Record Observations for Training • moover

Open the chapter list

This is Chapter 4 of 10 in the beginner path.

A behaviour model can only learn if the observations are recorded in a consistent way. In moover, the observation file is not meant to describe every single accelerometer sample. Instead, it describes periods of time during which a known behaviour was observed.

We’ll do three things in this chapter:

define what one observation row means
look at the minimum fields the file needs
check the most common problems before training

One row means one observed bout

The easiest way to think about the observation file is this: each row describes one labelled behaviour period for one animal.

For example, if animal A was grazing from 10:00:00 to 10:03:00, that is one row. If the same animal was then resting from 10:03:00 to 10:08:00, that is another row.

The minimum information you need

At training time, moover needs to know four things from each row:

which animal the observation belongs to
which behaviour label was observed
when the behaviour started
when the behaviour ended

Here is a simple example.

obs_example <- data.frame(
  id = c("Cow_01", "Cow_01", "Cow_02"),
  label = c("grazing", "resting", "grazing"),
  start_unix_ms = c(1712128800000, 1712128980000, 1712128800000),
  end_unix_ms = c(1712128980000, 1712129280000, 1712129100000)
)

knitr::kable(obs_example)

id	label	start_unix_ms	end_unix_ms
Cow_01	grazing	1.712129e+12	1.712129e+12
Cow_01	resting	1.712129e+12	1.712129e+12
Cow_02	grazing	1.712129e+12	1.712129e+12

That table is deliberately simple. In your own file, the columns can have different names if you map them in the spec, but those four ideas must still be present.

Why label spelling matters

The model does not know that Graze, grazing, and GRAZING might mean the same thing to a human reader. To the model, those are different labels unless you standardise them.

A good beginner habit is to choose one spelling for each behaviour and use it everywhere.

Time needs the same care as labels

Observation times need to line up with the accelerometer times. If they are in the wrong timezone or mixed formats, the model may look as though it is performing badly when the real problem is simply that the labels are not lining up with the movement data.

This is why moover converts times internally to UTC milliseconds. It gives the later steps one consistent time scale to work with.

A short QC checklist before training

Before you start a training run, it is worth checking these questions:

Do the animal ids match the ids used by the accelerometer data?
Are the behaviour names spelled consistently?
Do the start and end times make sense in the same timezone as the accelerometer recording?
Are there overlapping bouts that should be resolved first?

Those checks are not glamorous, but they usually matter more than the choice of model settings in an early project.

Move through the tutorial
Previous chapter: Chapter 3. Prepare Your Accelerometer Files
Next chapter: Chapter 5. Build Your First Behaviour Model