Sensors Across Large Sites: Where to Measure and How to Avoid Useless Data

Start with the decision, not the device

On a large territory a point sensor describes a tiny volume or a narrow view, not the site. Readings taken two metres apart can differ more than the seasonal change you are chasing, because soil texture, shade, slope and airflow all shift quickly. The practical failure is usually not physical accuracy but relevance: fleets grow, dashboards fill, and nobody can say which value changed which decision.

Treat "useless data" as an operational term, not a technical one. A series is useless if, over a full season or a defined review cycle, it never crossed a threshold, never changed an alert, and never started or stopped a task. Under that definition useless data is not harmless — it consumes budget for hardware, batteries, connectivity and analysis time.

Budget around the metric that matters: the cost of one actionable reading and the share of series that actually changed a decision in the last cycle, rather than the number of installed nodes.

• Write down which decision each node serves and in which zone
• Set target thresholds and a review cadence that removes silent nodes
• Measure cost per actionable event, not cost per installed unit

Delineate the unit of management before you measure

The cheapest way to avoid useless points is to refuse to measure variability you cannot act on. If a field is watered and managed as one unit, monitor it as one unit at a single representative spot instead of scattering probes to document horizontal differences that no actuator can correct.

Where zones genuinely receive different treatment — separate drip laterals, variable-rate application, independent sprinkler circuits — each becomes a decision unit that deserves a representative point plus depth coverage. Put at least one sensor inside the active root zone and one below it, so you can separate "the field is dry" from "water drained past the roots."

Let a GIS pick candidates before you leave the office

Remote layers are cheap and cover the whole territory in a way a handheld probe never will. Elevation and slope from a digital elevation model, soil maps and several seasons of satellite vegetation indices show where conditions actually vary. Overlay them in a GIS and the landscape divides into internally similar strata — and that is what you sample, not an arbitrary grid.

Published field workflows now formalise the sequence: project scoping, choosing the environmental drivers that matter, then an algorithm that distributes a chosen number of points across the mapped environmental space, checks statistical power and visualises the chosen sites against the layers. Open-source tools take a budget, a target area and a list of drivers and return candidate coordinates plus diagnostics.

Pair the map with people. Operators, agronomists and surveyors know where access is, where water gathers, which corners dry first and where sensors get damaged or stolen. Local knowledge corrects satellite noise and belongs in the process as an explicit step, not as an afterthought.

Size the fleet from variance and accuracy, not area

There is no universal "one sensor per hectare". On homogeneous terrain a handful of carefully placed points can represent the space, while in a heterogeneous valley where soil and slope change over short distances the same count drowns. Stratified sampling — describing variation within each stratum and budgeting points across strata — is more efficient than a uniform grid, which over-samples monotony and misses edges and boundaries.

Air-quality network design makes the same trade-off explicit: dense, populated districts need more points to capture exposure and hotspots, while rural and open zones are served by fewer, strategically placed stations. The constraint is always cost against the spatial resolution you want. Reference-grade instruments are scarce and expensive, so hybrid designs put a few calibrated devices among many low-cost sensors and lean on calibration, validation and integrity checks.

Run a pilot before committing the full budget. A short deployment or a quick variance study tells you whether your candidate spacing reproduces the range of values you need to detect. Adding nodes after the pilot is far cheaper than re-engineering a fleet that never captured the signal.

Engineer the reading, the link and the mount at each point

Once coordinates are set, the local environment still decides whether you get data or noise. Vertical placement usually matters more than horizontal: for soil moisture, decide whether the goal is the root zone, total stored water or deep drainage, and set probe depths to match — a single shallow probe cannot tell you whether water reached the roots or left the profile.

The radio link is part of the measurement. Before installation, map coverage across the site and choose the network — LoRaWAN (private or operated), NB-IoT or LTE-M — from node count, terrain, data criticality and whether you want your own gateways. Mount nodes at height and vertically, away from metal and reinforced structures, and re-test with a field tester, because signal strength can change from one metre to the next.

Along perimeters think in overlapping layers rather than evenly spaced points. Fixed wide-angle cameras spaced so their fields of view overlap, with pan-tilt-zoom units at gates and access points, give detection continuity that uniform spacing cannot; validate the spacing with an on-site survey rather than trusting a datasheet.

• Common mistakes: poor radio coverage, sensors next to metal structures, wrong sampling frequency, underestimated battery life, no on-site validation
• Low camera mounts and ultra-wide lenses distort edge pixels and create blind spots
• For metal enclosures, run a closed-door network test before installation

Validate, calibrate — then retire data that stops earning its keep

Accuracy is a maintenance discipline. Calibrate against a reference device, validate readings against known conditions and put automated integrity checks in place to flag drift and anomalies. A sensor that drifts quietly for months is worse than no sensor, because decisions built on its bias are silently wrong.

Make review cadence the anti-uselessness routine. Each quarter or season, ask of every node: did this series cross a threshold, change an alert or trigger a task? If not, repair, reposition or retire it. The goal is a fleet where every data stream has an owner and a decision attached, and where nodes that no longer drive anything are removed before they cost another season of batteries and bandwidth.

Finally, check the jurisdiction: on industrial sites and in regulated monitoring the density, method and periodicity of an observation network are often fixed in a work program and in national codes, so reconcile your design with current legal requirements before purchase.