Find your next learning step
Audit a CFAR threshold by regime and through time, so one acceptable average cannot hide local false-alarm failure.
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 · Define the audit
Route 2 · Read the ledger
Route 3 · Monitor drift
Route 4 · Respect the limits
Route 5 · Verify and continue
CHAPTER 1 OF 12 · DEFINE THE AUDIT
1. Why an average CFAR rate can hide failure
A constant false-alarm-rate detector is designed around a nominal probability of false alarm, but the word constant is a model property, not a guarantee about every operating regime. A detector can achieve one acceptable average while producing too many false alarms near clutter edges, during rain, at particular ranges or after a receiver change.
Conditional coverage asks a sharper question: within each meaningful condition, how often does a null cell cross the threshold? The condition might be range band, sea state, waveform, scan sector, time block or a measured clutter statistic. The purpose is diagnosis, not the promise of perfect calibration in every tiny subgroup.
Place this audit beside the Radar root, posterior-predictive CFAR calibration and reference-window contamination. It checks whether a threshold that looked calibrated at commissioning remains calibrated where it is used.
CHAPTER 2 OF 12 · DEFINE THE AUDIT
2. Define the null stream before counting alarms
The audit requires cells or trials believed to contain no target. Labels may come from controlled recordings, trusted tracks, exclusion zones, simulation or delayed adjudication. If true targets contaminate the null stream, the estimated false-alarm rate is biased upward.
For condition g, record n_g eligible null trials and k_g threshold crossings. The empirical rate is k_g/n_g. Compare it with the nominal α using a binomial interval or a calibrated sequential procedure. Do not interpret one crossing as independent when neighbouring range–Doppler cells share processing, clutter and threshold estimates.
State the sampling unit. A cell-level target α, a plot-level false alarm, and a scan-level probability of at least one false plot are different quantities. Multiplicity can make a small cell probability correspond to frequent operational alarms.
CHAPTER 3 OF 12 · READ THE LEDGER
3. Worked ledger: four regimes with the same threshold rule
Suppose the nominal cell false-alarm probability is 1%. Four prespecified regimes each supply 400 adjudicated null cells. Their false-alarm counts are 2, 4, 8 and 14.
| Regime | Null cells n | False alarms k | Observed rate | Approximate 95% Wilson interval |
|---|---|---|---|---|
| Near range, calm | 400 | 2 | 0.50% | 0.14%–1.80% |
| Mid range, calm | 400 | 4 | 1.00% | 0.39%–2.54% |
| Far range, calm | 400 | 8 | 2.00% | 1.02%–3.90% |
| Far range, rain | 400 | 14 | 3.50% | 2.10%–5.79% |
Pooling gives 28/1600 = 1.75%. That average already exceeds 1%, but it still understates the operational concentration: the far-range rain regime is at 3.5%, seven times the near-range calm estimate.
CHAPTER 4 OF 12 · READ THE LEDGER
4. Rolling quantiles turn recent null scores into a threshold
Let S_t be a null score whose large values indicate a target. A simple rolling empirical rule sets the next threshold to a high quantile of the most recent clean scores. For α = 1% and a calibration window of 500, the threshold is near the fifth-largest score, with the exact order statistic chosen by the declared finite-sample convention.
A short window adapts quickly but has noisy tail estimates. A long window is stable but mixes old and new clutter regimes. Weighted quantiles can emphasise recent observations, yet their effective sample size may be far smaller than the raw window length.
Exclude guard regions, confirmed targets and cells affected by the same detection being audited. Otherwise the threshold learns from contaminated values and the subsequent coverage estimate is circular.
CHAPTER 5 OF 12 · READ THE LEDGER
5. Calibration maps need uncertainty as well as colour
A coverage map should show n_g, k_g, the estimated rate and uncertainty for every cell of the map. Colour alone encourages over-reading sparse regions. Prespecify the grouping variables and minimum sample sizes before looking for hot spots.
| Audit view | Signal it reveals | Required caution |
|---|---|---|
| Range-band rate | loss of calibration with range | unequal target-screening quality |
| Weather-conditioned rate | clutter-regime failure | few rare-weather samples |
| Rolling time rate | gradual drift or step change | overlap creates dependence |
| Score-quantile trace | threshold movement | tail quantiles are noisy |
| Alarm run length | bursts and persistence | not a binomial count |
Use a neutral state for estimates whose intervals include the target, a warning state for persistent material excess and an insufficient-data state for sparse groups. Do not colour every random fluctuation as a defect.
CHAPTER 6 OF 12 · MONITOR DRIFT
6. A drift alarm needs a defined decision rule
One practical rule evaluates non-overlapping blocks of null trials. Flag a warning when the lower confidence bound for the false-alarm rate exceeds α in two consecutive blocks, or when a predeclared sequential statistic crosses its boundary. The repeated-block requirement reduces reactions to one noisy window.
Record both the statistical signal and the operational action: freeze automatic threshold updates, widen the diagnostic sample, inspect receiver health, or switch to a validated fallback detector. An alarm without an action path becomes decorative monitoring.
If windows overlap, successive estimates are correlated. A conventional 95% interval at every time point does not create a 5% overall false-alert rate. Calibrate the monitoring rule by simulation under realistic dependence and scan length.
CHAPTER 7 OF 12 · MONITOR DRIFT
7. Changed context: a clutter edge enters the window
Assume a rolling detector was calibrated on homogeneous far-range clutter. A coastline edge then enters half of the reference windows. The cell-under-test distribution and the reference statistic no longer share the commissioned relationship.
If the next 400 adjudicated null cells produce twelve alarms, the observed rate is 3%. The correct response is not merely to raise every threshold. First split the data by edge geometry and check whether the excess is local, whether target masking also increased, and whether the selected reference cells remain representative.
The transfer test is two-sided: after any repair, verify false alarms and probability of detection with injected or labelled targets. A threshold can restore nominal false alarms by becoming so conservative that useful targets disappear.
CHAPTER 8 OF 12 · MONITOR DRIFT
8. Conditional checks complement predictive PIT checks
Probability-integral-transform diagnostics examine whether predictive probabilities behave like their reference distribution. Conditional false-alarm ledgers examine a specific tail event. They should agree, but they answer different questions.
A nearly uniform aggregate PIT can coexist with an upper-tail excess in one regime and a deficit in another. Conversely, a 1% exceedance rate can hide badly shaped predictive probabilities below the threshold. Use the posterior-predictive calibration route to connect both views.
Where drift makes old calibration data less representative, recent or similarity-weighted quantiles may help. Weighting is not magic: disclose the weights, effective sample size and conditions under which guarantees weaken.
CHAPTER 9 OF 12 · RESPECT THE LIMITS
9. Limits and common misconceptions
Conditional coverage cannot be guaranteed for every possible subgroup without structure and enough data. Testing many discovered slices creates false discoveries. Groups should be operationally meaningful and declared before the audit.
The observed alarm rate is not automatically a pure estimate of Pfa when null labels are uncertain, cells are dependent or thresholds share reference data. Wilson intervals handle binomial sampling better than a simple Wald interval, but they do not solve those design problems.
CFAR does not mean a fixed number of false plots per hour. Cell count, correlation, clustering, plot extraction and track confirmation all sit between cell-level Pfa and operator workload.
CHAPTER 10 OF 12 · RESPECT THE LIMITS
10. Evidence and reproducibility route
The engineering purpose of adaptive thresholding and false-alarm control is set out in the IET chapter Constant False Alarm Rate Detectors. A modern detector study that explicitly derives conditional false-alarm probabilities under reference statistics is A Nonlinear Transform-Based Variability Index CFAR Detector for Doppler-Extended Targets.
For weighted quantiles under drift, Conformal Prediction Beyond Exchangeability gives a rigorous account of weighting and the cost of nonexchangeability. The connection here is methodological: it motivates transparent recency weighting, not a claim that an arbitrary CFAR stream inherits conformal guarantees.
Archive null-label rules, score definition, reference-window geometry, target α, order-statistic convention, block length, dependence assumptions, alert boundary, exclusions and software version. Retain the raw k and n behind every plotted rate.
CHAPTER 11 OF 12 · VERIFY AND CONTINUE
11. Frequently asked questions
Why not inspect only the overall false-alarm rate?
Because opposite errors can cancel. One regime may be too conservative while another produces an operationally expensive excess.
How many null cells are enough?
Enough depends on α, dependence and the precision required. Rare tail probabilities need far more effectively independent trials than their percentage notation suggests.
Should a drift alarm automatically raise the threshold?
No. First identify whether the cause is clutter change, contamination, hardware drift, label error or processing change, and then verify detection loss as well as false alarms.
Can rolling quantiles guarantee constant Pfa?
Only under stated assumptions and a validated implementation. Drift, dependence, contaminated calibration data and small effective sample size can all break nominal behaviour.
CHAPTER 12 OF 12 · VERIFY AND CONTINUE
12. Final transfer check and connected reading
Recompute the four rates in the worked ledger and explain why pooling hides the far-range rain problem. Then halve the effective sample size to represent dependence and describe how uncertainty changes even though the observed percentages do not.
Continue through the Radar root, Bayesian CFAR thresholds, posterior-predictive calibration, variability-index switching and reference-window contamination. Together they connect threshold theory to ongoing calibration evidence.
Top · Contents · Previous · How X Works