The Principles of Bio-Logging

Bio-logging is the practice of fixing autonomous recorders to free-ranging animals and letting the animal do the sampling. A tag is neither a camera trap nor a satellite collar: it rides on or inside the body for days or months, logging at 1–200 Hz and returning continuous records of movement, orientation, depth, temperature, heart rate or sound. The same commitments hold regardless of the hardware. First, the animal is the instrument platform, so the sampling geometry moves with it. Second, behaviour is measured continuously instead of observed occasionally: a researcher watching a foraging seabird sees minutes, while an accelerometer on the same bird can log four weeks of wingbeats. Third, every record must remain interpretable without an observer present, which pushes calibration and validation upstream into the deployment plan. The cost of that design is structural. Nothing can be repaired after release, so sensor choice, memory and gain are all fixed blind to the data they will eventually produce.

Every sensor answers a narrow question, and the deployment budget rarely allows all of them at once. Tri-axial accelerometers measure specific force, capturing gait, wingbeat frequency and posture; magnetometers add heading relative to the Earth's field; gyroscopes capture rotation but drift. Pressure sensors convert depth and altitude into dive profiles and flight height, while temperature, light and conductivity describe the environment rather than the animal. GPS gives accurate positions at a heavy energy cost; time-depth recorders give cheap, precise depth at almost none. Physiological channels — heart rate, ECG, body temperature — narrow the gap between behaviour and energetics. Sound recorders and camera loggers carry the largest data burden per hour. The principles of bio-logging point the same way: a study picks two or three channels that map onto a testable hypothesis, then keeps the rest out. A tag that measures everything at high resolution usually measures nothing for long enough to matter.

The Nyquist rule is non-negotiable: a signal sampled below twice its highest frequency is aliased, and aliased acceleration looks like plausible behaviour that never happened. A penguin's wingbeat at 5 Hz needs at least 10 Hz sampling, and field practice favours 25 Hz to keep the waveform shape. Every doubling of rate, however, doubles memory and energy. A 12 Hz tri-axial stream produces about 3.1 million values per day, so nine days of deployment is roughly 28 million numbers before any position fixes are counted. Engineers therefore spend the budget deliberately: high-rate bursts triggered by an accelerometer threshold, down-sampling during travel, and duty-cycled GPS at 15-minute intervals instead of continuous fixes. On-board processing helps — a tag that reports summary statistics or pre-classified dive phases transmits less than one that ships raw samples. The principles of bio-logging do not change with better batteries; they simply move the constraint from energy to bandwidth and to the analysis hours waiting at the end.

Most data loss happens at the interface between tag and animal, not in the electronics. A widely used guideline caps package mass at about 3% of body mass for flying birds and 5% for terrestrial mammals, but mass alone understates the problem: frontal area adds drag, and a harness that shifts 2 cm can chafe through skin in a week. Glue mounts suit short deployments on moulting penguins; suction cups last hours to days on cetaceans; collars fit felids and canids but must be tested for snagging. Capture and handling are the costliest minutes of the whole project. Restraint time, anaesthesia and recovery should be documented, and permits from animal ethics committees normally specify handling limits, maximum tag mass and recapture intervals. Tag loss rates of 10–30% are common in published deployments, so a recovery plan matters as much as the design. A tag that records beautifully for six days and never comes back has produced nothing usable. Before deployment, researchers must review ethical-standards-and-attachment-protocols-for to ensure animal welfare and data integrity are maintained.

Raw acceleration is meaningless until it is calibrated and segmented. The processing chain usually runs: correct for sensor drift and gravity, split the record into bouts or dives, extract features such as overall dynamic body acceleration (ODBA), pitch, roll and stroke frequency, then classify. Hidden Markov models suit sequences because they encode transitions between hidden states; random forests and gradient boosting handle many features without much tuning. Neither is self-validating. A classifier trained on one colony or one season often collapses elsewhere, so ground truth is essential — video from collar cameras, direct observation, or a few hours of high-rate data checked by hand. Reported accuracies above 90% look impressive until the confusion matrix shows a rare but ecologically important behaviour being missed. Precision and recall per behaviour class matter more than a single headline figure. Publishing the labelled training set, the feature code and the model weights is now the cheapest way to make a classification reproducible.

Take an illustrative deployment. A 90 kg female southern elephant seal is fitted at a Patagonian haul-out with a 65 g package: tri-axial accelerometer at 12 Hz, pressure at 1 Hz, GPS every 15 minutes, no camera. She is recaptured after nine days, an unusually clean outcome. The memory card holds about 28 million acceleration values, 780,000 pressure readings and roughly 860 positions. Step one is quality control — the pressure channel is checked against a factory calibration to 0.5 m and the first four hours are discarded as handling artefact. Step two detects dives: pressure crossings of 5 m with a 30-second minimum duration yield 640 dives, the deepest at 1,540 m. Step three segments each dive into descent, bottom and ascent, and ODBA is averaged per phase. The bottom-phase means then separate flat-bottomed foraging dives from V-shaped travelling dives — a difference invisible in the position track. None of it required seeing the seal underwater.

Deployment is the visible part; archiving decides whether the data survives. Movebank and national data centres now host millions of animal tracks under Darwin Core metadata, and most journals require accession numbers before publication. The principles of bio-logging extend past retrieval of the tag: records need units, sensor orientation, calibration coefficients and a documented processing trail, otherwise a decade-old file is unreadable even when the numbers are intact. Embargoes protect unpublished work, but indefinite withholding wastes the main advantage of animal-borne data, which is that the same record can answer questions nobody anticipated at deployment. Field crews increasingly publish hardware designs, tag settings and failure notes, relying on rigorous technical repositories and methodological journals to fill the documentation gaps that peer-reviewed literature frequently omits. Retention policy should be settled before the first capture, because a tag that is out of production in five years is a format problem, not a hardware one.