Knowledge from the field and the lab

A single-beam echo sounder measures depth along the line the vessel sails, producing profiles that are interpolated into a surface. A multibeam system transmits a fan of beams — typically covering a swath three to four times the water depth — and can achieve full seabed coverage with no interpolation between lines.

The decision usually turns on three questions. First, coverage: if acceptance criteria require full seabed search — common for post-dredge clearance, wreck and obstruction checks, or charting to IHO Order 1a and better — multibeam is the only realistic option. Second, depth and geometry: swath width scales with depth, so in very shallow, uniform areas such as canals, reservoirs and near-shore strips the swath advantage collapses, and a single-beam grid at 5–10 m line spacing can be equally informative. Third, budget and programme: single-beam mobilises from small craft with simpler calibration and faster processing; multibeam needs patch-test calibration, sound-velocity management and heavier processing time.

On real port projects we often combine the two — multibeam for berth pockets and dredge areas where depth clearance must be proven, single-beam for periodic siltation monitoring where a repeatable grid is sufficient. If the deliverable is a volume rather than a chart, it is worth asking which method the tolerance actually requires. The answer is often cheaper than assumed.

RTK (real-time kinematic) and PPK (post-processed kinematic) both refine the drone's on-board GNSS positions to centimetre level using carrier-phase observations — the difference is when the correction happens. RTK applies corrections in flight over a radio or network link; PPK records raw observations and resolves them against a base station after landing.

For final accuracy the two are similar: with a sound base station and good satellite geometry, both routinely deliver 2–5 cm horizontal and 3–8 cm vertical accuracy on the camera positions. The practical differences matter more. RTK depends on an uninterrupted correction link, so long corridors, quarries with masked horizons and congested radio environments can drop fix and degrade whole blocks of imagery. PPK has no link to lose — every photo is corrected in post-processing — which is why we default to it on linear and remote sites.

Neither method removes ground control entirely. We still place independent checkpoints to verify accuracy, and at least one known mark to catch datum or antenna-height blunders, particularly in the vertical. A brief that once needed twenty or more ground control points now typically needs four to six marks, cutting field time substantially. On turnaround, RTK data can be processed the same day; PPK adds roughly an hour of baseline processing — rarely the critical path.

IHO S-44 (currently edition 6) defines minimum standards for hydrographic surveys, expressed as orders. Each order sets a total vertical uncertainty (TVU), a total horizontal uncertainty (THU) and a feature-detection requirement, all at the 95% confidence level. TVU is computed from two coefficients — for Order 1a they are a = 0.5 m and b = 0.013, meaning roughly ±0.5 m in shallow water, growing gently with depth.

Order 1a applies where under-keel clearance matters but some tolerance exists: it requires a full seabed search and detection of features larger than 2 m in depths to 40 m. Order 1b carries the same uncertainty limits but drops the full-search requirement — appropriate where seabed type and clearance make small features irrelevant. Order 2 is for deeper water where a general description of the seabed is adequate. Above 1a sit Special Order (TVU coefficient a = 0.25 m, 1 m feature detection) and Exclusive Order, used for berths and critical channels.

The practical point is acceptance criteria. A dredging contract that states "survey to IHO Order 1a" is defining how confident everyone must be in the declared depth — which drives equipment selection, calibration, line plans and processing. Specifying a tighter order than the decision needs adds real cost; a looser one can void the survey's value as contractual evidence. We help clients settle the order before pricing the work.

Survey receivers now track four global constellations — GPS, GLONASS, Galileo and BeiDou — plus regional systems. More satellites do not directly mean more accuracy; they mean better geometry and more redundancy, which is what keeps solutions reliable near structures, under partial canopy and at low elevation angles.

Accuracy comes from how errors are removed. A standalone position is accurate to a few metres because of satellite clock and orbit errors and atmospheric delay. Corrections from a base station or network remove most of these, because the errors are nearly identical at both ends of a short baseline. Centimetre accuracy specifically comes from carrier-phase measurement: rather than timing the code, the receiver measures the phase of the carrier wave (wavelength about 19 cm) and resolves the whole number of cycles — the ambiguity. Once fixed, horizontal positions are typically good to 8–15 mm plus 1 ppm of baseline length.

The vertical is always the weak axis — roughly 1.5 to 2 times the horizontal uncertainty — because every satellite is above the horizon. And a fixed solution can still be wrong: multipath or an incorrect antenna height will not show in the receiver's quality figures. Our procedures require independent checks on known marks at the start and end of every session. It is the cheapest insurance in surveying.

Most survey businesses lose more time between the point cloud and the drawing than they do in the field. The steps are predictable — classify ground, build a TIN, generate contours, cut sections, compute volumes, lay out sheets — and predictable steps can be automated.

A robust pipeline looks like this. First, classification separates ground from vegetation, structures and noise; on clean photogrammetric or LiDAR data, automated ground filtering is now highly reliable, with manual review focused on breaklines and edges. Second, surface building: a TIN constrained by breaklines — kerbs, crests, toes — rather than points alone, which is what keeps contours honest around engineered features. Third, derivatives: contours at specified intervals, cross-sections at chainage along an alignment, and cut/fill volumes against a design or previous surface, all generated from the same TIN so they cannot disagree with one another.

The audit trail matters as much as speed. Every volume figure should be reproducible, with the source surface, comparison surface, boundary and method recorded alongside the result. Automation can reduce repetitive work, but performance depends on the input data, quality controls and chosen workflow.

Machine-learning segmentation can assist with extracting selected features from suitable orthomosaics. Results vary with image resolution, lighting, geography, training data and feature definition, so no general accuracy or completeness claim should be assumed.

What still needs a person: classifying what the model found (a shed against a substation), features hidden by shadow, canopy or parked vehicles, small or rare classes with little training data, and any judgement about what is relevant to the deliverable. Vectors straight out of a model are also not CAD-ready — they need topology cleaning, layer mapping and datum-correct georeferencing before they behave in a drawing or a GIS.

AI-assisted extraction does not eliminate digitising; it changes part of the task into review and correction. Whether that saves time depends on the area, feature repetition, required accuracy and review burden. A technical assessment is needed before choosing it over conventional digitising.