Multimodal evidence fusion for biodiversity monitoring
Brief · published 6 May 2026 · Naturecode Project
Biodiversity monitoring runs on instruments that rarely speak to each other. A satellite sees a forest from above. A microphone hears the same forest at dawn. A water sample carries the DNA the forest sheds into a stream. A ranger watches a clearing for the species that hold the system together. Each instrument is honest. Together they are far more honest than any one of them alone.
This brief argues for multimodal evidence fusion as the working stance for biodiversity monitoring — and outlines what an open, practitioner-usable monitoring stack would look like.
The instruments
- Satellite imagery — extent, change, productivity, structure, surface temperature.
- Acoustics — species presence, soundscape integrity, dawn and dusk activity, cryptic species.
- Environmental DNA — what is in a water sample, a soil sample, a swab.
- In-situ telemetry — weather stations, soil moisture, river gauges, camera traps, tagged animals.
- Field observation — the people watching the place, week after week.
Each instrument carries its own bias, its own coverage gaps, its own noise floor. Fused carelessly they amplify each other's mistakes. Fused carefully they describe a system that no single instrument could.
What honest fusion needs
- Aligned in time and place. Observations that disagree at the same moment are signals — but only if you can find them.
- Aware of what is missing. A monitoring system has to know where it has not looked, not only what it has seen.
- Traceable through the model. When a fused estimate is produced, the contributions of each source should be inspectable, not hidden inside model weights.
- Open to other instruments. New sensing methods arrive every year. The stack has to absorb them without rebuilding from scratch.
- Useful to a working ecologist. A monitoring tool that only researchers can read is incomplete.
The work ahead
- Open benchmarks that test multimodal models on real biodiversity questions, not contrived ones.
- Pipelines that can be run by working programs — protected-area authorities, fisheries, restoration projects — without requiring a research team.
- Methods for combining sources whose ground-truth conventions disagree.
- Honest reporting of what fusion does and does not improve.