Reading a residuals report — what the numbers actually say

The residuals report is what separates a defensible accuracy claim from marketing copy. We covered the QA report structure and the accuracy budget in earlier pieces. This article is the third in the trilogy: a focused walk through what the numbers on the residuals page actually tell you, and the recognition patterns for the common problems they're often hiding.

What residuals actually are

For each independent control checkpoint, the residuals page records the vertical residual:

Δz = cloud_z − surveyed_z

The sign convention: positive means the cloud reports a height above the surveyed truth. Negative means the cloud is below truth. Both are non-zero by some amount because no measurement is perfectly accurate; the question is what the pattern of non-zeros looks like.

A horizontal residual (Δx, Δy) is computed similarly for checkpoints on identifiable hard features (pavement markings, manhole covers, sharp edges). Vertical residuals are the headline metric for most LiDAR work; horizontal residuals matter more for asset position and CAD linework.

The six lenses on a residuals page

Any reasonable residuals report gives you these six pieces of information. Each one is a different lens on the same dataset and tells you something the others don't.

1. RMSE — the headline

Root Mean Square Error is the standard aggregate accuracy metric:

RMSE = sqrt( (Δz₁² + Δz₂² + ... + Δzₙ²) / n )

Squaring before averaging means large residuals contribute much more than small ones — a single 100 mm residual contributes more to RMSE than ten 20 mm residuals.

RMSE is what gets quoted in tender responses and accuracy claims. A "±20 mm RMSE vertical" headline means the typical residual is around 20 mm, with the understanding that some checkpoints will be tighter and others looser.

What RMSE doesn't tell you on its own:

• Whether the cloud is biased (consistently above or below truth)
• Whether the residuals are evenly distributed or have a long tail
• Whether one bad area is dragging the aggregate down
• Whether bare-ground checkpoints and canopy checkpoints behave differently

For all of those, you need the other five lenses.

2. Mean — the systematic bias check

The mean residual:

mean(Δz) = (Δz₁ + Δz₂ + ... + Δzₙ) / n

A well-calibrated capture has a mean near zero — residuals are roughly symmetric around zero, so positive and negative cancel. A mean significantly different from zero indicates systematic bias — the cloud is consistently above or below truth by the mean value.

Common causes of non-zero mean:

• Control tie issue. The GNSS solution for the base station was off, so every point in the cloud is shifted by the same amount.
• Datum / projection mismatch. Capture used GDA94 control but delivered to GDA2020 without transformation, or vertical datum was wrong (AHD vs AVWS vs ellipsoid).
• Sensor calibration drift. Some sensors drift over time; recent boresight calibration is part of the pre-flight check.
• PPK base coordinate error. Most commonly the base station's own coordinate was incorrect at decimetre scale.

A mean residual of (say) 35 mm with an RMSE of 38 mm is diagnostic: the cloud is biased by 35 mm and the precision (spread around the mean) is much smaller. The whole problem is the bias.

A mean residual of 2 mm with an RMSE of 35 mm is the opposite: no bias, but residuals are spread widely around zero. Different problem, different fix.

3. The distribution and histogram

A histogram of residuals shows the shape of the residual distribution. The healthy shape is approximately normal (Gaussian) — bell-curve symmetric around the mean, tails falling off smoothly.

Three unhealthy shapes worth recognising:

• Skewed — residuals lean one way more than the other. Indicates an asymmetric problem (e.g., uphill bias different from downhill bias).
• Bimodal — two peaks. Indicates two regimes producing different accuracy (typically bare-ground vs canopy mixed in the same aggregate).
• Long-tailed — most residuals tight, but a few far out. Indicates localised problems that the RMSE conceals; the worst-case number reveals them.

If the report doesn't include a histogram, ask for it. The distribution shape is one of the easiest signals to read once you see it.

4. Max residual and the spatial location

RMSE is an aggregate; max residual is the worst single point.

A capture with RMSE 20 mm and max residual 35 mm is healthy — even the worst case isn't far from the typical.

A capture with RMSE 20 mm and max residual 180 mm has a localised problem. The aggregate looks fine because the problem is averaged out across all the other tight residuals. Find the location of the max residual and inspect the cloud in that area.

Common reasons for outlier residuals:

• A surveyed checkpoint near a discontinuity (edge of pavement, base of vegetation) where the cloud's reported height depends on which point happened to be closest
• A checkpoint in a heavily vegetated patch where ground classification mis-identified the true ground
• A checkpoint in an overlap zone between flight strips where strip alignment didn't fully reconcile
• A genuinely wrong surveyed mark (rare, but happens)

The diagnostic isn't to dismiss the outlier — it's to understand why it's there and decide whether the cloud or the survey is at fault.

5. Per-cover stratification

A reputable residuals report stratifies residuals by surface cover type:

| Cover | RMSE | Mean | Max | | ------------- | ----- | ----- | ----- | | Bare ground | 18 mm | -2 mm | 35 mm | | Sparse veg | 24 mm | 3 mm | 48 mm | | Moderate veg | 35 mm | 8 mm | 75 mm | | Dense canopy | 65 mm | 15 mm | 180 mm|

This stratification surfaces what the aggregate hides — DTM accuracy degrades with canopy density (see the accuracy article for why). The aggregate number across mixed cover may be acceptable; the engineering question is whether the accuracy in the areas your design touches is within tolerance.

If the report has only a single aggregate RMSE with no stratification, ask for it. The aggregate can be misleading.

6. Spatial distribution

The final lens is plotting residuals on the project area as small symbols sized or coloured by magnitude. The pattern reveals:

• Random spatial scatter — healthy. Residuals are noise around zero, distributed evenly.
• Clustering — all the largest residuals in one part of the project. Localised problem (poor control distribution, strip alignment issue, classification edge case in one region).
• Drift — residuals systematically positive at one end of the project and negative at the other. Indicates a tilt in the cloud relative to truth — usually a control tie or PPK trajectory issue.
• Edge effects — large residuals around the project boundary, small in the middle. Indicates the cloud is well-controlled internally but tied loosely at the edges.

The spatial map is the lens that turns a number into a story. If the report doesn't include one, the residuals are difficult to interpret beyond the aggregate.

Common patterns and what they mean

Three patterns worth recognising on sight:

Mean ≈ 0, RMSE tight (under 25 mm), distribution normal, even spatial spread. Healthy capture. The accuracy claim is defensible.

Mean significantly different from zero, RMSE close to mean, small spread around it. Systematic bias — cloud is tilted or shifted. Diagnose by checking control tie, PPK base coordinate, datum / projection transformation. Often fixable in post- processing without re-flying.

Mean ≈ 0, RMSE OK, but max residual several times the RMSE and isolated to one area. Localised problem in a specific region. Inspect the cloud in that area; may need targeted re-fly or manual classification edit.

Mean ≈ 0, RMSE good for bare ground, RMSE 2-3× higher under canopy. Normal pattern; accuracy degrades with canopy density. Acceptance depends on whether the design intersects the high-canopy areas. May warrant a re-fly with higher point density if it does.

Bimodal distribution with two peaks in the histogram. Two regimes mixed. Usually bare-vs-canopy, occasionally two sub-projects processed differently. Per-cover stratification should separate the peaks.

Spatial drift — positive residuals one end, negative the other. Cloud is tilted. Likely PPK or control issue. Re-processing the trajectory against a different base reference often resolves.

The reading checklist

For any residuals report, five minutes with the following checklist surfaces most problems:

1. Look at the RMSE. Is it within the project's stated accuracy spec?
2. Look at the mean. Is it close to zero, or does it show systematic bias?
3. Look at the histogram. Is it roughly symmetric and normal, or does it skew / split / tail?
4. Look at the max residual. Is it within 3-5× the RMSE (acceptable) or significantly larger (localised problem)?
5. Look at the stratification. Are bare-ground residuals tighter than vegetated, as expected? Is the design area's cover class within acceptable tolerance?
6. Look at the spatial map. Random scatter, clustering, drift, or edge effects?

If all six come back clean, accept the report. If any one surfaces a concern, you've got a specific question for the provider rather than a vague unease.

TL;DR

A residuals page is six lenses on the same dataset. RMSE alone is the marketing number; mean + distribution + max + per-cover stratification + spatial map are what make the accuracy claim defensible.

A report that includes all six gives you everything you need to validate the delivery in five minutes. A report that includes only RMSE and a paragraph of prose gives you a number and a hope. Worth specifying the format of the residuals page at scoping rather than discovering at delivery what got included.