Conditional-extremes CFAR models how neighbouring radar cells behave when an anchor cell is already extreme. A stationary fit assumes that marginal tails and conditional dependence remain the same across observations. In practice they may change with range, azimuth, scan mode, sea state, rainfall, terrain or processing configuration.
Pooling incompatible regimes can create a smooth-looking average that is wrong for every regime. Splitting too aggressively leaves too few exceedances and produces unstable tails. The task is to introduce covariates only where they improve tail calibration and transfer.
The preceding conditional-extremes CFAR guide develops the anchor normalisation and multi-cell return layer. Here the emphasis is non-stationarity and disciplined pooling.
CHAPTER 2 OF 12 · FIND THE CHANGING REGIME
2. Covariates can enter three different layers
A regime variable may change the threshold, marginal tail or dependence model. These are not interchangeable. A higher raw threshold in one range band may simply reflect a changed marginal scale; after standardisation, conditional dependence may be unchanged.
First estimate covariate-dependent high quantiles or transform margins within regime. Then model exceedance magnitude. Only after comparable marginal scales are established should alpha, beta or the residual distribution of the conditional-extremes model vary with covariates.
Jonathan, Ewans and Randell’s non-stationary conditional-extremes work uses covariates across threshold selection, marginal modelling and dependence. It supports the layered workflow; radar deployment still requires domain-specific validation.
CHAPTER 3 OF 12 · FIND THE CHANGING REGIME
3. Worked example: pooling reverses a tail conclusion
Suppose calm and rain regimes each contribute 500 anchor exceedances. In calm conditions, 25 neighbours exceed their threshold, giving a conditional rate of 0.05. In rain, 100 neighbours exceed, giving 0.20. A pooled model reports 0.125.
| Regime | Anchor exceedances | Joint exceedances | Conditional rate | Operational reading |
|---|---|---|---|---|
| Calm | 500 | 25 | 0.050 | Pooled model is too conservative |
| Rain | 500 | 100 | 0.200 | Pooled model is too permissive |
| Pooled | 1,000 | 125 | 0.125 | Correct for neither regime |
A threshold chosen from the pooled distribution may worsen the problem by mixing marginal and dependence changes. Stratify the diagnostic before deciding which model layer needs the covariate.
CHAPTER 4 OF 12 · FIT CONDITIONAL COVARIATES
4. Threshold stability comes before covariate complexity
Fit a ladder of high thresholds within each regime or across a smooth covariate grid. Track marginal shape estimates, conditional alpha and beta, residual quantiles and the final alarm probability. Select the lowest region where the target extrapolation is reasonably stable, not the threshold that produces the preferred false-alarm rate.
If regimes have different data volumes, a common percentile may imply different physical thresholds and effective sample sizes. Report both. Require enough independent clusters rather than counting every correlated cell as a fresh tail event.
The original Heffernan–Tawn paper emphasises stability of conditional extrapolations and residual independence above threshold. Those diagnostics remain necessary after covariates are added.
Smooth effects need regularisation
Continuous covariates such as range or wind speed can enter alpha, beta and marginal parameters through splines or periodic bases. Penalise unnecessary roughness and choose complexity with blocked cross-validation. A wiggly tail surface may be a sparse-data map, not physics.
Use transformations with interpretable limits. Constrain parameters so simulated tails remain coherent. Inspect partial effects with uncertainty bands and residuals, especially near covariate boundaries where extrapolation is easiest to mistake for evidence.
For categorical scan modes, partial pooling can share information while allowing real offsets. Do not force a mode with changed processing gain into the same marginal transform merely because its sample is small.
CHAPTER 5 OF 12 · FIT CONDITIONAL COVARIATES
5. Pool regimes only after invariance checks
Two regimes may share dependence parameters while keeping separate margins, or share beta while alpha changes. Build a hierarchy of candidate models from most pooled to least pooled. Compare tail-focused predictive scores, conditional quantile coverage and alarm-event calibration.
Pooling is justified when parameter differences are small relative to uncertainty and the target probabilities transfer to held-out episodes. Failure to reject a difference with sparse extremes is not proof of equality. Include a practical tolerance tied to CFAR error, not only a significance test.
The tail-copula CFAR guide offers an independent dependence description. Agreement across approaches strengthens a pooling claim.
CHAPTER 6 OF 12 · POOL WITH DISCIPLINE
6. Regime imbalance changes pooled performance
Reuse the calm rate 0.05 and rain rate 0.20. If rain contributes only 10 percent of training data, the pooled conditional rate is 0.065. If deployment has 40 percent rain, the true mixture rate becomes 0.11 even though neither regime changed.
| Rain share | Pooled conditional exceedance rate | Change from 10% training mix |
|---|---|---|
| 10% | 0.065 | Baseline |
| 20% | 0.080 | +0.015 |
| 40% | 0.110 | +0.045 |
| 60% | 0.140 | +0.075 |
This is why deployment monitoring should track regime frequencies as well as residual calibration. A model can drift through composition without any parameter moving within regime.
CHAPTER 7 OF 12 · POOL WITH DISCIPLINE
7. Changed context requires reweighting or refitting
When the operating mix changes, compute the target alarm probability using deployment regime weights. If a new regime lies within the covariate support and diagnostics remain stable, reweighting may suffice. If margins, residuals or processing change, refit and validate.
Never extrapolate a spline far beyond observed range or weather conditions without a conservative fallback. Map the support of every covariate and mark regions where effective tail sample size is low.
The conditional-coverage CFAR guide supplies rolling tests for alarm calibration. Run them by regime so opposing errors do not cancel.
CHAPTER 8 OF 12 · POOL WITH DISCIPLINE
8. Bootstrap whole episodes
Resample independent scans, dwells, storms or declustered episodes, not individual cells. Repeat threshold fitting, marginal transformation, covariate smoothing, dependence fitting and multi-cell simulation. This carries selection and regime uncertainty into the return layer.
Use nested resampling or honest held-out blocks for smoothing choices. Report Monte Carlo error separately from model uncertainty. When rare regimes contain few episodes, scenario bounds may be more honest than a narrow bootstrap interval.
Inspect leverage. One severe storm or interference episode can determine a smooth covariate effect. Refit without each major episode and show how the operational threshold changes.
CHAPTER 9 OF 12 · VALIDATE AND CHALLENGE
9. Validation must be conditional and operational
Hold out contiguous periods, sites or modes. Check probability-integral transforms for margins, independence of normalised residuals from the anchor, conditional quantile coverage and the exact multi-cell alarm event. A good overall score can hide a dangerous rain-regime failure.
Compare cluster size, orientation and lag structure. The extremogram CFAR guide helps choose covariate-sensitive neighbourhoods; the hidden-regular-variation guide challenges residual joint risk after apparent asymptotic independence.
Record performance at the trial unit used operationally: per cell, dwell, scan or track seed. Do not compare rates with different denominators.
CHAPTER 10 OF 12 · VALIDATE AND CHALLENGE
10. Evidence boundaries and common mistakes
The conditional-extremes and non-stationary sources support covariate-dependent tail modelling and threshold diagnostics. They do not establish that a chosen radar covariate is causal or stable after a processing update.
Common mistakes are inserting covariates only in dependence while margins drift, selecting thresholds once on pooled data, treating every cell as independent, fitting overly flexible splines, and validating only the average regime. Another is silently using training mixture weights for deployment.
Regime labels can themselves be uncertain. A weather classifier with errors should be audited because misclassification blends the very tails the model is trying to separate.
CHAPTER 11 OF 12 · CHECK AND TRANSFER
11. Questions readers usually ask
Should every regime get its own model?
No. Separate only the layers that materially differ, then pool what transfers. Partial pooling is often more stable than either complete pooling or complete separation.
Can range be treated as a smooth covariate?
Yes when the effect is smooth and the support is dense enough. Protect boundaries, processing transitions and shadow zones from automatic smoothing.
What if a rare regime has no joint exceedances?
Do not conclude zero risk. Borrow information under explicit assumptions, use conservative bounds or collect more data.
How often should thresholds be rechecked?
After material environment or processing changes and on a schedule tied to accumulated independent tail episodes, not merely calendar time.
CHAPTER 12 OF 12 · CHECK AND TRANSFER
12. A reusable non-stationary CFAR routine
Define regimes and covariates before inspecting tail outcomes. Diagnose thresholds and margins by regime, add covariates to conditional dependence only where needed, compare pooling structures with blocked tail validation, reweight for deployment composition and bootstrap full episodes.
Keep the work connected to the Radar root, the Gold Standard of Education and the How X Works hub. The reusable lesson is to pool only what remains invariant at the tail probability the detector must control.
Top · Contents · Previous · How X Works