LiDAR DEM Fails the Vertical Accuracy Specification
A 120 km² LiDAR project for a floodplain agency specifies a 1 m DEM with RMSEz ≤ 10 cm in open terrain and ≤ 20 cm in vegetated terrain at the 95% confidence level. The contractor delivers a classified point cloud and DEM. You select 60 independent RTK GNSS checkpoints: 35 in open pasture, 15 in deciduous forest, and 10 on paved roads. Results: open-terrain RMSEz = 18 cm with a mean bias of +12 cm, meaning the DEM is higher than the checkpoints; vegetated RMSEz = 26 cm with a mean bias of +9 cm. The contractor's report claims “NVA 18 cm at 95%,” which passes, and states the checkpoints were measured with a 2 cm RTK system. The DEM is referenced to NAVD88 via Geoid18; checkpoints were processed in NAD83(2011) ellipsoid heights and converted with the same geoid. A local benchmark check shows the DEM is 10 cm high relative to two stable NAVD88 marks.
1. Evaluate whether the DEM meets the specification. Address the contractor's NVA claim and the bias.
Read the worked answer
The specification is RMSEz ≤ 10 cm in open terrain. The measured open-terrain RMSEz is 18 cm, so the product fails on the stated criterion. The contractor's “NVA 18 cm at 95%” is a category error: NVA is a 95% statistic, not an RMSE. If RMSEz is 18 cm, the 95% value would be roughly 1.96 × 18 ≈ 35 cm, not 18 cm. The +12 cm mean bias is systematic. Check the geoid model, the benchmark, and the checkpoint processing. A DEM that is 10 cm high against two stable NAVD88 marks confirms a real bias. Removing a 12 cm bias leaves a random component of √(18² − 12²) ≈ 13.4 cm, which still exceeds 10 cm. The vegetated RMSEz of 26 cm also exceeds the 20 cm criterion. The product cannot be accepted on the stated specification without rework or a formally justified specification change.
2. What QA sampling and acceptance procedure should have been specified in the contract?
Read the worked answer
The contract should define the vertical accuracy statistic, the land-cover strata, the minimum number of independent checkpoints per stratum, and the exact acceptance thresholds. It should require checkpoints that were not used in calibration or classification, state the vertical datum and geoid model, and require a QA report with RMSEz, mean bias, and 95th-percentile error. For floodplain mapping, it should also require a check against stable benchmarks and a statement of any systematic bias. The contractor should not be allowed to substitute a 95% NVA figure for an RMSEz requirement. A pre-project checkpoint plan and a joint acceptance review would have prevented the dispute.
3. The contractor argues that the open-pasture checkpoints are invalid because of tall grass. How do you respond?
Read the worked answer
Compare like with like. Open pasture is open terrain under the specification, and the checkpoints were collected with RTK, which can be affected by vegetation but should be evaluated against the same conditions used in the survey. The mean bias of +12 cm and the independent benchmark check showing the DEM 10 cm high point to the DEM, not only to the checkpoints. Require a joint recheck using a new set of independent checkpoints, including bare-ground and road locations, and a stable benchmark. If the contractor wants a land-cover reclassification, require evidence that the checkpoints were actually in tall vegetation and that the DEM meets the reclassified criterion. A post-hoc explanation without new data is not sufficient.