The instrument transmits C-band microwave energy and measures the return. Illumination comes from the sensor, not the Sun.
Field Lab 08 · Sentinel-1 SAR
Map wetland inundation without confusing dark pixels for proof
Build a traceable Sentinel-1 evidence package from an ecological question to a validated candidate-inundation map. The lab teaches decisions, failure gates and uncertainty—not a one-click flood colour ramp.
The downloadable table and interactive records are deliberately synthetic. They reproduce realistic Sentinel-1 metadata and QA problems but are not real flooding, real Baltic plots or observations from the 2024 Estonia campaign.
Scientific task
Can a Sentinel-1 time pair support a defensible map of candidate event inundation?
Your answer must preserve the difference between a sensor measurement, a classification rule and a hydrological interpretation. You will finish with a map class, confidence layer, validation record and list of non-claims.
Before software
What Sentinel-1 contributes
Pixel values combine dielectric properties, roughness, structure and acquisition geometry. They are not direct water depth or soil moisture.
VV and VH describe transmit–receive orientation. Different channels can respond differently to surface and volume scattering.
Speckle is inherent to coherent imaging. Filtering changes spatial support and must be recorded rather than treated as cosmetic cleanup.
TRANSMITC-band pulse
→INTERACTwater · soil · stems · terrain
→RETURNamplitude / power
→DERIVEbackscatter + class
Measurement scale
Linear power and decibels are two representations
Move the same positive linear backscatter value through the logarithmic conversion. The physical evidence has not changed; its numerical representation has.
10 × log10(0.030)-15.23 dB
Average in linear space when the method calls for mean power. A mean of dB values is a different statistic.
Decision lab · synthetic training records
Would you compare this observation?
Classify each row before revealing the audit. “Review” means the row needs evidence or a different design—not that it is automatically unusable.
SAL0104 Jun
- Geometry
- ASC · 131
- Incidence
- 38.2°
- VV / VH
- −8.4 / −15.7 dB
SAL0316 Jun
- Geometry
- ASC · 131
- Incidence
- 38.3°
- VV / VH
- −7.7 / −14.8 dB
SAL0528 Jun
- Geometry
- DESC · 36
- Incidence
- 33.1°
- VV / VH
- −10.2 / −17.0 dB
SAL0610 Jul
- Geometry
- ASC · 131
- Incidence
- 45.9°
- VV / VH
- −8.0 / −14.5 dB
SAL0722 Jul
- Geometry
- ASC · 131
- Incidence
- 38.5°
- VV / VH
- −6.9 / −13.9 dB
0 / 5 decided
End-to-end procedure
Question → scenes → measurement → map → evidence
Work in order. Each gate names the required input, the product you should have at the end and the condition that stops the workflow.
Write the claim before choosing scenes
Define the surface-water or inundation change that the map is allowed to represent.
ACTION
- Draw the area of interest.
- Choose reference and event time windows.
- Name the map unit, minimum mapping unit and validation source.
- List permanent water, radar shadow, built surfaces and emergent vegetation as possible confounders.
WHERE
Project brief and GIS study-area layer.
WHY
A dark-backscatter map is not automatically a flood map. The claim determines the observations, exclusions and validation evidence needed.
INPUT
- Ecological or hazard question
- AOI
- Event chronology
- Independent reference plan
OUTPUT
- One-sentence claim
- AOI
- Reference/event windows
- Predeclared non-claims
CHECK
- The event window follows the event trigger.
- The reference represents a defensible non-event condition.
- The proposed validation is independent of the threshold-tuning samples.
IF THIS FAILS
The event timing, target class or independent validation source is unknown.
Build a comparable Sentinel-1 scene stack
Select GRD observations whose acquisition geometry and polarization support comparison.
ACTION
- Search Sentinel-1 GRD over the AOI and both time windows.
- Record platform, sensing time, mode, product type, polarization, orbit direction, relative orbit and incidence-angle range.
- Prefer the same mode, polarization, orbit direction and relative orbit for a simple change comparison.
WHERE
Copernicus Data Space Browser or catalogue.
WHY
Backscatter changes with look geometry as well as surface conditions. A date match alone is not a comparability test.
INPUT
- AOI
- Time windows
- Sentinel-1 catalogue metadata
OUTPUT
- Scene inventory
- Candidate reference/event stack
- Rejection log
CHECK
- AOI coverage is complete.
- Required polarization exists on every selected date.
- Orbit direction and relative orbit are consistent or the difference is explicitly modelled.
IF THIS FAILS
Coverage, mode, polarization or geometry differs in a way the analysis cannot defend.
Preserve the source products
Create an immutable input inventory before processing.
ACTION
- Download or reference the exact catalogue items.
- Record product identifiers, checksums when available, retrieval date and processing baseline.
- Keep inputs, processing intermediates and outputs separate.
WHERE
Project data register and read-only inputs folder.
WHY
A defensible map must be traceable to exact source observations and processing assumptions.
INPUT
- Selected GRD products
- Catalogue records
OUTPUT
- Input manifest
- Unchanged source products
- Processing workspace
CHECK
- Every raster has an exact product identifier.
- No processing is performed inside the source archive.
IF THIS FAILS
A scene cannot be traced to a stable identifier and acquisition record.
Calibrate and terrain-correct consistently
Apply one documented GRD preprocessing chain to every accepted scene.
ACTION
- Inspect the product and chosen service before applying operators already included upstream.
- Apply orbit information and border/thermal-noise handling where required.
- Calibrate every scene to the same backscatter coefficient.
- Apply the chosen speckle strategy before resampling when used.
- Orthorectify or use an analysis-ready RTC product, record the DEM and coefficient, then clip only after geometry is stable.
WHERE
SNAP graph, Copernicus Data Space processing service or another reproducible SAR processor.
WHY
Calibration addresses measurement scale; terrain correction addresses location and geometry. They solve different problems.
INPUT
- GRD products
- Orbit data when required
- DEM
- Processing graph and parameters
OUTPUT
- Co-registered calibrated backscatter rasters
- Processing log
- QA quicklooks
CHECK
- Coefficient and units match across dates.
- Projection, grid, pixel spacing, extent and NoData are explicit.
- No operator was duplicated merely because it appears in a generic recipe.
IF THIS FAILS
Dates have different coefficients, grids or unrecorded processing histories.
Keep linear power and dB mathematically distinct
Use the value scale required by the operation and label it correctly.
ACTION
- Confirm whether the raster stores linear power or decibels.
- For positive linear power, calculate dB = 10 × log10(power).
- Average power in linear space unless the method explicitly defines another statistic.
- Never relabel a linear raster as dB.
WHERE
Raster calculator or analysis notebook.
WHY
The logarithmic transform changes arithmetic and visual contrast. Ratios, means and thresholds are not interchangeable across scales.
INPUT
- Calibrated backscatter
- Scale metadata
OUTPUT
- Explicit linear or dB analysis layers
- Scale check table
CHECK
- Linear values are non-negative.
- Typical low-return areas become more negative in dB.
- No zero or negative linear value enters the logarithm.
IF THIS FAILS
The stored scale or coefficient cannot be established from metadata.
Create one comparison grid
Align the reference and event evidence without inventing spatial detail.
ACTION
- Choose a target CRS, grid origin, extent and pixel spacing.
- Resample once using a method appropriate to continuous backscatter.
- Create a common valid-data mask.
- Record whether spatial averaging or filtering changes effective support.
WHERE
GIS or scripted raster pipeline.
WHY
Pixel-wise change is meaningful only when pixels refer to compatible ground support.
INPUT
- Preprocessed dates
- AOI
- Terrain and permanent-water masks
OUTPUT
- Aligned stack
- Common mask
- Grid specification
CHECK
- Exact shape, transform, CRS and NoData match.
- No visible half-pixel shift occurs at stable boundaries.
IF THIS FAILS
Rasters overlap visually but fail an exact grid comparison.
Estimate a candidate-water rule
Use labelled evidence to estimate a rule; do not select a threshold from map appearance alone.
ACTION
- Separate tuning samples from validation samples.
- Inspect VV, VH and change distributions by class and land-cover context.
- Choose a transparent threshold or classifier.
- Record the feature, scale, operator direction and threshold.
WHERE
Notebook plus GIS sample layer.
WHY
Smooth open water often returns little energy to the sensor, but wind roughening, emergent vegetation, soil moisture, radar shadow and double-bounce can violate a simple dark-water rule.
INPUT
- Aligned backscatter stack
- Training labels
- Context masks
OUTPUT
- Candidate classification rule
- Tuning diagnostics
- Model record
CHECK
- Threshold units match the source scale.
- Reference and event data were processed identically.
- Permanent water and terrain shadow are handled explicitly.
IF THIS FAILS
The rule depends on the validation labels or collapses under plausible threshold changes.
Separate water evidence from flood interpretation
Apply chronology and exclusions to distinguish candidate event inundation from permanent water or radar artefacts.
ACTION
- Compare event candidates with the reference water state.
- Exclude or label permanent water separately.
- Flag radar shadow, layover, edge artefacts and low-confidence vegetation contexts.
- Apply the minimum mapping unit only after documenting its effect.
WHERE
Raster logic and GIS review.
WHY
Flooding is a temporal interpretation, not merely a low-backscatter state.
INPUT
- Candidate-water layers
- Reference state
- Permanent-water/context masks
- Event chronology
OUTPUT
- Candidate event-inundation layer
- Exclusion mask
- Confidence layer
CHECK
- Persistent water is not counted as newly inundated.
- Masked pixels remain distinguishable from classified dry pixels.
IF THIS FAILS
The map cannot distinguish new water, permanent water and no-data/uncertain areas.
Measure omission and commission separately
Validate the candidate map against independent reference observations.
ACTION
- Use a spatially and temporally relevant reference sample.
- Build a confusion matrix.
- Report class precision, recall and intersection-over-union or F1 alongside sample counts.
- Stratify errors by open terrain, vegetation, settlement and radar geometry when sample size allows.
WHERE
Validation notebook and map review.
WHY
High overall accuracy can hide failure on a rare inundated class.
INPUT
- Held-out validation labels
- Candidate map
- Stratum/context attributes
OUTPUT
- Confusion matrix
- Class metrics
- Mapped error samples
CHECK
- Validation samples were not used for threshold selection.
- Reference timing supports the event claim.
- Sample counts accompany every metric.
IF THIS FAILS
Validation is circular, temporally mismatched or too sparse for the published claim.
Run sensitivity and uncertainty checks
Expose the decisions that control mapped area and error.
ACTION
- Sweep a defensible threshold interval.
- Compare with and without the selected speckle or spatial filter.
- Test alternative context masks and minimum mapping units.
- Map stable classifications separately from decision-sensitive pixels.
WHERE
Reproducible analysis script.
WHY
A single crisp boundary can conceal sensitivity to threshold, filtering, reference timing and mixed pixels.
INPUT
- Candidate workflow
- Alternative parameter set
- Validation data
OUTPUT
- Sensitivity table
- Agreement/confidence raster
- Qualified interpretation
CHECK
- The reported inundated area includes an uncertainty or sensitivity statement.
- Parameter alternatives are scientifically plausible, not arbitrary.
IF THIS FAILS
Small defensible changes reverse the main conclusion without being reported.
Export an auditable evidence package
Deliver the map together with enough context to reproduce and challenge it.
ACTION
- Export calibrated reference/event layers, final class, confidence and masks as GeoTIFF or COG.
- Save scene inventory, processing graph, threshold record, confusion matrix and sensitivity table.
- Write a short interpretation with supported claims and explicit non-claims.
WHERE
outputs, qa and documentation folders.
WHY
A flood-coloured raster without provenance, validation and uncertainty is not analysis-ready evidence.
INPUT
- Accepted processing outputs
- QA records
- Validation and sensitivity results
OUTPUT
- Analysis-ready raster package
- Machine-readable manifest
- Decision log
- Portfolio brief
CHECK
- Every output states coefficient, scale, CRS, grid, time, polarization and NoData.
- The final map opens correctly outside the processing project.
IF THIS FAILS
A reviewer cannot reconstruct which scenes, coefficient, threshold and exclusions created the result.
Interpretation boundary
Dark at the event date is not enough
Select a sample to see why chronology, context and geometry must accompany a threshold.
Choose A–D. Predict whether it is a candidate, exclusion or review case before reading the explanation.
Analysis-ready handoff
Seven products, one defensible interpretation
Workstation pack
Download the lab evidence
Preserve the source-status column. Your portfolio must call these records synthetic.
Authoritative references