01 / LISTEN
Bioacoustics
Solar recorders listen around the clock. Every call, chorus and engine becomes a timestamped, searchable signal with coordinates attached.
[ 01 · THE SIGNALS ]
The forest never stops talking.
Every dawn chorus, every rain front, every chainsaw is a broadcast. Canopy puts a receiver in the understory and another one in orbit, then lets the two argue until only the truth is left.
02 / SUSPECT
Open-set detection
Most models only recognise what they were trained on. Canopy flags sounds it has never heard before, because the rarest signal is often the one that matters.
03 / CONFIRM
Satellite cross-check
Sentinel-2 passes confirm what the microphones suspect. Acoustic anomalies are matched against canopy loss within days, not months.
[ 02 · FIELD TEST ]
Run the detector. Right here.
This is not a mock. The page loads Canopy's edge model, a 94 KB CNN embedder and its open-set head, and runs it inside your browser. Nothing you play or record leaves this tab. Pick a signal, or give it one of your own.
INPUT SIGNAL
DETECTOR: anomaly-v1 · EMBEDDER: threat-cnn-forest-v1b · BASELINE: 569 forest background windows · RUNS LOCALLY VIA WASM
DETECTOR OUTPUT
--
ANOMALY SCORE
STANDING BY
0 DEPLOYED ALERT LINE AT 0.99 1
MODEL LOADS ON FIRST RUN
[ 03 · THE POSITION ]
Most systems classify what they know. Canopy listens for what it does not.
[ 04 · OPEN CHANNEL ]
Talk to the people building it.
Canopy is open source and already on real land. If you hold ground worth listening to, or you work on bioacoustics and want the protocol, the channel is open.
01 / PILOT
Put a unit on your land
Forest departments, reserves, plantations and farms. We arrive with hardware and leave with a detector calibrated to your site, not somebody else's.
02 / RESEARCH
Take the protocol
The leave-one-site-out evaluation, the audited splits and the fitted detectors are open. If you are measuring cross-site transfer, start from ours rather than rebuilding it.
03 / BUILD
Work on it
Detection research, field hardware, and getting alerts to people who do not read the language the model thinks in. All three are open problems.