Deploying Acoustic Detection Algorithms on Low-Cost, Open-Source Acoustic Sensors for Environmental Monitoring

Abstract

Conservation researchers require low-cost access to acoustic monitoring technology. However, affordable tools are often constrained to short-term studies due to high energy consumption and limited storage. To enable long-term monitoring, energy and space efficiency must be improved on such tools. This paper describes the development and deployment of three acoustic detection algorithms that reduce the power and storage requirements of acoustic monitoring on affordable, open-source hardware. The algorithms aim to detect bat echolocation, to search for evidence of an endangered cicada species, and also to collect evidence of poaching in a protected nature reserve. The algorithms are designed to run on AudioMoth: a low-cost, open-source acoustic monitoring device, developed by the authors and widely adopted by the conservation community. Each algorithm addresses a detection task of increasing complexity, implementing extra analytical steps to account for environmental conditions such as wind, analysing samples multiple times to prevent missed events, and incorporating a hidden Markov model for sample classification in both the time and frequency domain. For each algorithm, we report on real-world deployments carried out with partner organisations and also benchmark the hidden Markov model against a convolutional neural network, a deep-learning technique commonly used for acoustics. The deployments demonstrate how acoustic detection algorithms extend the use of low-cost, open-source hardware and facilitate a new avenue for conservation researchers to perform large-scale monitoring.

Keywords: acoustics; bioacoustics; conservation; ecology; machine learning.

Figure 1

AudioMoth: a low-cost, low-power acoustic monitoring device developed for a wide variety of conservation projects, deployed on a tree in a grip-sealed bag using a cable tie.

Figure 2

An 8 kHz tone extraction with a single Goertzel filter.

Figure 3

Overview of the data collection and bat detection process. Samples are collected and stored in a partition of the SRAM. Once full, the collected samples are fed into a 60 kHz Goertzel filter in windows of 256. A sliding window is run over the resulting series of responses. The sliding window takes the median response from five Goertzel filters and compares it to a threshold. It then either triggers a recording or moves the window along.

Figure 4

Bat detection analysis on a recording of a soprano pipistrelle bat, collected by a deployed AudioMoth device. The two plots each show median responses of two sliding windows, (A) with the correctly configured five Goertzel outputs per window; and (B) with too many, at 15 per window. Bat calls are missed when over half the windows consist of background noise, rather than the call itself.

Figure 5

A receiver-operating characteristic curve showing the true-positive and false-positive rate produced by a range of thresholds when used by the bat detection algorithm.

Figure 6

Spectrogram showing the first target sound: echolocation calls of a soprano pipistrelle bat.

Figure 7

Spectrogram showing the second target sound—the song of the cicadetta montana.

Figure 8

Overview of the sample collection and cicada detection process. Samples are collected in a single partition of the SRAM containing a maximum of 8192 samples. Once it is full, the samples are fed through two Goertzel filters in windows of 128 samples. The outputs from these filters are zipped together and the ratio of each pair is calculated. The median value of these ratios is compared to a threshold to decide either to record or return to sleep.

Figure 9

Receiver operating characteristic curve showing the performance of the New Forest cicada detector at various thresholds. The final threshold was chosen by capping the false-positive rate to 0.01 and taking the threshold which corresponded with the highest attainable true-positive rate.

Figure 10

Spectrogram showing the third target sound—a gunshot recorded in the rainforest at a distance of 255 m.

Figure 11

Overview of the data collection and gunshot detection process. Samples are collected and stored in three partitions in SRAM. Once two have been filled, the detection algorithm is run while the third fills. Once the third has been filled, the algorithm is run again on the most recent two partitions. If a gunshot does not fit into a single partition, each partition is used twice meaning the full gunshot will be in the next iteration.

Figure 12

Gunshot detection model state diagram showing possible movement between five states: silence (S), noise (N), impulse (I), decay (D), and tail (T).