Protecting elephants
with sound.

Extended infrasound rumble sequence recorded adjacent to a diesel generator station. NMF separation achieved 94% noise reduction, revealing sustained low-frequency contact calls from a family group of ~6 individuals.

Long-form contact call sequence along a heavily trafficked corridor road. Vehicle engine noise masked infrasound energy; NMF restored 3 simultaneous callers. Dominant fundamental at 18 Hz with harmonics to 108 Hz.

Continuous infrasound rumble from a matriarch-led group near a water pumping station. Generator harmonics at 50 Hz and 100 Hz masked the fundamental; NMF component extraction isolated 5 elephant vocalisation components.

Rumble calls recorded beneath a commercial flight path. Aircraft broadband noise swept from 200 Hz downward through the infrasound range. NMF successfully separated the 12–20 Hz elephant components despite overlapping spectral content.

Contact calls from a herd crossing a park road. Vehicle idling and movement noise created a 40–300 Hz noise floor. Separation preserved the 14–28 Hz infrasound fundamental and reveals a chorus of 4 distinct individuals.

Infrasound recording from the Mara during a tourist aircraft overflight. This session established early validation data for NMF-based separation. Dominant call component at 16 Hz with amplitude modulation characteristic of long-distance signalling.

Third aircraft overflight session in a series of controlled experiments. Repeated separation of the same herd group across multiple noise events enabled cross-validation of NMF model stability. Call signatures matched across all three sessions.

Inaugural aircraft noise study session in Tsavo East. Elephant herd of approximately 14 individuals responded with heightened contact calling during aircraft passage. Infrasound energy at 14–22 Hz peaked during the aircraft approach phase.

High-intensity rumble from a male in musth condition, partially masked by adjacent generator. The sustained 50 Hz generator harmonic overlapped with the 3rd harmonic of the elephant call. NMF resolved this using temporal envelope differences.

Simultaneous aircraft and ground vehicle noise — the most acoustically challenging test case in this dataset. NMF decomposed 4 independent noise components plus 2 elephant signal components. Represents the performance ceiling for single-channel separation.

Archival 1986 recording — among the earliest infrasound captures with simultaneous aircraft interference. Tape-transferred at 44.1 kHz. NMF applied retrospectively; demonstrates technique applicability to legacy datasets.

Contact call sequence near a ranger patrol route. Vehicle diesel engine produced strong 100–300 Hz noise floor. After NMF separation, 2 distinct callers resolved at 16 Hz and 21 Hz; temporal overlap suggests coordinated calling between individuals.

Recording from the Luangwa Valley during a high-altitude commercial flight path. Broadband aircraft signature swept through 0–400 Hz over 90 seconds. NMF converged in 12 components; 2 attributed to elephant vocalisation at 15 Hz and 30 Hz.

Third aircraft-noise session in the Okavango series. Seasonal flood water created unusual acoustic reflectance. Infrasound propagation extended beyond typical range — estimated caller distance of 1.2 km based on amplitude attenuation modelling.

1989 Amboseli archival session — third aircraft interference recording from a multi-day field campaign. Original K. Payne and W. Langbauer infrasound study. NMF retrospective application validates the original manual spectrogram analysis from the period.

Greeting ceremony calls partially obscured by tourist vehicle traffic. The characteristic multi-individual greeting sequence shows overlapping fundamentals at 14 Hz, 18 Hz, and 23 Hz — consistent with 3 individuals reuniting after separation.

Dry season rumble sequence during high visitor traffic period. Despite noise conditions, NMF separation achieved clear isolation of a sustained 17 Hz rumble with 6 harmonics. Spectrogram shows characteristic frequency modulation of a close-contact rumble.

Second vehicle-noise session in the Tsavo longitudinal study. Consistent herd group identification across sessions enables individual-level analysis of repeated infrasound signatures. This recording matches the dominant caller from session 1 (sel_162).

First aircraft session from the 1989 Amboseli field campaign. Shorter aircraft transit produced a focused noise pulse. Retrospective NMF analysis reveals a rumble call almost entirely obscured in the original recordings — not previously identified.

Serengeti woodland boundary recording; denser vegetation reduces direct-path propagation. Aircraft noise cleaned via NMF; residual elephant signal shows an unusually wide frequency spread (0–150 Hz) typical of forest-edge conditions with acoustic scattering.

Extended infrasound session adjacent to a diesel generator station — second selection from the same recording block. NMF revealed an overlapping rumble sequence from a sub-group moving east; two callers distinguished by 3 Hz fundamental frequency offset.

Subsequent selection from the Tsavo corridor vehicle session. The road noise profile shifted as vehicles decelerated; NMF adapted its component weights accordingly. Reveals a sustained contact call at 19 Hz lasting nearly the full recording duration.

Second NMF selection from the Chobe generator-noise block. The matriarch rumble identified in sel_82 continues here at reduced amplitude, consistent with the group moving away from the microphone. Background estrus calling faintly visible at 28 Hz.

Follow-on selection from the Hwange aircraft study. Aircraft has passed; residual rumble from the herd continues. NMF shows the group resumed normal foraging calls at 14–20 Hz after a 45-second post-flyover silence captured in the previous selection.

Extended vehicle-noise selection from the Chobe riverfront block. Herd contact-call chorus peaks midway through; NMF isolates 5 simultaneous callers with fundamentals spanning 14–26 Hz. Temporal clustering suggests coordinated movement signalling.

Continuation of the Mara aircraft isolation session. Aircraft engine frequency drops as it recedes; NMF successfully tracked the changing noise profile. Reveals a second-caller rumble at 22 Hz not detected in the prior selection — likely a juvenile.

Next selection in the Tsavo triple-flyover series. Herd call intensity increased 3 dB relative to the prior overflight event — consistent with escalating arousal response. NMF isolated the same dominant caller signature as sessions 1 and 2.

Second selection from the inaugural Tsavo flyover study. The aircraft has departed; infrasound energy from the herd drops from peak but a sustained low-amplitude rumble continues — characteristic of post-disturbance reassurance calling.

Follow-on selection from the musth male recording. The 50 Hz generator harmonic persists; NMF now confidently separates it given the prior-selection model. The musth rumble's amplitude modulation depth increased by 40% — consistent with approaching a female group.

Follow-up to the most challenging mixed-noise case. Both aircraft and vehicle noise persist at lower amplitudes. NMF separation quality improves as noise becomes more stationary; elephant call components at 15 Hz and 31 Hz now clearly resolved.

Second selection from the Gonarezhou vehicle-patrol study. Contact calling between two separated sub-groups — NMF unmasks a directional call response pattern consistent with coordinated re-grouping behaviour at the 17 Hz and 22 Hz fundamentals.

Continuation of the Luangwa aircraft study. The broadband sweep from the prior selection has passed; steady-state aircraft noise remains. NMF isolation quality at its best in this session — shows a clear 15 Hz rumble call with 5 harmonics up to 75 Hz.

Second NMF selection from the 1989 Amboseli third-session archive. Aircraft noise quieter here; separation resolves a previously hidden 20 Hz call. Cross-referencing with original field notes confirms this was a known 'Let's go' rumble — a rare archival recovery.

Second selection from the Queen Elizabeth greeting-ceremony block. Vehicle traffic has cleared; NMF reveals the tail end of a greeting sequence — individual calls converging on a shared 16 Hz fundamental with diminishing amplitude as the group settles.

Continuation of the dry-season Queen Elizabeth session. NMF shows the rumble transitions to a higher-pitched contact call at 24 Hz — consistent with the group beginning directed movement. A brief musth rumble at 8 Hz is faintly visible in the lower spectrogram band.

Follow-on selection from the second Tsavo vehicle session. The dominant caller from sel_192 continues; NMF now shows a second, quieter individual joining at 20 Hz — possible calf given the higher fundamental frequency relative to the adult at 15 Hz.

Second NMF selection from the 1989 first aircraft session. The aircraft noise pulse has attenuated; the rumble call previously partially masked is now fully resolved. Frequency modulation sweep from 16 Hz to 18 Hz over 4 seconds — characteristic of long-distance calling.

Subsequent Serengeti woodland selection. Aircraft echo off a tree-line is visible as a delayed spectral ghost; NMF correctly attributes this to the noise component. The elephant call at 13–18 Hz persists throughout — an unusually long sustained call for open savannah.

Combined vehicle and generator noise — a dual anthropogenic source test case from Chobe. NMF decomposed 3 noise components and 2 elephant signal components. The vehicle contributes 40–300 Hz noise while the generator adds discrete harmonics; separated cleanly in the time-frequency domain.

Baseline generator-noise session establishing NMF training parameters for the full 2000-24 dataset. Consistent generator tone at 50 Hz provides a clean noise reference. Elephant call at 16 Hz is faintly present — confirms caller was within 500 m of the microphone.

June 1989 Amboseli archive — shorter aircraft pass produces a narrow-band noise spike rather than a broadband sweep. NMF isolates the 15 Hz elephant rumble with exceptional clarity; this selection is frequently used as a benchmark for infrasound separation algorithms.

Fourth vehicle-noise selection from the Gonarezhou patrol study. A single short contact call is visible at 18 Hz — brief, high-amplitude pulse typical of close-proximity alerting. Vehicle noise reduces to near-silence at the end of the recording.

Short aircraft-noise event in Samburu. The flyover is brief; NMF required only 6 components to achieve clean separation. The residual elephant signal at 17 Hz lasts the full 69 seconds — an unusually long sustained call from a single individual.

Combined background environmental noise and distant vehicle. The lowest signal-to-noise ratio in this dataset. NMF required 10 components; separation confidence is moderate. Despite this, a 16 Hz rumble call is recoverable — demonstrating NMF robustness at the noise floor.