From Sound to Satellite How We Combine Bioacoustics and Remote Sensing for Biodiversity Insights

Topic was presented at Munich Datageeks - May 2026

Abstract

This talk introduces Huula, a Munich-based biodiversity-monitoring startup, presented by two members of its team: one focused on hardware and product, the other on machine learning and data science. The talk opens with the evolution of Huula's edge device, the BioT, a solar-powered bio-acoustic sensor built from scratch through several rough prototype iterations into its current production design. The machine-learning speaker then explains how bird calls are classified directly on the device using lightweight, mobile-optimized image-classification models run on audio spectrograms, refined with a hidden Markov model that turns noisy per-snippet predictions into a clean presence/absence signal over time. A cloud-side validation and expert-labeling loop feeds corrected data back into model retraining. Two case studies follow: acoustic biodiversity monitoring for a solar park under German nature-protection regulation, and satellite-based remote sensing for rainforest conservation in the Amazon. The talk closes with early attempts to combine bio-acoustic and remote-sensing data, and broader reflections on building hardware, iterating quickly, and treating the surrounding data pipeline as seriously as the model itself.

About the speaker

David is part of the founding team at Huula and works on the company's hardware and product side. He notes he does not come from a hardware background himself, but had to learn it out of necessity because nothing suitable existed to buy — Huula's BioT sensor was built from scratch, starting from an early duct-taped prototype and iterating through several hardware revisions to reach its current, more manufacturable design. He closes the talk with reflections on hardware development and general startup-building principles.

Konrad is a mathematician by training and leads machine learning and data science work at Huula. His part of the talk covers how bird calls are detected from audio on constrained edge hardware, how a hidden Markov model is used to add temporal context to raw model predictions, and how satellite remote sensing is used to monitor forests, culminating in early ideas for combining the two data modalities.

Transcript summary

Huula and the evolution of the BioT device

Huula builds hardware for biodiversity monitoring, starting with birds because they are a strong proxy for ecosystem health: they are highly sensitive to environmental stress and, conveniently, vocalize frequently, which makes them easy to record. A large community of hobbyist ornithologists has also freely labeled bird-call data online, which made it possible to train detection models. The company does not intend to stop at birds, but that is where it started.

Because no suitable off-the-shelf hardware existed, Huula had to design its own device, the BioT, entirely from scratch. The device evolved through several visibly rough iterations: an early, duct-taped prototype good enough only to demonstrate the concept and convince early believers; a second version deployed on an actual customer site, still visually unrefined; a further version built without a clear enclosure lid; the first version the team designed properly in-house, which included an updated solar panel and more manufacturable internals; and the current production design, which the speaker physically shared with the audience after the talk. The team expects this current design to be superseded within about a year and considers the device's evolution an ongoing iterative process, aimed at scaling from around 100 deployed BioTs today to roughly 1,000.

From raw audio to a computer-vision problem

Audio is a difficult data modality to work with directly: a single 3-second recording snippet contains around 140,000 samples, far more than is practical to push through a machine learning model on constrained edge hardware. The standard approach — used by Huula and the wider industry — is to convert each audio snippet into a spectrogram, a time-by-frequency representation of the sound. Once represented this way, bird-call detection becomes effectively a computer-vision problem, a much better-studied space with many existing model architectures to draw on.

Huula applies an image classifier (a convolutional neural network) to each spectrogram, producing a confidence score for whether each bird species of interest is present or absent in that 3-second snippet.

Constraints of running AI directly on the device

Running the AI model on-device, rather than in the cloud, is a deliberate choice driven by two factors: streaming continuous audio from hundreds, thousands, or eventually tens of thousands of devices to the cloud would be prohibitively expensive, and processing raw audio (which may contain non-consenting human speech) off-device raises privacy concerns. Huula is therefore committed to processing audio locally and streaming only the resulting detections.

This creates a very tight design space on microcontroller-class hardware: RAM is measured in megabytes rather than gigabytes, at a time when mainstream models are moving from billions toward trillions of parameters — meaning Huula's models must be roughly a thousand times smaller, in the millions of parameters. Model execution time is also constrained, since a 3-second audio snippet must be processed in under 3 seconds to keep up with incoming data. Quantization is another limiting factor, since on-device chips only accelerate very narrow data types such as int8. Finally, the hardware cannot support more complex model operators such as attention layers or squeeze-and-excitation blocks — only basic layers and ReLU activations are available.

Given these constraints, Huula uses models from the MobileNet family, which are well suited to microcontroller deployment. This choice was informed by plotting every model in the timm (PyTorch Image Models) library by parameter count against ImageNet top-5 accuracy, then identifying which of those models are actually viable within Huula's tight memory and compute budget — a small subset dominated by MobileNet variants.

Adding temporal context with a hidden Markov model

A single 3-second snippet only captures a momentary observation, but bird presence is temporally correlated — a bird detected now is likely still present moments later. To capture this, Huula models bird presence as a simple two-state statistical system (present or absent), with estimated probabilities of transitioning between states from one time step to the next (staying present, leaving, reappearing, or staying gone).

Because the underlying image classifier produces false positives and false negatives, the raw per-snippet predictions are noisy. To address this, Huula applies a hidden Markov model, a well-studied statistical model that allows efficient, closed-form inference and lets all its parameters be learned directly from data, while still incorporating time into the prediction. In practice, applying the hidden Markov model to a day's worth of raw per-snippet predictions and then extracting the most likely underlying sequence of states (rather than just smoothing the raw output) produces a much cleaner signal of when a bird was actually present — this cleaned-up sequence is what Huula uses as its final prediction.

The full data pipeline, from raw audio to customer-facing insight

Huula's end-to-end pipeline starts with raw audio captured on the device. An edge AI model performs two tasks: detecting which birds are present, and running a high-recall human-voice detector; if there is even a slight indication of human speech, the audio is discarded immediately for privacy reasons. Audio without detected human speech is stored in a local, on-device detection database.

Roughly every one to two hours, each device syncs its local detection database to Huula's cloud database. There, a further validation layer checks whether each detection is plausible for that location and time of year (accounting for migratory birds that are only present in certain seasons), and in some cases pulls the underlying audio for expert biologists to double-check ambiguous detections and confirm the model is not "hallucinating" birds. Only after this validation does data reach the customer-facing product, referred to internally as the Huula platform.

Crucially, this validation process feeds back into model improvement: once detections have been cross-validated by location, time, and expert listening, that corrected data is used to retrain and improve future iterations of the model, creating a continuous feedback loop.

Expert labeling interface

To support this feedback loop, Huula built a labeling interface for expert biologists, showing the spectrogram of a detection alongside relevant metadata. Experts can either confirm ("greenlight") a detection or annotate what is wrong with it. This interface has proven useful both for the biologists doing the validation work and for Huula, since it generates additional high-quality training data for future model iterations.

Results in numbers

Huula's daily bird detection volume has grown substantially: from roughly a couple of thousand detections per day the previous year to around 15,000 detections per day on a good day this season. Total recorded audio volume has reached roughly 10 terabytes — for comparison, a recently reported Spotify audio archive was around 300 terabytes, suggesting Huula's volume could reach comparable scale within a couple of years. Streaming this volume of raw audio over typical IoT mobile network contracts would be prohibitively expensive, reinforcing the case for on-device processing.

A common question about small, edge-deployable ("tiny ML") models is how much accuracy is sacrificed compared to larger models. The benchmark used for comparison is BirdNET, from Cornell Lab (together with other institutes), widely regarded as the gold standard for bird-call detection — a model roughly 24 times larger than what Huula runs on its hardware. Rather than comparing raw accuracy, Huula fixes precision at a high standard (around 99%, meaning that when the system reports a bird is present, it should be correct 99% of the time) and then compares recall at that precision level. On this metric, Huula's much smaller model is competitive with BirdNET: both achieve roughly 40% recall at 99% precision.

Case study: acoustic monitoring for a solar park

In Germany, building a solar park requires strict monitoring and regulatory compliance: which bird species were present before construction, which are present afterward, which left, and how the operator must accommodate for that. The conventional approach is for biologists to visit the site five to six times over a single spring breeding season, each visit lasting around two hours, following a detailed species-specific protocol (for example, needing to consult a lengthy, separate reference for each individual species, such as the blackbird). This is a labor-intensive, largely manual process — and a clear opportunity for Huula's approach.

Huula deployed a grid of BioT sensors across a solar park, chosen specifically because a grid (rather than a single sensor) allows triangulating the origin of a bird call once it is picked up by at least three sensors, enabling localization rather than just presence detection. In this deployment, the team recorded around 77,000 calls from the skylark, one of the more prominent species observed. Mapping these localizations revealed clear spatial patterns, including a cluster near the top of the solar park and activity in peripheral regions, as well as a clear seasonal pattern where detections drop off after July, coinciding with the end of the breeding season.

These raw detections are then translated into the specific reporting format regulators require, using rule-based logic (for example, a rule requiring at least two observations with a minimum of seven days between them) implemented as SQL-like queries that cluster detections into distinct breeding sites. This produces the deliverable nature authorities actually need: where birds are breeding, how many are breeding at each site, and how this changes over time. The skylark is a special case in this reporting, since it breeds twice within a single season (a first and second brood). A second example species, the yellowhammer (Goldammer in German), was also observed clustering its breeding activity around trees within the solar park, illustrating how individual species carve out habitats even within a relatively small, engineered ecosystem.

Remote sensing case study: Amazon rainforest conservation

Alongside bio-acoustics, Huula also offers remote sensing insights to all its customers, most notably demonstrated through a recent project with a conservation organization working in the Amazon rainforest. The underlying data source is ESA's Sentinel-2 satellite imagery, at roughly 10 meters per pixel resolution with a new image roughly every five days. In the example shown, satellite imagery revealed a largely undisturbed tropical rainforest area alongside a heavily deforested region near a highway, where forest had been cleared for plantations and other uses.

The organization's core question was which parcels of land would be most valuable to place under conservation: protecting land already far from any threat has limited impact, whereas protecting land adjacent to active deforestation has a much larger effect. Huula's response combined several layers of analysis:

1. Simple spectral indices — starting with the Normalized Difference Vegetation Index (NDVI), a straightforward calculation based on the near-infrared and red color bands that is highly sensitive to chlorophyll. This requires no machine learning at all but already clearly distinguishes healthy forest (strongly vegetated) from plantations or rivers (which show little to no vegetation signal), and is often useful to customers on its own.
2. Land-cover segmentation — a relatively small, shallow neural network (in the tradition of segmentation architectures like U-Net, though much smaller than typical large foundation models are actually needed for this domain) classifies each area of the image into roughly ten discrete categories, such as water, forest, and low vegetation.
3. Forest connectivity — a metric proposed by the European Commission's Joint Research Centre, calculated by aggregating the forest classification map and scoring each location based on how much forest surrounds it; more surrounding forest yields a higher connectivity score.
4. Change detection over time — comparing forest maps from different years (for example, 2024 versus 2025) highlights where forest has disappeared (marked in red) versus where new forest has appeared (marked in green), directly surfacing active deforestation.
5. Deforestation risk prediction — a predictive time-series model estimates which regions are most likely to be deforested over the next three to five years, showing near-zero risk for well-protected interior forest and much higher risk near existing clearings and infrastructure such as highways.
6. Above-ground biomass via LiDAR — for tracking the health of existing, undisturbed forest, biomass is estimated using LiDAR data from the GEDI instrument in space. Because GEDI only captures data along narrow individual tracks rather than a dense grid, Huula uses machine learning models to extrapolate from these sparse GEDI tracks, combined with optical satellite imagery, to produce a full-coverage biomass map — giving the customer an inventory of stored carbon and overall forest health across the parcel.

Combining bio-acoustics and remote sensing

Huula is still in the early stages of exploring how its two main data types — hyperlocal bio-acoustic detections and spatially distributed remote sensing data — can be combined, but two directions are already showing promise.

The first is extrapolating species habitats. In an example just north of Munich, an array of microphones detects a kingfisher at only two of several sensor locations. Without any additional context, this alone suggests a small, localized habitat around those two sensors. However, cross-referencing this with remote sensing data reveals the Isar river running past those exact sensor locations, making it far more plausible that the bird's actual habitat is a much narrower band that runs elongated along the river, rather than an isolated area around the two detection points.

The second direction is exploring "what-if" scenarios: correlating bird observation counts (for example, per day or week) with environmental factors such as temperature and precipitation reveals noisy but real trends. This makes it possible to explore, for example using climate projections, what might happen to species such as skylarks or blackbirds under a given amount of warming — helping to understand how resilient a given ecosystem is to climate change, model how ecosystems might shift under continued warming, and identify concrete interventions (such as planting specific tree species or reintroducing other species) that could increase an ecosystem's resilience.

Startup and hardware lessons

Returning to close the talk, David shares several broader takeaways from building Huula, starting with hardware. Hardware is fundamentally trickier to work with than the digital systems most attendees are used to: software failures are traceable and infinitely repeatable, and while not always easy to debug, they are debuggable. Hardware deployed in the field is a different kind of problem — a device can simply stop connecting to the cloud with no way to intervene remotely, or fail due to physical causes (such as water damage on a contact) that may no longer be diagnosable by the time the device is returned and opened up. This was a significant adjustment, since David does not come from a hardware background himself.

Several broader startup lessons, not necessarily unique to Huula, were also highlighted: doing things that don't scale early on (in Huula's case, this specifically meant frequent field visits); working only with paying customers from the start, since a customer who isn't paying provides no reliable signal about whether the product is actually valuable to them; and building systems early even while knowing they won't hold up long-term, then being willing to iterate quickly and discard work — a mindset the speaker connects to the growing usefulness of agentic development workflows, which make this kind of fast iteration and willingness to throw work away significantly easier.

A recurring theme throughout the talk is that a raw bird call detection is not, by itself, a biodiversity insight. The team initially assumed, naively, that sufficiently accurate models might let them simply trust raw detections directly in the cloud-facing product. In practice, this doesn't hold: even a 1% error rate across a million detections means 10,000 potentially wrong data points, and if a customer questions whether a rare species was genuinely detected, being unable to substantiate that erodes trust. This is why systems such as the hidden Markov model, spatial extrapolation, and expert-in-the-loop validation matter — as summarized in the talk, the surrounding pipeline and system matter as much as the model itself.

For the remote sensing side, the lesson is to always start with simple, interpretable approaches (such as basic vegetation indices) first, since they are fast to produce, generate immediate learnings, and provide a baseline for comparison — with the greatest value coming from combining multiple data layers. Huula is currently working toward combining remote sensing layers with INC2 data as an extension of this principle, though the same logic already holds even when working with satellite data alone. The speaker frames this as "remote sensing scales, in-situ sensing grounds" — remote sensing provides broad geographic coverage, while ground-based sensors like the BioT validate what is actually happening. This is also seen as the path to economic viability: Huula does not believe it needs a BioT sensor placed every kilometer of the earth, but rather enough coverage to learn reliable patterns, after which a much wider, sparser sensor grid should suffice — which in turn lowers hardware costs at scale and strengthens the overall business case.

Vision: sensing, reasoning, and action layers

Huula's longer-term vision is structured around three layers. The sensing layer is what the company is currently building and scaling — the BioT is the most concrete example, but the term also includes other sensing sources such as satellite imagery, which Huula doesn't own directly but integrates, and eDNA data, which the company already collects and stores even though it isn't yet actively used. The rationale for storing this data now, even before it is put to full use, is the belief that accumulating this breadth of data will eventually provide the foundation needed to build a much stronger reasoning layer — noting that today's landscape of "foundation models for earth observation" remains underexplored and tends to rely on satellite data alone, whereas Huula's ambition is to build a foundational model for life on Earth more broadly, a goal expected to take several years.

The action layer is what Huula does with the resulting insights: using them to help positively influence the ecosystems being measured. Part of this involves incorporating decades of existing ecological science — tooling and indices for which the team is currently building out — using techniques similar to those discussed elsewhere at the event. The long-term ambition is to combine Huula's own data with this existing scientific knowledge to produce concrete, actionable measures that customers can take to improve outcomes, which in turn improve the ecosystems Huula measures — closing the loop into a long-term flywheel. The speaker is clear that this remains a long-term vision that the company is still far from completing.