Every terrain-referenced navigation system shares the same weakness, and it is worth being plain about it: it can only recognise ground that has something to recognise. Over a river bend, a road junction or a mine bench, the match is sharp. Over calm open water, a dry salt lake or a paddock of uniform crop stretching to the horizon, there may be very little to match at all. What separates a trustworthy system from an untrustworthy one is not whether it has this limit, because they all do, but how it behaves when it reaches it. This article looks at the difficult ground, how a system can tell that it is uncertain, and what it should do next.
Featureless and repetitive are different problems
It helps to separate two kinds of difficult ground.
Featureless ground offers little information. Calm water, the flat white crust of a salt pan, a freshly worked paddock or a closed crop canopy all look much the same from one hundred metres to the next. The matcher has too little to lock onto, and the natural result is a broad, weak match or no match at all. That is a problem, but an obvious one.
Repetitive ground is more dangerous. Plantation forestry in regular rows, parallel dune fields, vineyards and orchards, and grids of identical sheds or pens contain plenty of features, but the same pattern appears in many places. The matcher can find a strong match in the wrong place, because the wrong place looks just like the right one. That is a confident error, and it is the one that matters most.
The special cases
Water is the classic hard case. Calm water is nearly featureless and reflects the sky, sun glint can saturate parts of the image, and wind-driven waves create texture that moves with the weather rather than staying fixed to the map, so matching against it would be worse than useless. Shorelines, by contrast, are excellent features, provided the reference allows for tide and water level. On coastal and river work the practical answer is often to plan the route so that a shoreline stays in view.
Salt pans and dry lakes combine a uniform surface with high brightness and, on hot days, strong heat shimmer and mirage. Many inland lakes in Australia are dry for years at a time and then fill after heavy rain, so the reference imagery may show a dry pan while the aircraft sees water, or the reverse.
Uniform crops change through the season. A crop with visible row structure early on can close into a featureless canopy later, and a paddock that was stubble in the reference imagery may be green on the day.
How a system knows it is uncertain
A position fix without an honest uncertainty is dangerous, because the autopilot’s navigation filter weights every measurement by how uncertain it claims to be. Terrain matching offers several natural signals of confidence:
- Sharpness of the match. A good match stands out clearly from its neighbours. A broad, flat match means many nearby positions fit almost as well.
- Ambiguity. If the second-best candidate position scores nearly as well as the best, and it is somewhere else entirely, the fix is ambiguous however good the best score looks.
- Supporting evidence. A fix supported by features spread across the whole image is stronger than one resting on a handful of features clustered in one corner.
- Consistency with motion. A new fix should agree with where the inertial and airspeed data say the aircraft ought to be. A fix that implies an impossible jump deserves suspicion.
A well-designed system combines signals like these into an uncertainty that grows smoothly as the ground becomes less informative, and it is tested to confirm that the uncertainty is calibrated: when it states an error bound, errors should fall inside it about as often as claimed. Over repetitive terrain, keeping more than one candidate position alive until the ambiguity resolves is safer than committing early to the best-looking one.
Falling back gracefully
When the matcher has nothing useful to say, the honest move is to say so: report a large uncertainty or withhold the fix, rather than emit a weak position with a small error bar. The autopilot’s navigation filter then carries the aircraft on its other sources.
For a fixed-wing aircraft, those other sources are reasonably good over short periods. Inertial measurements, airspeed, heading and barometric altitude, combined with the wind estimate the aircraft built up while it had good position, support dead reckoning that holds up for a while and then degrades as the wind shifts and small errors accumulate. How long it holds depends on the sensors and the conditions, which is why it should be measured on the operator’s own platform rather than assumed.
The mission logic should then decide what happens, based on the reported uncertainty rather than a single healthy-or-failed flag:
- If the uncertainty is still well inside what the mission can tolerate, continue and expect a fresh fix at the next distinctive feature.
- If it is growing towards the limit, consider climbing to widen the camera footprint and bring more features into view, or turning towards known features such as a road or a shoreline.
- If it exceeds the limit, execute the planned contingency, which should itself be designed to take the aircraft back over feature-rich ground.
- In every case, tell the remote pilot clearly which source is carrying the navigation and how uncertain it is.
Planning around quiet ground
Much of this can be managed before take-off:
- Review the route over the reference imagery and identify low-feature and repetitive legs.
- Keep those legs as short as practical, and plan transects that periodically cross roads, fence lines, creek lines or shorelines.
- Over sparse country, choose an altitude that keeps several distinct features in each frame rather than the lowest permitted height.
- Use recent imagery of the actual area, captured in a similar season, because over plain country subtle features are all there is.
- Treat each long, quiet leg as a known risk in the flight plan with its own contingency, not as a surprise.
Where TerrainSLAM fits
TerrainSLAM provides absolute position by matching a downward camera’s view against pre-loaded terrain data, onboard, and compensates for heat shimmer, haze and changing light, which are common companions of salt pans and inland plains. It adapts its processing across altitudes, from low passes to higher survey flights. Like any terrain-based system, it depends on the ground offering something to match. When you evaluate it, or any alternative, ask to see how it behaves over the quiet ground on your own routes as well as the distinctive ground, and we are happy to build that into a tailored demonstration.