Calibration, validation, and reference measurements in proximal sensing

It’s relatively easy to set up a spectrometer or a camera in the field and start collecting data. The fun part begins when you look at the images and measurements and start asking what they actually tell you.

Perhaps you want to calculate an index such as the Green Chromatic Coordinate (GCC) or the Crop Water Stress Index (CWSI). But how reliable are those pixel values from your camera? Do you need to calibrate your sensor? How can you check whether your results are trustworthy? And what can you use as a reference?

These are important questions, and they bring us to three terms that often appear together in proximal sensing: calibration, validation, and reference measurements.

They are closely related, but they don’t mean the same thing. Understanding the difference can make your measurements more meaningful and your research much more convincing.

Why the terms matter

Calibration, validation, and reference measurements are sometimes used interchangeably in everyday scientific conversations. It’s easy to see why: all three help us understand whether our measurements can be trusted. But each plays a different role.
A useful way to think about them is this:

• Calibration establishes or adjusts a relationship between a sensor’s response and a known quantity.
• Validation tests how well a measurement method or model performs using independent data.
• Reference measurements provide the observations against which a sensor, model, or derived product can be compared.

Let’s make this more concrete. Imagine you have a thermal camera mounted above a crop canopy. You want to use its images to estimate canopy temperature and assess water stress.

First, you check the camera’s temperature readings against a suitable reference target. That’s part of instrument calibration. Next, you compare your canopy-temperature estimates with independent temperature measurements collected in the field. That’s validation. Those independent field measurements are your reference measurements. The same field campaign might contribute to several parts of this process, but it’s important to be clear about which role each measurement plays. Once you get the hang of these distinctions, it’s much easier to design a sensible measurement strategy.

Calibration means relating a response to something we know

Calibration is about establishing the relationship between a measurement system’s response and a known or controlled quantity. At the instrument level, this might mean checking whether a sensor reports the right temperature, whether a spectrometer measures wavelengths correctly, or whether a probe responds appropriately to a known range of values.

Take a thermal camera again. Some cameras include an internal calibration mechanism, such as a reference shutter or an internal target with known thermal properties. The camera uses this mechanism to check or correct its response. Depending on the design, you might hear a little click when the mechanism activates.

A field spectrometer offers another familiar example. Before measuring vegetation, you may use a white reference panel to account for the illumination and instrument response. These panels are designed to have known, well-characterized reflectance properties. The resulting reference spectrum helps you calculate reflectance from the measured signal. A flat-looking spectrum can be a reassuring sign, but don’t assume that a perfectly straight line is always the goal. Real reference panels have wavelength-dependent reflectance, and their behavior depends on the panel’s properties and the instrument’s processing.

Calibration can also mean building a model

In proximal sensing, calibration isn’t limited to the instrument itself. We often use the word when building a model that relates sensor measurements to a property we’re interested in.

For example:

• Spectral measurements and leaf nitrogen: Collect leaf samples, measure their nitrogen content in the laboratory, and relate those measurements to spectral features. You can then build a model to estimate leaf nitrogen from new spectral observations.
• RGB images and canopy development: Calculate vegetation color metrics from camera images and relate them to independently measured canopy cover, leaf-area index, or phenological development.
• Soil electrical conductivity and soil properties: Compare conductivity measurements with soil samples to investigate relationships with texture, salinity, or moisture. The relationship will depend on the sensor, soil conditions, and the property being estimated.
• Canopy temperature and crop water stress: Use thermal observations alongside suitable plant, weather, and reference measurements to develop a method for estimating water stress.

The key point is that the model learns a relationship from observed data. The relationship may be useful, but it doesn’t automatically apply everywhere.

A model developed for one crop, sensor, field, season, or set of environmental conditions may perform differently under another. Calibration gives us a starting point, but we still need to investigate how well the relationship holds up.

Validation: Does this work beyond the data we used?

So you’ve calibrated your sensor or developed a promising model. Great! But how do you know whether it will work on observations it hasn’t seen before? That’s where validation comes in. Validation evaluates the performance of a method using data that are independent of the data used to develop or fit it. This distinction matters because a model can look excellent when tested on the very observations it was trained on, yet perform much less impressively in the real world.

Imagine collecting spectral measurements from 100 leaves and using their laboratory-measured nitrogen content to develop a prediction model. You could randomly divide the observations into a training set and a test set. Use the training set to develop the model, then evaluate its predictions against the reference measurements in the test set. That is a useful starting point, provided the split reflects the structure of the data. If several leaves come from the same plant, for instance, randomly distribute those leaves between training and test sets may allow the model to exploit similarities between them.

A stronger test of transferability might hold out entire plants, plots, fields, or even a growing season. This helps reveal whether the model works beyond the particular samples it learned from. The right validation design depends on what you want to do with the results.

A model intended to compare plants within one field may not need to demonstrate the same level of transferability as a model intended for use across multiple farms and growing seasons.

A camera-based vegetation index might reliably capture the timing of spring green-up in a particular ecosystem. That doesn’t automatically mean it can estimate leaf-area index accurately. Likewise, a thermal model might distinguish relatively stressed and unstressed areas within a field but struggle to produce accurate absolute estimates of crop water stress across different weather conditions. Neither result is necessarily a failure. The important question is whether the method performs well enough for its intended application.

Good validation tells us not only how well a method works, but also where, when, and under what conditions we can reasonably expect it to work.

Reference measurements: Useful evidence but far from “magical truth”:

You’ve probably come across the term “ground truth”. It’s widely used in remote sensing, and for good reason: we often need field observations to help interpret what a drone, aircraft, or satellite sees. But there’s a small catch. Field observations are measurements too, and measurements are never entirely free of uncertainty.

A soil sample collected from one point may not represent the full area observed by a sensor. A leaf-clip measurement may describe a small section of a canopy rather than the whole plant. A field measurement taken in the morning may not be directly comparable to a camera observation made hours later, especially when the target changes rapidly. Even careful fieldwork requires choices about where to sample, how many samples to collect, which instrument to use, and how to handle variation between observations.

That’s why I often prefer the term “reference measurement”. It reminds us that field observations provide valuable evidence for comparison without implying that they are automatically perfect or error-free.

A good reference measurement should be as independent, reliable, and representative as practical. Ideally, it should also be traceable to recognized standards where relevant, well documented, and matched to the quantity, spatial scale, and timing of the sensor observation. For example, if you want to validate canopy temperature from a thermal camera, a carefully positioned and suitable temperature reference can be more informative than a field observation taken at a different time or from a completely different part of the canopy.

The goal here isn’t to find an imaginary perfect measurement. It’s to build the strongest comparison you reasonably can and understand its limitations.

Proximal Sensing: A valuable link to remote sensing

One of the things I find particularly exciting about proximal sensing is how it connects detailed field observations with measurements from drones, aircraft, and satellites. A field camera can monitor canopy color and seasonal development day after day. A spectrometer can capture detailed information across wavelengths that an ordinary RGB camera cannot distinguish. Soil sensors can track changes in moisture over time, while thermal cameras can reveal patterns in canopy temperature across a field. These observations can add context to remote-sensing data that would otherwise be difficult to interpret.

For example, a PhenoCam may collect images several times a day, while a satellite passes over the same area only periodically. The camera can reveal what happens between satellite observations, helping us understand whether a change in a satellite-derived vegetation index reflects gradual seasonal development, a sudden weather event, or something else.

Drone surveys can provide another useful perspective, offering detailed spatial observations between field measurements and satellite observations. However, drone flights are often conducted as occasional surveys rather than continuously, so their temporal coverage may be much more limited than that of a fixed field camera.

So, can we simply use proximal measurements as reference data for remote sensing? Sometimes, yes, but not automatically. The quantities being measured, the spatial footprints, viewing geometry, timing, spectral response, and processing methods all matter. A satellite pixel might include several fields, different canopy layers, shadows, bare soil, water, or infrastructure. A field sensor, on the other hand, may observe only a small and carefully selected part of a single target. Even if the sensors are looking at the same place, they may not be measuring the same thing at the same scale. This is why matching observations across scales is such an important part of a robust sensing strategy. Proximal measurements can provide excellent supporting evidence, but their suitability as references needs to be demonstrated rather than assumed.

Uncertainty is part of the story, not a reason to give up

Every sensor-derived estimate and every reference measurement comes with some uncertainty. It might come from the instrument itself, the calibration procedure, the sampling design, differences in timing, natural variability in the target, environmental conditions, or assumptions built into a model. The important thing is not to eliminate every possible source of uncertainty. That’s rarely realistic. It’s to understand which sources matter and how they affect the conclusions we draw.

Depending on the application, useful ways to describe uncertainty are repeated measurements, calibration residuals, confidence intervals, error statistics, uncertainty budgets, sensitivity analyses, or a clear discussion of the method’s limitations. It’s also worth remembering that a single accuracy statistic rarely tells the whole story. A model may have a strong correlation with reference observations while still producing errors that are too large for a particular decision.

For example, a model that successfully ranks plots from relatively dry to relatively wet soil may still be unsuitable for estimating absolute soil-water content. Whether that matters depends on how the results will be used. A clear description of uncertainty often tells us much more than an impressive-looking number on its own.

What makes a convincing calibration or validation study?

A good study doesn’t need to be complicated for the sake of it. It needs to make the measurement process understandable and the conclusions defensible.

When reading or designing a study, I like to look for a few essentials:

• The target quantity: What are we actually trying to measure or estimate?
• The sensor and its response: What does the instrument record, and how are the raw measurements processed?
• The reference method: How are the comparison measurements collected, and what uncertainty do they have?
• The sampling design: Where, when, and how often are measurements taken?
• The matching procedure: Do the sensor observations and reference measurements represent comparable places, times, and quantities?
• The model-development process: Which data are used for calibration, and how are model choices made?
• The independent evaluation: How is performance tested on data that were not used to develop the model?
• And the limitations: Where might the method perform less well, and what should users be cautious about?

It’s tempting to focus on a high correlation coefficient or a low prediction error, especially when presenting exciting new results. But those numbers only make sense in the context of the study design, the range of observations, the independence of the data, and the intended use of the method.

The aim isn’t to build the most impressive model. It’s to understand when a measurement system provides dependable evidence and when we should be more careful.

Reference systems are not just reference numbers

Perhaps the most useful way to think about reference measurements is as part of a wider observation system. Field samples, weather stations, camera records, sensor-maintenance logs, crop-management information, drone flights, and satellite products can all contribute to a more complete picture. Each source has its own strengths and limitations, but together they can help us understand why a sensor records a particular pattern and whether that pattern reflects something meaningful in the real world. This is where proximal sensing really shines. It can provide detailed, repeated observations close to the target, helping us connect processes happening in the field with patterns observed at broader spatial scales.

And that’s the bigger picture: calibration helps us establish meaningful relationships, validation helps us test whether those relationships work, and reference measurements give us the evidence we need to make the comparison.