Evaluate a UAV Survey Before Scientific Analysis

Chapter 4 Practicum — Evaluate a UAV Survey Before Scientific Analysis

Assessment purpose

This is the Chapter 4 synthesis assessment, not Lesson 2.26. You are the remote-sensing scientist accepting a UAV survey handover for ecological analysis. Your responsibility is to decide which evidence is usable, which needs review and which is unsuitable—not to force every file into one stack.

Allow 7–10 hours across focused sessions. Reuse your mission, radiometric, georeferencing, reconstruction, product and raster-QA functions. Keep raw inputs immutable and preserve every stopped workflow as a documented decision.

Scenario

The research team supplies:

• mission_metadata.csv and image_metadata.csv;
• photogrammetry_report.json;
• separate GCP and withheld check-point residual tables;
• RGB orthomosaic preview;
• Red, Green, Red Edge and NIR bands;
• clean and deliberately damaged DSM variants;
• study boundary and synthetic field polygons;
• calibration-panel availability and irradiance metadata.

The handover contains eight deliberate conditions: variable illumination, one weakly georeferenced area, one band-alignment problem, one orthomosaic seam/ghost, one DSM spike/pit pair, a temporal mismatch warning, inconsistent NoData metadata and one ambiguous reflectance scale.

Every asset is synthetic and CC0. No file contains private imagery, personal information or published Baltic plot coordinates.

Decision principle: An attractive product is not analysis-ready until acquisition, geometry, radiometry, support, timing and provenance meet the intended-use contract.

Check your understandingThe calculated nominal GSD is 2.19 cm. What can the final report claim from that value alone?

Choose one answer

Only the approximate image sampling distance under the stated geometry and assumptionsThe orthomosaic is horizontally accurate to 2.19 cmEvery 2.19 cm object can be detected

Check answer

Required delivery structure

uav_practicum/
├── mission_audit.csv
├── georeferencing_report.csv
├── raster_alignment_report.csv
├── radiometric_qa.csv
├── uav_qa_matrix.csv
├── uav_stack_manifest.csv
├── extraction_table.csv
├── qa_map.pdf
├── UAV_PRODUCT_QA_REPORT.md
└── uav_practicum.ipynb

Store downloaded fixtures under data/raw/. Use data/interim/ only for named diagnostic or alignment candidates. Put accepted derivatives under outputs/. Never overwrite the fixture pack.

Stage 1 — establish identity and integrity

1. Compare every file SHA-256 with manifest.json.
2. Record synthetic/open status and refuse any claim that coordinates are real field sites.
3. Parse every CSV and JSON; open every GeoTIFF and GeoJSON.
4. Record software versions and environment.
5. Separate observed metadata, derived calculation, assumption, decision and missing evidence into distinct columns.

Stop if a checksum changes unexpectedly, a file cannot open, coordinate reference is missing where transformation is required, or product identity cannot be established.

Stage 2 — audit the mission

6. Describe platform category, payload, mission time, positioning and intended products.
7. Calculate sensor pixel pitch and approximate GSD using consistent units.
8. Calculate image footprint dimensions.
9. Calculate nominal forward and side spacing from overlap.
10. Calculate trigger interval at documented speed.
11. State the pinhole, nadir and level-surface assumptions.
12. Record rolling-shutter and motion implications.
13. Compare UAV date with field date and list contextual evidence required to accept the mismatch.

Do not describe nominal GSD as positional accuracy or minimum detectable feature.

Stage 3 — audit images and radiometry

14. Plot exposure, irradiance, saturation and blur by capture time and flight line.
15. Identify the blurred image, exposure transition and high-saturation image.
16. Record panel evidence and its limitations.
17. Compare the clean Red raster with uav_radiometric_gradient_demo.tif using a fixed stretch and column profiles.
18. Inventory the stored quantity, scale, units, NoData and valid range of every band.
19. Classify Red Edge scale as blocking review until authoritative metadata are available.
20. Produce radiometric_qa.csv with consequence and action.

Check your understandingThe Red Edge values resemble reflectance multiplied by 10000, but no authoritative scale is supplied. What is required?

Choose one answer

Keep quantitative Red Edge derivatives blocked and request verified product metadataDivide by 10000 because the pattern is familiarNormalise each image independently to make it look similar

Check answer

Stage 4 — evaluate photogrammetric reconstruction

21. Map each processing-report field to feature matching, tie points, camera model, bundle adjustment or missing downstream evidence.
22. Calculate aligned-image fraction.
23. Interpret the 0.42-pixel reprojection error as internal fitted evidence only.
24. Review the 2.8% focal-length change without applying a universal limit.
25. Link the reported weak region to mission coverage, image conditions and later check-point evidence.
26. List dense-cloud, surface, orthorectification and mosaic settings absent from the handover.

The report passes only if its diagnostic meaning and missing evidence are correctly stated. A low reprojection residual never substitutes for check-point validation.

Stage 5 — assess georeferencing independently

27. Keep GCP and check-point tables separate.
28. Verify point ID, role, residual sign and units.
29. Calculate mean east/north/vertical bias, component RMSE, planimetric RMSE, vertical RMSE and maximum residual.
30. Plot check-point residual vectors with a declared arrow scale.
31. Identify the south-east outlier/local deformation and retain it.
32. Record horizontal and vertical reference limitations.
33. Produce region-specific accept/review/unsuitable decisions in georeferencing_report.csv.

Check your understandingWhy must the large south-east check-point residual remain in the assessment?

Choose one answer

It may reveal local block deformation that directly affects analysis in that regionIt should be removed because it increases RMSEControl-point residuals already prove the edge is correct

Check answer

Stage 6 — audit orthomosaic and DSM

34. Inspect RGB metadata and state that it is an 8-bit display preview, not reflectance.
35. Quantify the brightness difference around the seam and locate ghosted texture.
36. Overlay field polygons and identify affected analytical support.
37. Compare clean and defective DSM rasters under identical styling.
38. Detect the 18.5 m spike and -3 m pit, then inspect local context.
39. Record that the DSM is an upper reconstructed surface with undocumented vertical datum.
40. Do not calculate canopy height because a validated DTM is absent.
41. Add every finding to uav_qa_matrix.csv and the QA map.

Stage 7 — verify multispectral raster integrity

42. Run the complete grid contract across Red, Green, NIR, shifted NIR, Red Edge and DSM.
43. Demonstrate that aligned NIR passes and shifted NIR fails despite matching CRS, resolution and shape.
44. Compare NoData values and create variable-specific masks.
45. Check accepted reflectance-proxy ranges and flag Red Edge scale.
46. Inspect local registration around sharp features in QGIS.
47. Define the target grid and any resampling/correction evidence required.
48. Produce raster_alignment_report.csv.

Do not relabel the shifted band’s transform as a correction. Do not calculate indices until content registration and the grid contract pass.

Stage 8 — build only the accepted spectral subset

49. Use aligned Red, Green and NIR from the synthetic pack.
50. Intersect masks and apply denominator safety.
51. Calculate NDVI and GNDVI in float.
52. Verify finite valid values and the expected mathematical range.
53. Leave Red-edge NDVI blocked and record the required metadata.
54. Integrate the clean DSM only with its vertical limitation.
55. Create a reason-coded QA mask.
56. Reopen and compare every derivative with declared expectations.
57. Produce uav_stack_manifest.csv for accepted, blocked and review layers.

Stage 9 — extract to field support

58. Transform field polygons to EPSG:3301; do not relabel them.
59. Choose and document centre-based, all-touched or area-weighted raster support.
60. Report candidate and valid cells, valid fraction, statistic, units, source checksum, raster QA status and field/UAV dates.
61. Flag plots intersecting seam, weak georeferencing or invalid support.
62. Produce and reopen extraction_table.csv.

Extraction does not validate an index as an ecological proxy. It only creates a traceable link from accepted raster support to synthetic polygons.

Stage 10 — final product decisions

Classify each product and relevant region:

• acceptable — evidence meets the declared use;
• review — bounded missing evidence or correction remains;
• unsuitable — a requirement is not met for this use.

At minimum decide separately for RGB orientation, Red/Green/NIR spectral use, Red Edge, NDVI/GNDVI, DSM upper-surface use, south-east plot overlay and field-date comparison.

Every decision must cite evidence, affected support, scientific consequence and next action. “Looks good” is not acceptable evidence.

Required QA map

qa_map.pdf must show:

• study boundary and synthetic plots;
• orthomosaic seam and ghost region;
• check-point residual vectors;
• weak south-east area;
• DSM spike/pit locations;
• affected or excluded plot support;
• legend, CRS, scale and source note;
• labels or patterns so colour is not the only signal.

Reopen the PDF and verify text, labels and geometry at normal reading size.

Final report structure

UAV_PRODUCT_QA_REPORT.md must include:

1. Mission description
2. Sensor description and measurement quantities
3. Flight geometry, GSD and assumptions
4. Radiometric evidence and unresolved scale
5. GCP/check-point and horizontal/vertical evidence
6. Photogrammetric reconstruction QA
7. Orthomosaic seam/ghost diagnosis
8. DSM interpretation and artefacts
9. Multispectral grid, masks and indices
10. Spatial extraction support
11. Temporal support and field compatibility
12. Known limitations
13. Product/region decisions
14. Corrective action and owners
15. Provenance, software and checksums

Assessment rubric

Dimension	Weight	Full-credit evidence
Mission and sensor reasoning	15%	Correct GSD/footprint/spacing with assumptions; sensor and product meanings remain distinct
Photogrammetry understanding	20%	Complete software-independent chain; internal diagnostics are interpreted within limits
Georeferencing QA	20%	Control/check separation, correct statistics, residual map and local/vertical interpretation
Raster and multispectral QA	20%	Alignment, scale, masks, safe indices, DSM and output round trips are demonstrably correct
Scientific interpretation	15%	Product- and region-specific decisions connect evidence to ecological consequence and timing
Documentation and provenance	10%	Immutable inputs, checksums, versions, manifest, limitations and actions are complete

Automatic revision required

Revision is required if:

• GSD is presented as positional accuracy;
• orthomosaic is described as raw imagery;
• reprojection error is treated as absolute validation;
• GCPs and check points are interchangeable;
• reflectance scale is assumed;
• bands enter calculations without alignment validation;
• DSM is called bare ground or direct vegetation height;
• an index uses shifted bands;
• temporal mismatch is ignored;
• visual appearance is the only QA evidence.

Professional Mistakes — UAV and Photogrammetry

Use this table as a final failure-mode review. Add a row to your QA matrix for every mistake relevant to the handover.

Mistake	Why it happens	How to detect it	How to prevent it	Scientific consequence
GSD = accuracy	Both use distance units	Accuracy claim cites pixel size only	Report GSD, effective resolution and independent residuals separately	Boundaries and plot overlays appear more certain than supported
Orthomosaic = raw photograph	Mosaic looks photographic	No surface, seamline or source-image history	Preserve the reconstruction and contribution chain	Derived pixels are interpreted as direct instantaneous observation
High overlap = guaranteed quality	One plan percentage is easy to compare	Blur, texture, motion and achieved coverage are absent	Audit sharpness, matches, geometry and actual footprints	Reconstruction failure remains despite many images
All surveyed points used as GCPs	Fitted residuals become smaller	No withheld role exists	Reserve independent, distributed check points	External accuracy is untested and optimistic
Low reprojection error proves position	Internal diagnostic is called error	No check-point residuals accompany the claim	Label internal versus external evidence	Shifted/warped products are accepted
Vertical accuracy ignored	Maps focus on planimetry	Vertical datum and RMSE Z are missing	Audit height reference, transformation and checks	DSM differences receive unsupported height meaning
Rolling shutter ignored	Frames look sharp at first glance	Shutter/readout and motion are undocumented	Record shutter and validate motion modelling	Systematic geometric distortion enters the block
Motion blur ignored	Mosaic blending hides weak frames	Source-image sharpness not reviewed	Predeclare sharpness rules and inspect frames	Matches and fine boundaries degrade
Changing illumination ignored	Colour balancing looks smooth	Exposure/irradiance varies with flight order	Track light, settings and image contribution	False spatial/spectral gradients appear
Panel guarantees reflectance	Reference target sounds definitive	Panel timing/exposure/condition absent	Follow protocol and retain residual limits	DN is presented as comparable reflectance
Automatic exposure unreviewed	Images look visually balanced	exposure/gain varies across block	Audit metadata and use documented correction	Between-image values are incomparable
Bands assumed perfectly aligned	Files share dimensions	transforms or edges disagree	Numeric contract plus local registration checks	Spectral arithmetic mixes ground footprints
NDVI calculated before registration	Formula is simple	index halos follow object edges	Make alignment a blocking gate	False vegetation patterns are created
DSM treated as DTM	Both are elevation rasters	ground-classification evidence absent	Document surface generation and validate terrain	Canopy/buildings are interpreted as ground
DSM treated as direct vegetation height	Height appears above ground	no aligned DTM or vertical validation	Require compatible surfaces and uncertainty	Plant structure is overstated or biased
Seamlines ignored	Mosaic looks continuous at normal zoom	contribution/brightness boundaries untested	Inspect fixed stretches and source map	Radiometric or geometric discontinuity biases analysis
Block-edge distortion ignored	central area dominates global metrics	residuals increase near boundary	acquire margins and map local error	Edge plots have worse position than reported
Water/moving vegetation ignored in SfM	software still returns points	noisy clouds, holes or ghosts align with target type	anticipate failure and mask/model explicitly	Invented or unstable surfaces enter products
Field/UAV date mismatch ignored	dates are close numerically	process events/tide/grazing absent	define temporal compatibility before extraction	raster and field data represent different states
Bad areas hidden	final map looks cleaner	exclusions lack IDs, reason or original support	map and document every excluded region	users assume complete coverage and selection bias is concealed

Reflection and submission

Answer in private notes:

1. Which product passed geometric QA but failed another category?
2. Which internal diagnostic was most tempting to overinterpret?
3. Which finding affected only one region rather than the whole survey?
4. Why was stopping Red-edge NDVI the correct scientific decision?
5. What new evidence would change your final classification?

Submission

Upload the complete uav_practicum/ folder or project archive, the verified QA map, your notebook, all tables and the final Markdown report. Include one 350–500 word executive summary that a research lead can act on without opening the code.

Portfolio artifact

Artifact 2.D — Professional UAV Survey Assessment

Add the complete assessment to the Professional UAV Product Audit and Processing Report. It is the Chapter 4 portfolio handover and the acquisition-to-analysis bridge for the later Satellite Earth Observation chapter.