InfestoWatch: Early Pest Infestation Detector with Edge AI

1. Introduction

Forest pests like bark beetles cause massive ecological and economic damage. Unfortunately, most infestations are detected too late – after the tree has already been compromised. InfestoWatch is a solar-powered sensor node that uses Edge AI to detect early warning signs from trees, using a combination of VOC gas sensing and acoustic vibration detection.

By analyzing tree-emitted gases and internal chewing sounds, this low-power, deployable solution helps foresters catch infestations early and take action before the damage spreads. The device runs a machine learning model locally, allowing it to work completely off-grid in remote forests or plantations.

Pest outbreaks—especially from bark beetles—destroy millions of trees every year and cost billions in forestry damage. But by the time infestations are visible, it’s often too late. I wanted to create a low-cost, low-power, and deployable solution that helps foresters detect trouble early, enabling targeted intervention and saving trees before the damage spreads.

Most current solutions focus on satellite or drone imagery, which are either delayed or costly. I felt there was a gap for a real-time, ground-level sensing solution using Edge AI.

2. System Overview

InfestoWatch combines

• Sensing:It uses a BME688 gas sensor mounted on a tree to measure VOCs like ethanol, α-pinene, and other stress markers emitted by the tree.
• Processing on the Edge:A microcontroller runs a TinyML model (trained using Edge Impulse) to classify the readings as "Healthy" or "Infested."
• Alerting:If an infestation pattern is detected, the node blinks an LED, logs the event, and sends a LoRa alert to a nearby gateway and further to a web app /smartphone.

It can run on battery + solar, making it suitable for long-term deployment in forest environments.

3. Pest Infestation Simulation for Model Training

Tosimulate pest infestation stress on trees for VOC data collection using the Bosch BME688 development kit.

MY first goal is to create conditions that mimic early tree stress or pest activity (like bark beetle infestation) to collect VOC profiles using our 8-sensor BME688 board.

According to my research, I can use Bosch AI-Studio or edge impulse to Collect gas signatures and Train models visually, assign labels like “pest-infested bark, ” “damaged, ” “control, ” etc.

• Simulate infestation conditions: bark damage, alcohol sprays (for ethanol), pine extract
• Use the BME688 dev kit to collect real-time gas profile curves
• Label and train model in Bosch AI-Studio, export to Edge Impulse
• After Training, port the final model to your Nano 33 BLE Sense.
• Use just 1 BME688 sensor on the deployment node to keep it power-efficient.

Simulated Experiment Plan

Prepare Test Surfaces - Real tree bark samples or ntreated wood blocks

To simulate stress triggers I plan to take lightly pierce bark (simulate beetle boreholes). Apply a small drop of:

• Ethanol (mimics beetle infestation stress VOC)
• Turpentine or pine essential oil (α-pinene analog)

3. Model Training (TinyML Fusion)

Dataset Collected paired data:

• gas_resistance (kΩ from BME688)
• piezo_rms (average analog vibration amplitude)
• Labels: HEALTHY, INFESTED
• Uploaded CSV to Edge Impulse with two input features.

Impulse Setup

• Input block: Raw features (2)
• Processing: Flatten
• Learning block: Fully Connected Neural Network (FCNN)

4. Hardware Setup

5. Firmware & Logic

Arduino reads both sensors: Computes RMS from piezo analog signal

• Computes RMS from piezo analog signal
• Reads gas resistance from BME688
• Packs values into a feature vector
• Feeds vector to TinyML model for inference
• If prediction = INFESTED, blink LED and send LoRa alert.

6. Deployment

• Enclosure
• Solar setup
• LoRa/MQTT notification

7. Results

Status: Initial tests on bark samples + pine resin ongoing

Planned to record the following

• Elevated VOC levels
• Detectable vibration patterns from simulated chewing
• Fused model triggers alert within 15 seconds of condition onset

Implementation of the concept has just started. The hardware is being tested, and model training is in progress. Stay tuned — we’ll be sharing real-world tests, images, and demo videos soon!