Investigating Target Class Influence on Neural Network Compressibility for Energy-Autonomous Avian Monitoring

Biodiversity loss poses a significant threat to humanity, making wildlife monitoring essential for assessing ecosystem health. Avian species are ideal subjects for this due to their popularity and the ease of identifying them through their distinctive songs. Traditionalavian monitoring methods require manual counting and are therefore costly and inefficient. In passive acoustic monitoring, soundscapes are recorded over long periods of time. The recordings are analyzed to identify bird species afterwards. Machine learning methods have greatly expedited this process in a wide range of species and environments, however, existing solutions require complex models and substantial computational resources. Instead, we propose running machine learning models on inexpensive microcontroller units (MCUs) directly in the field. Due to the resulting hardware and energy constraints, efficient artificial intelligence (AI) architecture is required. In this paper, we present our method for avian monitoring on MCUs. We trained and compressed models for various numbers of target classes to assess the detection of multiple bird species on edge devices and evaluate the influence of the number of species on the compressibility of neural networks. Our results demonstrate significant compression rates with minimal performance loss. We also provide benchmarking results for different hardware platforms and evaluate the feasibility of deploying energy-autonomous devices.

💡 Research Summary

The paper addresses the challenge of performing real‑time bird‑song identification on ultra‑low‑power microcontroller units (MCUs) for autonomous, energy‑self‑sustaining wildlife monitoring. Recognizing that traditional point‑count surveys and cloud‑based acoustic analysis are costly, labor‑intensive, and unsuitable for large‑scale deployments, the authors propose an edge‑AI solution that processes audio directly on the device, thereby eliminating the need for continuous data transmission and reducing overall energy consumption.

A comprehensive dataset was assembled primarily from Xeno‑Canto, selecting 500 bird species based on data availability and regional relevance (German, then broader European, then global). For each species, 250 recordings were randomly chosen, and an additional non‑bird class was created by merging the 49 environmental categories of the ESC‑50 dataset. Recordings shorter than two seconds were discarded, silent sections were removed, and the remaining audio was segmented into 2‑second chunks (up to 30 chunks per recording). Each chunk was normalized and transformed into a mel‑spectrogram (48 kHz sampling rate, 64 mel bands, FFT window 512, hop length 384, frequency range 150 Hz–7.5 kHz). This preprocessing yielded roughly 2 500 chunks per species with considerable variance, reflecting realistic field conditions.

To increase robustness, four augmentation techniques were applied with a 50 % probability per chunk, up to three augmentations per chunk: vertical frequency shifts (±5 %), horizontal time shifts (±25 %), SpecAugment‑based time warping, and addition of background noise (20 %–80 % intensity). These augmentations simulate environmental variability such as changes in vocal output, ambient noise, and recording conditions.

The model architecture chosen is MCUNet‑in4, a lightweight convolutional neural network specifically designed for MCU inference. It consists of an initial convolutional layer, 17 MobileInvertedResidualBlocks, and a final linear classifier. The first layer was adapted to accept a single‑channel mel‑spectrogram, and the output layer was sized to match the number of target classes (including a non‑bird class). Pre‑trained ImageNet weights were transferred to all layers except the first and last.

Compression was performed using an interleaved pruning and post‑training quantization pipeline (based on the method of