BeeBetter · Hamze Hammami

The Problem

Honeybee colonies are disappearing at an alarming rate.

Honeybees are the planet's most critical pollinators, responsible for the majority of crop yields and wild flora. Over the past decade their colonies have seen accelerating decline due to biological threats, environmental stress, and a beekeeping industry still relying on technology from the 1860s.

40% yearly increase in colony decline across reported regions

50% drop in honey production over the same period

80%+ of beekeepers reported fewer honeybees within a ten year period

Key Threats

Colony decline is driven by several compounding threats. The Varroa destructor mite spreads viruses and causes physical deformities that progressively weaken colonies. Invasive hornets and beetles attack bees directly and damage hive structures. When a colony loses its queen, bees exhibit distinct acoustic distress patterns that, without early detection, lead to hive abandonment. These biological pressures are compounded by seasonal stress, with losses peaking in winter when hives are hardest to inspect.

The Technology Gap

"The Langstroth hive was invented in the 1860s and has remained the standard choice for beekeepers worldwide. Why hasn't this technology advanced?"

While innovations from the same era have long been superseded, beekeeping practice remains a basic wooden box requiring manual inspection and often left entirely unattended. IoT systems, machine learning, and robotics are the natural next step. The absence of automated monitoring means threats go undetected until it is too late for the colony, and beekeepers have no real-time visibility into hive conditions between visits.

BeeBetter Core

A multi-modal machine learning system for hive health and hazard detection.

BeeBetter combines computer vision and audio classification through three models working together: YOLOv8 for bee and predator detection, VGG-Net for individual bee health classification, and a custom CRNN for queen state detection from hive acoustics. IoT sensors track temperature, CO₂, weight, and air quality, feeding a continuous early-warning system for colony threats.

bee 0.96

Visual Detection

YOLOv8 correctly identifies the honeybee among four visually similar insects. A wasp, moth, and fly are flagged as impostors despite their yellow-and-black markings.

Acoustic Analysis

Short-Time Fourier Transform reveals clear frequency differences between a queen-right and a queen-less colony, forming the input to the CRNN classifier.

CRNN Architecture

Three convolutional blocks extract spectral features from the STFT input, followed by a bidirectional LSTM that captures temporal patterns and outputs a queen state classification.

Queen State Detection · Live

Since the initial paper the CRNN has been further trained and validated. The video below shows actual real-time queen state detection running on hive audio recordings.

BeeVe · Sound Module Extension

BeeVe

Early Human Studies

Humans have been listening to bee sounds for over a century. Early attempts to record animal acoustics date to the late 19th century, and a landmark moment came when Karl von Frisch decoded the waggle dance in the 1940s, demonstrating that bee behaviour encodes structured spatial information and earning the Nobel Prize in 1973. Later work by Kirchner (1993) confirmed acoustic communication within hives, and researchers including Ferrari et al. (2008) and Kanelis et al. (2023) identified acoustic correlates for swarming and queen loss using supervised classifiers trained on labelled recordings. BeeBetter's CRNN follows this same direction. The gap is that all prior methods assume labels, predefined conditions, and human annotation. They answer specific questions but cannot discover anything the researcher did not already know to look for.

Can honeybee buzzing be decoded without any labels?

Honeybee buzzing is not a language. It arises from mechanical muscle vibrations reflecting the collective physiological condition of the colony, not communicative intent. No existing bioacoustic framework applies to it, yet evidence suggests the signal carries structured information about colony health.

BeeVe asks whether repeatable acoustic patterns exist in this signal and whether they can be discovered entirely without supervision. Using PaSST as a frozen audio encoder, a Vector-Quantized Variational Autoencoder is trained on five hours of hive recordings with no labels used at any stage. The result is a discrete codebook, a learned vocabulary of recurring acoustic patterns that emerges purely from data.

UMAP Latent Space

Architecture · Methodology

From raw hive audio to non-semantic acoustic states.

The pipeline begins with the UrBAN dataset, five hours of real-world hive recordings, and passes each file through PaSST, a frozen spectrogram transformer pretrained on AudioSet, to extract 1295-dimensional feature vectors at 23ms intervals. A Vector-Quantized Variational Autoencoder then learns to compress these into a finite codebook of discrete tokens entirely without labels. The resulting vocabulary is decoded and validated post-hoc against known queen status, revealing structured separation without any supervision.

Training Objective

The VQ-VAE is trained using a composite loss consisting of a reconstruction term and a combined quantization term. The outer weight λ = 0.1 ensures the reconstruction term remains the dominant training signal, with quantization kept as a regulariser to prevent the codebook from collapsing in early training. Training proceeds in two phases: for the first ten epochs only the reconstruction loss is active, allowing the encoder and decoder to establish a meaningful representation before quantization is introduced. After epoch 10 the full loss is applied.

Unsupervised Learning

No labels are used at any stage of training. The codebook is updated by exponential moving average with decay α = 0.99. A diversity loss encourages uniform utilisation of the codebook through entropy regularisation, preventing codebook collapse. Queen status is used only for post-hoc evaluation.

Core Methods

UrBAN Dataset Scientific Data 2025

Acoustic data equivalent to approximately five hours of hive recordings was sampled from the UrBAN dataset. The UrBAN dataset contains well over 1000 hours of honey-bee hive audio collected under real-world conditions. Annotations are used only to ensure that recordings from multiple conditions are represented in the experiment and are not used as labels for training.

PaSST · Patchout Spectrogram Transformer arXiv 2110.05069

Audio is loaded at 22,050 Hz before feature extraction. The model variant passt_s_swa_p16_128_ap476 is used with pretrained weights and no fine-tuning. Timestamp embeddings are extracted at approximately 23 ms intervals, yielding one 1295-dimensional feature vector per frame.

Model Architecture

Encoder

Compresses PaSST embeddings into a 128-dimensional continuous latent representation through five fully connected blocks (1295 → 1024 → 512 → 512 → 128) with LayerNorm, GELU activation, dropout, and a residual connection at the 512 → 512 stage.

Vector Quantizer

Maps the latent representation to the nearest entry in a learned codebook C = {e₁, e₂, …, eₖ}, where K ∈ {32, 64} depending on the experiment. Encoder outputs and codebook entries are L2-normalized before distance computation. Gradients bypass the non-differentiable quantization step via straight-through estimation, and the codebook is updated by EMA with decay α = 0.99.

Decoder

Reconstructs the original 1295-dimensional space from the quantized representation through four stages (128 → 512 → 512 → 1024 → 1295) with LayerNorm, GELU, dropout, a residual connection at the first stage, and a final linear projection without activation.

Experimental Results

Three experiments. Consistent structure.

Three full-scale experiments were conducted to assess stability across codebook sizes and random seeds. E1 baseline uses five hours of data with a codebook of 64 and seed 0. E1 baseline seed1 repeats the baseline with a different random seed. E2 small codebook uses three hours of data with a codebook of 32.

State Validation

The UMAP projection shows queenright frames forming a single connected manifold (green) while queenless frames appear as three isolated clusters (red) spatially separated from it, with no label supervision at any stage of training. JSD values range from 0.609 to 0.688 across all experiments, confirming that the two conditions occupy distinctly different regions of the learned token space. The scattered red points within the green mass represent the queenless outliers, below 2% across all experiments, indicating the model can generally differentiate between the two conditions.

2D latent projections coloured by queen status

State Validation Metrics Across All Experiments

Experiment	Condition	JSD	Active Tokens	Entropy (bits)	Top Token (%)	Silhouette	QNL Outliers
E1 baseline	queenright	0.609	13/64	2.042	39.04	0.046	1.57%
E1 baseline	queenless	0.609	5/64	1.134	58.00	0.046	1.57%
E1 baseline seed1	queenright	0.688	13/64	2.210	27.68	0.016	1.57%
E1 baseline seed1	queenless	0.688	6/64	1.187	56.30	0.016	1.57%
E2 small codebook	queenright	0.663	16/32	2.398	19.94	0.188	1.70%
E2 small codebook	queenless	0.663	6/32	1.247	56.45	0.188	1.70%

Queenless Sub-states

The PCA projection of queenless-only embeddings reveals three spatially distinct regions corresponding to sub-states A, B and C, confirming that the queenless condition is not acoustically uniform. Sub-state A (57.6%) occupies the left of the PC1 axis, Sub-state B (22.0%) sits in the middle, and Sub-state C (20.4%) extends to the right. PC1 alone captures 73.9% of the variance within the queenless condition. The UMAP projection overlays all embeddings, with queenright in grey and the three queenless sub-states as coloured clusters. Each sub-state forms an isolated region completely separated from the queenright manifold, discovered without any supervision.

PCA projection of queenless embeddings coloured by sub-state

UMAP projection coloured by QNL sub-state assignment

Sub-state Statistics Across All Experiments

Experiment	Sub-A Size	Dom. Token	Purity	Sub-B Size	Dom. Token	Purity	Sub-C Size	Dom. Token	Purity
E1 baseline	57.6%	T0	97.5%	22.0%	T10	53.6%	20.4%	T19	90.8%
E1 baseline seed1	57.4%	T1	97.9%	23.0%	T8	88.7%	19.5%	T5	73.1%
E2 small codebook	57.1%	T12	97.7%	21.6%	T3	89.3%	21.3%	T11	41.9%

Token Temporal Structure

The chi-squared independence test rejects the null hypothesis of token independence at p ≪ 0.001 across all experiments. Outgoing transition entropy averages 2.08–2.42 bits against a maximum possible 3.70–3.91 bits (H/H_max = 0.56–0.65), indicating structured but non-deterministic transitions. Approximately half of all frame-to-frame transitions are self-transitions (51–58%), and when transitions do occur they follow structured pathways rather than distributing randomly across the codebook.

Token transition probability matrix · baseline experiment

Temporal Token Sequence Statistics

Experiment	Active Tokens	Self-transitions	Transition Entropy H (bits)	H/H_max	Chi-squared
E1 baseline	15/64	51%	2.35 ± 0.86	0.60	p ≪ 0.001
E1 baseline seed1	13/64	54%	2.08 ± 0.74	0.56	p ≪ 0.001
E2 small codebook	13/32	58%	2.42 ± 0.58	0.65	p ≪ 0.001

Explore the BeeVe Visualizer

Interactive 3D latent space, token gallery, real audio playback and transition analysis.