A bright target can exceed a detector or digitiser range. The stored value then clips at a ceiling, and a still brighter scene may produce oversaturation, charge spill, nulls, oscillation or crosstalk in nearby detectors. The maximum digital number is therefore not an ordinary measurement of the true radiance.
Nonlinearity begins before hard saturation in some instruments. A linear calibration may look excellent over moderate radiance while high-radiance residuals grow systematically. Correcting supported nonlinear response is different from inventing values after clipping.
The Gold Standard is to retain quality flags and raw counts, map the supported response range by detector and band, fit only with independent calibration levels, quarantine affected neighbours and refuse quantitative recovery where information has been lost.
A practical orientation
Use this guide when fires, sunglint, clouds, city lights or reflective surfaces create unusually high counts. The worked example uses a 12-bit detector, three calibration levels and a high-radiance quadratic residual.
Choose a reading route
Understand the detector boundary
Chapters 1–3 separate clipping, oversaturation and nonlinearity.
Build calibration evidence
Chapters 4–6 map response, flags and neighbouring-pixel effects.
Work and transfer the example
Chapters 7–9 calculate a correction and change exposure and scene.
Chapter contents
- 1. Separate radiance, response and stored value
- 2. Recognise ordinary clipping
- 3. Distinguish oversaturation and blooming-like spill
- 4. Test linearity across radiance levels
- 5. Fit nonlinear response only inside support
- 6. Map detector, band and time dependence
- 7. Worked example: curvature before a 12-bit ceiling
- 8. Propagate the correction into physical products
- 9. Transfer check: change exposure, band and scene type
- 10. Use bounds and alternate observations for saturated targets
- 11. State limits and preserve evidence
- 12. FAQs and the next reading route
1. Separate radiance, response and stored value
Scene radiance reaches the instrument, the detector converts photons into charge or another electrical response, electronics amplify and offset that response, and the analogue-to-digital converter stores a finite code. Each stage has its own range and uncertainty.
A Level-1 product may scale digital numbers back to radiance, but that operation cannot restore information the detector never recorded. Preserve the raw code, calibration coefficients, detector identity and flags so a derived radiance can be traced to its supported range.
| Layer | Quantity | Possible boundary |
|---|---|---|
| Scene | Spectral radiance | Target outside design range |
| Detector | Charge or voltage | Full well or nonlinear gain |
| Electronics | Amplified signal | Clipping or recovery transient |
| Digitiser | Integer code | Maximum representable DN |
| Product | Radiance/reflectance | Flagged or invalid conversion |
2. Recognise ordinary clipping
If a 12-bit digitiser represents codes from 0 to 4095, every response above its upper threshold can be stored as 4095. The identical code can therefore represent many true radiances. It is a lower bound on brightness, not a precise endpoint.
Count saturated pixels by band and detector, retain the quality bit and inspect whether the plateau is spatially consistent with a bright target. Do not average clipped values as though they were unbiased measurements.
- Maximum-code frequency.
- Band and detector identity.
- Quality flag.
- Spatial and temporal context.
- No exact inversion beyond clipping.
3. Distinguish oversaturation and blooming-like spill
At still higher input, detector physics or readout electronics can produce null values, oscillations, persistence or crosstalk. Charge generated in a saturated pixel may affect adjacent pixels, creating a distorted footprint larger than the true bright source.
The affected neighbourhood depends on sensor design and readout direction. Use instrument documentation and anomaly flags rather than applying a universal one-pixel buffer. A dark or null streak beside a fire can be an instrument artefact, not a cold surface.
| Pattern | Possible mechanism | First evidence |
|---|---|---|
| Flat maximum | Clipping | DN ceiling and flag |
| Null after bright target | Recovery failure | Readout order and anomaly record |
| Adjacent bright trail | Charge spill or crosstalk | Neighbour/detector pattern |
| 0–max oscillation | Oversaturation response | Raw sequence and band comparison |
4. Test linearity across radiance levels
A linear calibration uses response equal to gain times radiance plus offset. Test residuals at several radiance levels spanning the supported range. A small residual at the midpoint does not rule out curvature near the high end.
Use onboard diffusers, lamps, lunar observations or stable terrestrial sites according to the instrument programme. Separate changes in the calibration source from changes in the detector.
| Calibration level | Sensitivity | Failure missed alone |
|---|---|---|
| Low | Offset and dark response | High-end compression |
| Middle | Nominal gain | Endpoint curvature |
| High, unsaturated | Nonlinearity | Low-end bias |
| Near ceiling | Onset of clipping | Source uncertainty |
5. Fit nonlinear response only inside support
A quadratic term can describe modest gain curvature when independent calibration levels support it. The coefficient must be estimated on unsaturated observations and validated on separate scenes or calibration events.
A better fit does not make extrapolation safe. Once the response is clipped or unstable, several radiances produce the same or corrupted code. Flagging and alternative observations are more honest than a polished invented value.
- Fit on unsaturated data.
- Retain linear comparator.
- Validate on independent levels.
- Propagate coefficient uncertainty.
- Do not extrapolate through clipping.
6. Map detector, band and time dependence
Focal-plane detectors can age differently. Track residuals by detector and band rather than averaging them into one sensor coefficient. A mean correction can improve the fleet while worsening the outlying detector that matters for one swath position.
Close calibration regimes when gains, electronics, processing or anomaly masks change. Historical scenes should be reproducible with the coefficient set valid at acquisition time.
| Dimension | Why it matters | Required record |
|---|---|---|
| Detector | Individual responsivity | Detector index |
| Band | Different material and range | Spectral band |
| Time | Ageing and updates | Validity interval |
| Temperature | Electronic response | Instrument state |
| Processing version | Different flags/correction | Product lineage |
7. Worked example: curvature before a 12-bit ceiling
The released linear calibration predicts DN equal to 3.2 times radiance plus 64. At radiances 200, 600 and 1000 arbitrary units, predicted codes are 704, 1984 and 3264. Observed means are 702, 1968 and 3190, showing a growing negative residual at high radiance.
A quadratic correction fitted to independent unsaturated levels reduces held-out residuals from minus 2.3% to minus 0.3% at radiance 1000. At radiance 1250, the linear prediction is 4064 and the observed code is 3950, still below the 4095 ceiling but close enough that uncertainty and detector variation matter.
At radiance 1300, the linear prediction is 4224 and the stored value is 4095. That pixel is saturated. The quadratic model must not be used to claim an exact radiance, because the code no longer uniquely identifies the input.
| Radiance | Linear DN | Observed DN | Interpretation |
|---|---|---|---|
| 200 | 704 | 702 | Near-linear |
| 600 | 1984 | 1968 | Small compression |
| 1000 | 3264 | 3190 | Supported nonlinearity |
| 1250 | 4064 | 3950 | Near ceiling; high uncertainty |
| 1300 | 4224 | 4095 | Clipped; no exact inversion |
8. Propagate the correction into physical products
A radiance correction can change reflectance, brightness temperature, indices, thresholds and retrieval inputs. Reprocess a representative product set and report which downstream values change materially, not only the detector residual.
Keep saturated pixels masked through later calculations unless the downstream algorithm explicitly handles bounds. A vegetation index computed from one saturated and one valid band can appear numerically ordinary while carrying a broken ratio.
| Downstream quantity | Risk | Check |
|---|---|---|
| Reflectance | Biased scale | Reference sites |
| Brightness temperature | Nonlinear Planck conversion | Band-specific propagation |
| Band ratio/index | One-sided clipping | Per-band flags |
| Classification | Threshold crossing | Changed-label audit |
9. Transfer check: change exposure, band and scene type
Shorten exposure or use a lower-gain acquisition of the same stable target. Valid radiances should agree within uncertainty after calibration, while formerly clipped pixels may return to range. Change band and confirm that the correction is not copied across detectors with different response.
Compare bright cloud, desert, sunglint and fire scenes. Their spectra and spatial patterns differ. A correction trained on a smooth diffuser may not describe a transient oversaturation artefact caused by an intense compact source.
| Changed context | Expected invariant | Failure exposed |
|---|---|---|
| Shorter exposure | Same scene radiance | Clipping or persistence |
| Different band | Independent calibrated value | Copied coefficient |
| Diffuse to compact source | Supported response only | Blooming/crosstalk |
| New processing version | Reproducible lineage | Silent flag change |
10. Use bounds and alternate observations for saturated targets
For a clipped pixel, report that true radiance exceeds the saturation threshold after accounting for calibration uncertainty. If another band, shorter exposure, sensor or time is unsaturated, it may constrain the target through a separate physical model.
Do not present a cross-sensor estimate as recovered truth without its spectral, temporal and geometric assumptions. A wildfire can change between acquisitions, and sunglint geometry can change faster than the revisit.
- Lower bound retained.
- Quality mask propagated.
- Neighbourhood anomaly buffer justified.
- Alternate observation assumptions stated.
- No exact value from a many-to-one code.
11. State limits and preserve evidence
A nonlinear correction cannot repair saturation, charge spill or missing samples unless an independently validated instrument-specific method exists. A saturated pixel may also contaminate neighbours that never reached the maximum code.
Preserve raw counts, calibration events, source radiance and uncertainty, detector/band index, temperature, readout order, coefficient versions, residuals, flags, affected-neighbour rules, downstream tests and exclusions.
- Maximum DN is not maximum radiance.
- Unflagged neighbour can still be affected.
- Linear residual is not automatically detector ageing.
- Correction range must be declared.
- Historical product remains reproducible.
12. FAQs and the next reading route
Remote sensing depends on knowing not only the value but also whether the instrument was capable of measuring it. Saturation is a statement about that capability boundary; nonlinearity is a calibration problem inside the boundary.
Read the Remote Sensing root, the stored-value guide and the on-orbit ageing guide alongside this article. The Gold Standard parent and How X Works hub connect the same practice across subjects.
Can I convert DN 4095 to an exact radiance?
Not when 4095 is the clipping ceiling. It represents a range of inputs above the supported threshold.
Is blooming the same as saturation?
No. Saturation is loss of response range; blooming-like spill or crosstalk affects neighbouring samples or detectors.
When is a quadratic correction acceptable?
When it is fitted and validated on unsaturated calibration levels within a declared detector, band and time range.
Continue through the verified ecosystem
Sources and scope note
- NASA — Nonlinear Detector Response of Aqua MODIS Land Imaging Bands — used for high-radiance gain nonlinearity, multi-level calibration and quadratic correction within supported response.
- USGS Landsat Missions — Oversaturation — used for clipping, nulls, oscillation and crosstalk patterns over very bright targets.
- USGS Landsat Missions — Landsat Radiometry — used for detector responsivity, uncertainty and ongoing radiometric calibration.
The 12-bit detector, calibration coefficients and radiances are original teaching examples. NASA and USGS sources support the instrument-response principles; they do not validate the example coefficients.