Embero is an advanced, multi-tiered AI and IoT ecosystem designed to turn local schools and communities into proactive guardians against wildfires. Specifically optimized for the micro-climate and topography of Montenegro, Embero shifts the paradigm from passive disaster response to active, student-led climate resilience and community awareness.
Every summer, students in Montenegro witness the devastating impact of wildfires firsthand, watching local forests burn around their towns and schools. The statistics are catastrophic:
- Economic & Environmental Loss: Montenegro loses over 100 million EUR annually to forest fires.
- Damaged Ecosystems: Between 2011 and 2023, a staggering 190,348 hectares of land were destroyed.
- The Human Factor: 95% of these fires are caused by human activity (carelessness, clearing debris, agricultural burning).
Despite Serbia being seven times larger, it suffers ten times fewer fires, proving that Montenegro's vulnerability is primarily a systemic issue due to the lack of an early detection framework. As members of our school's eco-club at JU Srednja mješovita škola Golubovci, we refused to stay helpless. We realized that while large infrastructure takes years to deploy, local schools and communities can use decentralized AI and hardware to monitor their immediate environments, learn from real-time data, and stop disasters before they spread.
Embero bridges the gap between complex IoT engineering and grassroots community action. It functions through four interconnected pillars:
-
Solar-Powered Smart Sensor Stations (24/7 Ground Grid): Decentralized physical stations deployed in high-risk zones around the community and school perimeters. They monitor environmental micro-climates continuously, tracking sharp spikes in temperature, drops in humidity, and anomalous increases in
$CO_2$ and smoke particulates. - Multi-Tiered Autonomous Drone Fleet (Aerial Scouting): When ground sensors log data anomalies, the system triggers targeted drone missions (using tethered, small tactical, or heavy-duty frames depending on weather conditions) to verify risks and prevent costly false alarms.
- YOLOv8 Edge & Server AI Vision Pipeline: Video and image feeds from ground stations and drone RGB/Thermal cameras are passed through an optimized YOLOv8 model trained to classify and isolate actual smoke and flame patterns from shifting mountain fog, clouds, or heavy road dust.
- The School & Community Action Dashboard: Instead of routing data exclusively to government agencies, Embero populates an interactive dashboard built for school eco-clubs, student leaders, and teachers to visualize risks and take proactive community safety steps.
- Processing & Core Logic: ESP32-S3 / Arduino Giga R1 WiFi acting as the main edge compute nodes.
- Telemetry Core: LoRaWAN RN2483 (868MHz transceiver) enabling secure long-range data transfer completely independent of cellular networks.
-
Sensor Payload: SHT31/DHT22 (Temperature & Humidity), Senseair S8/MH-Z19C (
$CO_2$ Infrared Sensor), Sharp GP2Y1010AU0F (Optical Dust/Smoke Sensor). -
Power Subsystem: Highly durable
$LiFePO_4$ batteries connected to a 20-30W solar panel system. This chemistry provides thermal stability and eco-friendly recycling cycles. -
Smart Ingestion Sampling Algorithm: To preserve
$LiFePO_4$ battery life during low-solar winter days, the ESP32-S3 utilizes a dynamic sampling algorithm. When environmental baselines are stable, the station remains in ultra-low-power Deep Sleep, transmitting telemetry once every 15 minutes. If a$CO_2$ delta spike exceeding$50\text{ ppm}$ per minute is detected, the node immediately forces an "Emergency Power State," entering a real-time 30-second transmission loop. - Micro-Climate Sensor Calibration: Gas sensors (MQ-135 and Senseair S8) use automated baseline calibration (ABC). Custom compensation matrices run natively on the edge to filter out atmospheric pressure fluctuations and high humidity deltas common across mountainous terrain using real-time values from the SHT31 sensor.
- Frame: Lightweight, impact-absorbent Carbon Fiber.
- Propulsion: High-RPM Racerstar BR2205 2300KV Brushless Motors matched with APC Slow Fly 10x4.5 propellers optimized for stable, high-payload data-gathering flights.
- Power: ZOHD LIONPACK 5000mAh 4S Li-ion pack providing 20-25 minutes of steady operational flight.
- Optics Subsystem: Dual-camera payload combining a RunCam Split 2 HD visual sensor and an AMG8833/FLIR Lepton thermal array.
- Autonomous Launch Trigger Mechanism: The drone is housed inside an automated all-weather roof enclosure at the school site. Upon central cloud cross-validation of a ground risk event, the bay opens, and the flight controller (Pixhawk/Arduino Core) takes an automated flight path payload pushed over MAVLink directly to the target node's coordinates.
-
Mission Fail-Safe Protocols: In the event of a total LoRaWAN or local WiFi signal loss over dense mountain terrain, the onboard flight computer switches automatically to a hardware-interrupt Return-to-Home (RTH) routine utilizing a dual GPS/Glonass array with an accuracy of
$\pm 0.5\text{m}$ .
⚠️ Important Prototype Note: Embero is a wildfire detection system. For this hackathon, we built a GitLab Orbit skill prototype that demonstrates how environmental AI alerts can auto-generate GitLab issues. The YAML skill successfully defines the workflow and logic; full live API integration is the next step on our roadmap.
The agent is implemented as a configuration layer designed to translate critical environment thresholds into actionable software tasks.
The skill tracks data logs from real-time environmental streams and prepares structured issue creation payloads whenever anomalies cross safe bounds:
| Data Source | What the Skill Tracks |
|---|---|
| Solar Tracker | Temperature, humidity, light levels, LDR sensor arrays |
| Drone AI Camera | YOLOv8 smoke/flame detection confidence scores |
| Ground Sensors |
|
The decision-making layer evaluates conditions using cross-validated thresholds to prevent false positives:
| Condition | Risk Weight | Intended Action Trigger |
|---|---|---|
| Temperature > 60°C | Critical | Auto-Generate High-Priority GitLab Issue |
| Humidity < 15% | High | Log Local Climate Risk Warning |
| AI Camera detects smoke/flame | Critical | Flag Visual Alert Overlay |
| Temperature 45–60°C | Moderate | Trigger Secondary Drone Scouting Stream |
The structural logic, YAML definition workflow, and template structures are entirely open-source. School eco-clubs can fork this repository, adapt the threshold values to match their local geography, and contribute optimized climate rules back to the network.
Automated emergency systems face massive challenges with false alarms, which drain public trust and local resources. Embero solves this through strict Multi-Source Cross-Validation and Human-in-the-Loop (HITL) governance:
- Cross-Validation Layer: The central AI module will not escalate an alert based on a single variable. For an official risk alarm to trigger, the system requires a simultaneous, localized data match based on our custom evaluation formula:
The system triggers a high-alert flag only when optical smoke detection
-
Edge AI Optimization & Temporal Persistence: To compute visual telemetry at the edge, a quantized YOLOv8-nano model optimized to INT8 precision ONNX format runs locally on a low-power drone compute stick (or Raspberry Pi Zero 2 W). This achieves a solid
$\approx 15\text{ FPS}$ inference rate while pulling under$2.5\text{W}$ of power. To maximize reliability, the classification stack utilizes temporal persistence validation: a smoke/flame class must be detected in 5 sequential frames over a 5-second sliding window to rule out moving mountain fog, sun flares, or transient road dust. -
Human-in-the-Loop Checkpoint: Embero does not automate evacuations or public panic. The AI handles the heavy cognitive lifting—processing video streams, mapping vectors, and tracking real-time GIS propagation data. However, the dashboard presents this curated information directly to the supervising school coordinator or teacher, keeping a human expert firmly in charge of confirming and escalating the alert.
-
Mitigating Thermal Drift in Rocky Landscapes: During peak summer heat in parts of Montenegro like Podgorica and Golubovcima, stone surfaces often radiate heat upwards of
$50^\circ\text{C}$ , causing standard thermal cameras to trigger false alerts. We designed a differential thermal mapping algorithm that tracks rapid local temperature gradients over a compressed area rather than absolute heat values, isolating actual open flame kinetics from ambient summer heat. -
Canopy Penetration & Dense Vegetation: Early fires under heavy pine canopies are completely invisible to standard visual sensors. Embero circumvents this by utilizing NDIR
$CO_2$ gas tracking on the ground to pick up chemical plumes rising through the foliage, which then prompts the thermal-overlay camera on the drone to isolate hotspots invisible to the naked eye. - Remote Network Failures: Wildfire zones rarely have stable 5G coverage. We bypassed this limitation by choosing LoRaWAN as our primary transmission protocol, ensuring data moves through low-frequency radio waves directly to a community gateway.
- Reframing the End User: We originally designed Embero as a top-down government utility. Overhauling the concept for the global hackathon allowed us to see its true potential: embedding the UI directly into the hands of student eco-clubs creates instantaneous, hyper-local climate accountability.
All technical assets are open-sourced and structured across the following directories in this repository:
- Full wiring diagrams and PCB pinout maps for the Arduino Giga R1 and ESP32-S3 CAM base processing boards.
- C++ code tailored for the Arduino IDE, optimizing low-power deep sleep cycles (
$<15\mu A$ ) between sensor sampling intervals. - Custom compensation matrices for the Senseair S8 NDIR CO2 sensor and the Sharp optical dust sensor to filter out baseline humidity variations.
- Implementation of OTAA (Over-The-Air Activation) join sequences and data payload serialization scripts to compress telemetry packets into lightweight, 12-byte arrays.
- Jupyter notebooks containing training configurations, dataset augmentation parameters (synthetic smoke injection, color jittering), and hyperparameters used for our model.
- Conversion workflows transforming the native PyTorch weights (
.pt) into optimized, quantized INT8 ONNX formats ready to be executed on edge microcontrollers and drone compute sticks. - Real-time stream processing infrastructure that manages RTSP video inputs from the drone, applies frame structural analysis, and runs spatial inferencing.
- Clean, component-based user interface source code utilizing Tailwind CSS for fluid, fully responsive visual components.
- Production code integrating Leaflet.js maps with active GIS overlays to display live wind data, localized cloud patterns, and school-specific threat parameters.
- Built-in Node.js server scripts providing endpoints for secure MQTT sensor input ingest, storing historical log deltas, and updating real-time emergency states.
- Production YAML workflows establishing the GitLab Orbit integration schema.
- Structural mock input logs mapping sensor anomalies to automated issue tracking structures.
- Automation templates outlining issue descriptions, markdown severity reports, and coordinator routing.
Embero is structurally engineered to be incredibly cost-effective, proving that saving a country's ecosystem does not require multi-million dollar investments up front:
- ESP32-S3 Core Board + Camera Module — 15 EUR
- Senseair S8 Infrared
$CO_2$ Sensor — 25 EUR - LoRaWAN RN2483 RF Transceiver Module — 18 EUR
- SHT31 High-Accuracy Temp & Humidity Sensor — 5 EUR
- Sharp Optical Dust/Smoke Sensor — 8 EUR
- 20W Rugged Solar Panel + MPPT Charging Board — 30 EUR
- High-Cycle
$LiFePO_4$ Battery Pack (12V 6Ah) — 25 EUR - Weatherproof IP66 Chassis & Mounting Assembly — 12 EUR
- TOTAL NODE PRODUCTION COST: 138 EUR
- Phase 1 – The Pilot (2026): Deploying a mesh of 50 stations and 5 coordinating drones across 3 high-risk municipalities (Podgorica, Nikšić, Bar) requires a total budget of only 10,000 EUR.
-
Phase 2 – National Scaling (2028+): To completely blanket the 8,000
$km^2$ of vulnerable forestry in Montenegro requires roughly 4,000 interconnected sensor stations. The entire national grid scales to an estimated 725,000 EUR.
📊 The Bottom Line: Protecting the entire country via Embero costs less than 3 days of the structural and environmental damage Montenegro currently loses to wildfires every year (100M EUR).