Forest Guardian: an alarm network independent of the internet and the power grid

The problem

Jordan's reserves, such as Dibeen, Ajloun, and Azraq, cover large areas with no cellular coverage or power, while the number of rangers is limited. As a result, poaching during the migration season and illegal logging in winter are discovered only by chance or after the fact. By the time a fire's smoke is visible from afar, it has usually spread. And any solution that relies on the cellular network can't run in exactly the places that need it most.

How it works

1. ListeningA simple detector monitors the ambient sound level while the node sleeps.
2. ActivationAny sudden loud noise wakes the model; the model doesn't run continuously.
3. ClassificationThe model classifies the sound as a gunshot, a chainsaw, or a normal sound, with a confidence score.
4. Smoke detectionThe smoke sensor and weather readings add an early fire warning.
5. PassingThe alarm is sent as a short message that hops from node to node until it reaches the gateway.
6. ReceivingThe gateway synchronizes time, removes duplicate messages, alerts the ranger, and shows the alarm on a map.

Project features

• Recognition happens on the node itself, with no internet, cloud service, or subscription.
• The system handles two classes of different difficulty: a gunshot is momentary and hard to recognize, while a chainsaw is continuous with a stable signature and easier to recognize; comparing the system's performance on both is a research result in itself.
• The mesh network carries the alarm from distant nodes by hopping between nodes, with no communication tower needed.
• The node relies on deep sleep and wakes on a sudden sound, which lets it run on a small solar panel.
• The system never stores or sends any audio recording; it only sends the model's classification result, protecting residents' privacy.

Expected challenges

These are the main obstacles to expect during the build, laid out clearly so you know them before you start.

• ⚠ Acoustic monitoring also picks up human voices, so no audio must be stored, and deploying nodes near populated areas requires prior permission.
• False alarms are the biggest challenge: distant cars, thunder, and breaking branches sound like gunshots. Reducing them requires training the model on sounds from the installation site itself rather than foreign recordings.
• The energy budget is the core of the project: running the model continuously drains the battery, so waking on sound is a basic requirement, not just an optimization.
• A network that rebroadcasts every message to all nodes drains every node's battery with each message, so it needs repeater nodes on high ground instead of flooding.
• Range drops a lot under dense trees, from several kilometers to hundreds of meters, which hopping between nodes compensates for.
• Locating the sound source by time-difference of arrival requires sync better than one millisecond, since one millisecond of error is about 34 cm; this feature is outside the project's core scope.

The consultation gives you a solution for each one.