Why Durability Over Accuracy

Most sensor guides and community projects focus on accuracy — how close a reading is to a reference instrument at the moment of calibration. WeSense takes a different approach: we prioritise long-term stability and low maintenance over initial accuracy.

This is a deliberate design choice, and some people may instinctively push back on it. This page explains why.

The Problem with Accuracy-First

Calibration Degrades

A sensor calibrated to laboratory accuracy today will drift. The question is not if but how fast and how much. For a permanent, unattended sensor network:

• Most people will never recalibrate their sensors
• Any design that requires periodic calibration will fail at scale
• An "accurate" sensor that drifts unchecked gives worse data than a stable sensor that was never calibrated

The contrast is stark. The Sensirion SHT4x datasheet specifies humidity drift of less than 0.25 %RH/year and temperature drift of less than 0.03 °C/year — essentially negligible over a 5-year deployment. Compare that to budget sensors like the AHT20, which don't publish long-term drift specifications at all — a telling omission.

For CO2, the Sensirion SCD30 uses dual-channel NDIR technology that provides hardware-level drift compensation, making it inherently stable without user intervention. Single-channel sensors like the MH-Z19B and CM1106-C rely entirely on ABC (Automatic Baseline Correction) algorithms that assume regular exposure to fresh outdoor air — an assumption that fails in 24/7 indoor spaces like bedrooms or offices.

The Painful Calibration Cycle

Many community sensor projects recommend regular recalibration against reference instruments. In practice:

• Reference instruments are expensive and not widely available
• The process is time-consuming and requires technical knowledge
• Compliance drops rapidly after the first few months
• The network degrades as uncalibrated sensors contribute drifting data

What Drift Looks Like in the Real World

A 320-day field evaluation of Plantower PMS5003 sensors at the University of Utah found that one sensor exhibited significant drift partway through the study, with dust deposition on the photodetector causing declining response to light scattering. A subsequent study identified performance changes in the PMS5003 that correlated with extended deployment duration.

Meanwhile, the South Coast AQMD evaluation of approximately 400 PurpleAir sensors over three years found dramatic variability in performance driven by seasonal trends and particulate source type — exactly the kind of inconsistency that makes accuracy-at-calibration meaningless.

By contrast, Sensirion's SPS30 particulate matter sensor is rated for a 10+ year lifetime with built-in contamination resistance — it costs more upfront, but it's the kind of sensor that a permanent network needs.

Why Stability Wins

Emergent Accuracy

A network of thousands of slightly imprecise but stable sensors achieves emergent accuracy — the statistical aggregate is more accurate than any individual sensor. This only works when sensors are consistent over time.

Recent research supports this approach. A 2025 study in Nature npj Climate and Atmospheric Science demonstrated a trust-based consensus calibration framework where sensors that consistently agree with reference standards and reliable peers receive higher weighting, while those showing drift are down-weighted. The network self-corrects — but only when individual sensors are stable enough for the algorithm to distinguish drift from real environmental variation.

The EPA Air Sensor Toolbox provides extensive evaluation data showing that low-cost sensor networks can approach reference-grade accuracy when properly understood — but the key finding across their evaluations is that sensor-to-sensor consistency matters more than individual sensor accuracy.

Government Stations Provide the Baseline

Government reference-grade monitoring stations (which WeSense also ingests) provide the accuracy baseline. Community sensors provide the density. You don't need every sensor to be reference-grade when you have reference stations for cross-validation.

What We Look for in a Sensor

Property	Priority	Why
Long-term stability (low drift)	Critical	Data quality over years, not moments
Lifespan	High	Sensors should last 3-5+ years without replacement
Maintenance requirements	High	Must be zero or near-zero
Power efficiency	Medium	Enables solar/battery deployments
Initial accuracy	Lower	Correctable via cross-calibration with reference stations
Cost	Medium	Lower cost enables denser networks

See Recommended Sensors for our specific sensor choices based on this philosophy.