Experimental Validation of a Wildfire Early Warning System Based on a CO2 Sensor Network

Forest fire prevention, forecast, and control are becoming increasingly popular issues, in large part because of climate change. While several early warning systems use remotely sensed images collected by optical and non-optical sensors, as well as supervised AI (Artificial Intelligence) algorithms to detect fires early on, the development and dissemination of reliable, low-cost sensors together with the advancement of the IoT (Internet of Things) paradigm make it possible to apply monitoring techniques relying on widespread ground-based sensor networks.

This paper illustrates an innovative technique where smart CO₂ sensors were used to capture smoke produced by combustion and discriminate an alert through AI techniques. In more detail, a small-scale field experiment was conducted where 44 CO₂ sensors were deployed on a hillslope, triggering a small controlled fire. The sensors were connected via LoRaWan (Long Range Wide Area Network) technology and a gateway to an online platform that included an optimized database and an interactive management interface. Several environmental variables were monitored during the experiment, most notably wind speed and direction. In addition, 3 unsupervised AI algorithms were tested to discriminate alerts (Long-Short Term Memory - LSTM; AutoEncoder on CO₂ absolute values and AutoEncoder on CO₂ differences between two consecutive measurements) and compared with a classical alert system based on thresholds calibrated on each sensor, using the maximum CO₂ recorded in the 5 days prior the experiment, in absence of fires.

Several sensors detected anomalies in CO₂, particularly those placed downwind. The results highlighted the capabilities of AI to better discriminate the alert with respect to the classical no-AI system. More specifically, the application of AI-based methods could also bring the alert on many sensors forward with respect to the no-AI method. Future deployments of such a system will be carried out in a broader area, employing more than double the number of sensors and combining them with other detection technologies (e.g., remotely sensed RGB and IR images) and AI techniques.

 

Acknowledgments: This study was funded by The Next Generation EU—Italian NRRP, Mission 4, Component 2, Investment 1.5, call for the creation and strengthening of ‘Innovation Ecosystems’, building ‘Territorial R&D Leaders’, and Project Tech4You—Technologies for climate change adaptation and quality of life improvement, n. ECS0000009.  This work reflects only the authors’ views and opinions, neither the Ministry for University and Research nor the European Commission can be considered responsible for them.