Terrain variation, occlusion, and the hidden limitations of constant-altitude drone mapping.
You planned the flight at 100 metres. The drone held that altitude. The images are sharp. The overlap is correct. The processing report looks clean.
So why does the final terrain model still contain areas that are poorly reconstructed: slopes that look smeared, gaps behind ridges, contours that don't quite make sense?
Because "100 metres" over a mountain isn't the same problem as "100 metres" over a flat construction site.
We've written about whether a drone dataset is fit for engineering use. There is an earlier question, and in the mountains it is the harder one: was the dataset captured correctly in the first place? Often the answer is decided by geometry before the drone leaves the ground.
Constant Altitude Isn't Constant Ground Distance
Most flight-planning apps measure height from the takeoff point by default. That makes "100 metres" a height above where you launched, not a height above the ground beneath the drone. The two are the same only where the terrain happens to sit at launch elevation.
Over a ridge 40 metres higher than the launch point, the camera is only 60 metres from the ground. Over a valley 40 metres lower, it is 140 metres away. The drone flew a level line; the camera-to-ground distance changed by more than a factor of two.
The alternative is terrain-relative flight, usually called terrain-following, where the aircraft adjusts its height along the route to hold a steady distance above the ground, guided by an elevation model. Pix4D's own acquisition guidance for terrain with height variations recommends this: a single flight that follows the terrain, so the distance between the drone and the ground, and therefore the GSD, stays constant across the project. It advises against flying at one constant altitude regardless of terrain changes, because the GSD then differs across the site. Where the flight app can't follow terrain, its fallback is multiple flights at different altitudes, keeping the GSD in the overlapping areas between missions within a factor of two of each other.
The GSD Specification Can Become Misleading
Suppose the flight is planned for 5cm GSD at launch level. GSD scales in proportion to camera-to-ground distance, so as an illustration:
- On the ridge, 60 metres from the ground, the GSD drops to 3cm.
- At launch level, 100 metres away, it is the planned 5cm.
- In the valley, 140 metres away, it grows to 7cm.
The report may still say "5cm GSD". The effective ground sampling distance isn't one number across the survey anymore. Overlap shifts with it. With the same camera trigger spacing, the image footprint shrinks as the ground gets closer, so an 80% front overlap planned for launch level falls to roughly 67% over the ridge, while the valley floor gets more overlap than planned. Low-overlap ridges are where the software struggles to find enough shared features, a failure mode we broke down in why the processing failed on a correctly flown survey and in how many overlap photos you actually need.
This ties into our piece on why 5cm GSD does not mean 5cm accuracy, but it is a different problem. That article is about the gap between resolution and positional accuracy. This one is upstream of it: in rugged terrain, the resolution itself isn't uniform to begin with. A 2023 study on flight network design for mountainous terrain describes the consequence: large-scale differences between images, gaps, and errors caused by extreme elevation differences reduce the quality and accuracy of photogrammetric products in mountainous areas.
Put together, the chain looks like this:
Terrain variation → changing camera-to-ground distance → changing GSD and overlap → changing image scale → weaker matching and reconstruction → occlusion → incomplete terrain representation → weaker DTM and contours.
The Hidden Problem: The Camera Can't See Behind the Slope
The last link in that chain is the interesting one, because no amount of careful processing can fix it. A camera pointing down (nadir) only records surfaces it has a line of sight to. On a slope, the terrain itself blocks that line of sight. Depending on the geometry, a mountain slope can hide:
- the opposite slope
- drainage channels
- road edges
- cliff faces
- structures
- gullies
- retaining walls
- cut slopes
That changes how the problem should be diagnosed. The cause isn't necessarily "we didn't take enough photographs". It can be "the camera never had a line of sight to the surface". You can have hundreds of exposed, sharply focused, correctly overlapped images and still have holes in your understanding of the terrain, because more images from the same angle don't add a viewpoint that was never there.
Why Nadir-Only Photography Isn't Always Enough
This is why steep-terrain work usually brings in oblique imagery alongside the nadir pass. Researchers mapping landslide-prone mountain slopes in Korea concluded that vertical imaging alone is insufficient on steep terrain, and recommended terrain-following flight planning combined with oblique imaging, which increases overlap on steep slopes and improves 3D reconstruction. A study of UAV flight missions in steep terrain, published in Remote Sensing, reached a similar conclusion from a rockfall site: the authors argue that terrain-following requires a ground control station with a terrain-following mode and a good elevation model, and that the best way to capture steep terrain is a combination of vertical and oblique images. Another study, Vega et al. (2022), tested flight height and off-nadir imagery on complex terrain and found that off-nadir images in the 20° to 35° range, flown in two perpendicular directions, were useful, with above-ground-level flights at 90/70 overlap giving the best accuracy results.
In practice, that means crosshatch (grid) missions and additional flight lines: the same slope seen from more than one direction and more than one angle, so a face hidden from one pass is visible in another.
One caution: terrain-following is only as good as the elevation model it follows. A preprint on terrain-following flight planning points out that coarse, freely available terrain models can carry average elevation errors of about 10 metres, and up to 30 metres in mountain areas. A drone following a poor model can end up closer to the slope than planned or much farther away, so the quality of the DEM matters for safety as well as for GSD.
Mountain Roads: The Hardest Case
This is where all of the above becomes concrete. A road through mountainous terrain can contain, within a few hundred metres:
- steep slopes above and below the carriageway
- sharp elevation changes
- cut and fill zones
- retaining walls
- vegetation
- bridges
- drainage channels
- river crossings
You aren't mapping a flat polygon. You're mapping a corridor through a three-dimensional environment, and almost every element on that list is something an engineer designs against. A cut slope can hide the road edge behind it from a straight-down camera. A retaining wall is a vertical face that a nadir pass barely records. Water is poorly reconstructed by photogrammetry, and dense vegetation can hide the ground entirely. Those are the limits we flagged in our post on engineering-ready deliverables, and in the mountains, they show up together on the same project.
Our walkthrough of a hydropower survey in Pakistan, from 10,000 drone images to engineering-ready data, shows the kind of ground we're working with: terraced farmland, steep cliff faces, and a river that's always moving. This kind of terrain doesn't give you room for error. A flight plan that's not built for it will fall apart. Ground control is just as important here. Control points and checkpoints need to be spread out across all the heights and slopes, not just the easy flat areas. We talked about this in the pre-survey checklist.
Before You Fly a Mountain Site, Ask:
- Is the planned height measured from the takeoff point or from the ground below?
- If the flight follows the terrain, which elevation model is being used and how accurate is it?
- How will the ground sample distance and overlap change as the drone moves across the site?
- Which slopes or cliff faces will a camera pointing straight down never be able to see?
- Are oblique or crosshatch passes planned to capture those hidden areas?
- Are control points and checkpoints spread across the full range of elevations, not just the low spots?
The Actual Takeaway
A drone survey doesn't fail in the mountains because the drone is faulty. It fails because the ideas behind a flat-ground flight plan, like constant height, constant camera-to-ground distance and a camera that sees everything, don't hold up. The fix starts before you even lift off. Use terrain-relative flight, understand how GSD and overlap will change, and add extra angles so nothing is missed. If you're working on a corridor, slope or mountain site and want the capture planned around the real terrain, not just a default setting, that's the job our topographic survey and drone imaging and processing teams do before the first image is taken.
