Why Drone Photogrammetry Struggles Over Water — And What Surveyors Should Do Instead

Why Drone Photogrammetry Struggles Over Water — And What Surveyors Should Do Instead

Reflections, moving surfaces, and the riverbed a camera can't see

The drone flew all the way up a river corridor without incident. The orthomosaic is sharp. The banks, the access road, and all the rock outcrops reconstruct well.

Then you open the point cloud and view the water.

It's noisy. Patches are missing. Some elevations contradict their neighbours, and other lengths look perhaps a bit too smooth. That might not be a big deal if the project only requires an attractive map. If it is used to feed a hydraulic model, as part of some flood assessment or a group of river cross-sections, this becomes very relevant.

The drone took thousands of photos. That does not mean that it recorded reliable elevation information for everything, though.

Water has been flagged as a limit in our previous posts on engineering-ready deliverables and mountain flight planning. It needs an explanation of its own, because in a river project the water is usually what the client cares most about.

What Water Does to the Process of Matching

Photogrammetry takes the same feature in photographs of a scene captured from different positions and tracks how that feature moves between images to compute its place in 3D. A rock, road marking or building corner stands still and looks the same from one exposure to the next.

Water does neither. In the time between two consecutive photos, the surface of a river can change due to:

  • reflections of the sky, clouds, trees, and surrounding slopes
  • ripples and morphing waves
  • water flowing and physically moving between exposures
  • Sun glint and changing illumination, which introduce noise and prevent correct feature matching. This may lead to erroneous elevation models or even no elevation at all
  • a uniform, repetitive texture that will give almost no indication for the program to latch onto

Esri's own Drone2Map troubleshooting advice calls water notoriously hard to match between drone images, and says the result can be mottled, discoloured, or jagged sections of the orthomosaic.

Research backs it up, even under good conditions. A 2026 study of a clear, fast-flowing river in Austria compared three image-based methods against a LiDAR reference. Structure from motion only created sparse, inconsistent points, and multi-view stereo reconstructed the riverbed, not the water surface. The authors concluded that current image-based approaches are not enough for robust water surface reconstruction in that kind of setting.

That's the very important distinction: a surface can be fully visible in an image yet not be suitable for robust reconstruction.

A River May Look Complete in the Orthomosaic but Be Incorrect in 3D

An orthomosaic is a geometrically corrected image. In it, a river may appear visually continuous, subject to water conditions and processing choices. A digital surface model (DSM) is different: it holds elevations calculated from reconstructed geometry. When the water gives the software too few stable features, the DSM can end up with gaps, irregular points, local distortions, or interpolated areas that look plausible but are not supported by real measurements.

With many processing tools, users can mask a waterbody and interpolate across it, which makes the surface look nice. But interpolation fills a hole. It doesn't measure the water. This is the same gap we discussed in terms of DSM artefacts, between an output that looks clean and one that is trustworthy.

A pretty orthomosaic does not ensure that the elevation model beneath it is trustworthy.

The Water Surface Is Not the Riverbed

Even when a water surface does reconstruct, you are still getting the water level rather than the bed beneath it. An RGB camera records light coming directly from the scene, and in opaque, turbid, or reflective water it just can't see the bottom.

Here is one exception you should be aware of. In clear, shallow water, research demonstrates that through-water photogrammetry can work, but only if reflection and refraction corrections are applied. Without them, the bed looks shallower than it really is, because light bends at the surface; that work uses a refractive index of 1.34 for water. For near-vertical imagery, a 2 metre deep channel would appear approximately 1.5 metres deep as a rough guide. Standard processing over a working river doesn't deliver that. It's a special workflow with conditions.

This is important whenever a project requires:

  • river cross-sections for hydraulic modelling
  • channel geometry for flood studies
  • sediment or scour assessment
  • bridge foundation and clearance analysis
  • reservoir capacity or underwater terrain
  • reliable elevations for drainage and water-resource design

For these kinds of jobs, a drone surface model of the visible surroundings is only part of the survey. The other bits have to be measured, not gleaned from the way the pictures appear.

What Goes Wrong Downstream

Imagine a corridor survey with banks, access roads and adjacent slopes and a length of what appeared to be shallow water. The land reconstructs well. The water is punctured, uneven, or just reflects matching errors. If those areas are treated as ground data without a check, the error travels.

A hydraulic model requires measured channel cross-sections and bed elevations. But a continuous, interpolated surface is not a replacement. A water depth has two halves: a water-level reference and a reliable bed elevation, which is seldom found from an aerial image alone.

This does not mean drone mapping is a bad fit for river projects. The error is treating different types of spatial information as if they were interchangeable.

What Surveyors Should Do Instead

Where It Can Shine, Use the Drone

Drone imagery is particularly well suited to revealing the ground around the river: bank edges, embankments, access roads and nearby structures, exposed rock and sediment bars, and erosion features. Plan accordingly with overlap, surveyed control, and checkpointing to validate the output. But consider the quality of data for water to be a problem area, and don't treat it as if the quality is equivalent to that of exposed land.

Add Conventional Measurements Where They're Needed

Accessible points along the banks can be measured with RTK GNSS and total stations. Those observations help establish control, check the mapped features independently, and connect this data to the project's coordinate and elevation datums. The method depends on access, river conditions and safety, and also on the accuracy the project requires.

Measure the Bed Directly

When the project requires riverbed elevations, that means bathymetric surveying. Depending on depth, flow and access, the options include:

  • wading measurements in suitable, safe conditions
  • a single-beam echo sounder, which works out depth from the travel time of a sonar pulse, along survey lines
  • multibeam sonar where detailed underwater coverage is required, and conditions allow it
  • other hydrographic methods suited to the site and the deliverables

However, any technique then requires that depths be referenced to a suitable vertical datum and converted into bed elevations using some measure of the water level. You want the equipment to be guided in accordance with the survey specification and site conditions, not merely by the availability of a drone.

On the Naran-Batakundi hydropower survey, for example, the Kunhar River bed was measured with a dual-frequency sonar, with river cross-sections and depths recorded alongside coordinates, while the drone and topographic survey covered the surrounding terrain.

Merge Everything into One Dataset

The best river surveys combine the methods rather than require one to do everything:

  1. Map all the observable surfaces and riverbank features with your drone.
  2. Set up GNSS or total-station observations for control and independent checks.
  3. Collect bathymetric measurements wherever underwater geometry is needed.
  4. Wrap everything in a single coordinate reference system and vertical datum.
  5. Check the overlaps and joins. Merging terrestrial and underwater data might create discontinuities where the two datasets meet.
  6. Keep a log of areas you measured directly, reconstructed photogrammetrically, interpolated, or left unmeasured.

That last step is what makes the dataset defensible.

Can a Different Sensor Make It Work?

Sometimes, within limits. Bathymetric LiDAR uses green laser pulses that can penetrate water, with usable depth ranging from under a metre in turbid water to more than 10 metres in clear water. This is unlike typical drone-based topographic LiDAR, which uses near-infrared light that reflects off the water surface rather than penetrating to the riverbed.

The result can be strong when the conditions fit. A Scottish evaluation of topo-bathymetric LiDAR found that over 95% of the bed was captured on the River Dee, with a mean discrepancy against conventional topographic LiDAR of only 0.04 m (helped by a gravel bed and good water clarity). The study, however, highlights the difficulty of timing these surveys with ideal flow conditions. And lasers, unlike echo sounders, do not penetrate fine sediments, so the points they record represent water depth without accounting for the thickness of silt or mud.

Most types of multispectral or general optical imagery can help characterise water or identify environmental features, but that does not necessarily mean it is a useful measure of the underlying terrain you seek to characterise.

So the question should not be which sensor has the highest resolution. It's which method can provide the information the project needs under the conditions at the site.

Deliverables From an Engineering-Ready River Survey

The team and the client must agree on what is expected in the final set before starting fieldwork. An engineering-ready river package can vary by project but may include:

  • a georeferenced orthomosaic of the visible corridor
  • a DSM or terrain model for the exposed land
  • bank lines and selected landforms
  • for some rivers, cross-sections and bathymetric points
  • control and independent checkpoint reports
  • a well-defined coordinate reference system and vertical datum
  • a data-gap map or limitations statement showing where measurements are not reliable
  • documentation on the integration of the aerial and conventional datasets

Not every piece is needed in every river survey. There are different requirements for a visual inspection than for a detailed hydraulic model, and thus the deliverables will vary based on the intended use. The idea is to be clear about where the limits of the data apply, as opposed to simply producing complete-looking maps and allowing that to imply full measurement coverage.

The Last Question: What Did You Actually Measure?

Drone photogrammetry is very powerful for river corridors, dam sites, bridge approaches, exposed channels, and the terrain around them. But water changes the measurement problem. The water surface is difficult to reconstruct due to reflections, movement, and an absence of persistently stable features to match, and the submerged riverbed generally needs a separate measurement approach.

That doesn't mean we need to give up on drone surveys. It means using them where they perform well, being clear about where their outputs become uncertain, and combining them with the appropriate survey methods when the engineering question asks for more.

A complete-looking river map isn't necessarily a complete river survey. The difference lies in knowing which surfaces were reconstructed, which elevations were measured, and which parts of the project still need evidence.

If your project involves a river, reservoir, or drainage corridor and you're not sure which parts need a drone, which need a survey crew and which need sonar, that is the planning conversation our topographic survey and drone imaging and processing teams have before any data is collected.

Back to blog

Leave a comment