Despite having many indicators of a specific area’s location using meteorological instruments and live data, wildfires are only ever detected and responded to once they have already begun. Due to this fact, responders onto the scene can only tackle an already growing wildfire, and in worse cases a rapidly growing wildfire. What if there was a way to predict a wildfire before it begins and use this information to tackle the fire before it has begun? Or better yet, use the information for deploying autonomous responders? It is known that the main contributors of wildfires are increased temperatures, coupled with the excess exposure to sunlight, dryness and stronger wind conditions, which creates the perfect opportunity for dry wood to collide with other dry wood and ignite due to already high temperature, or for a tree branch to suddenly ignite under focused/intense sunlight.
WildFire Protection is a web service that aims to use machine learning models trained from historical recorded wildfire data to create a system that can predict future wildfires using four independent variables; temperature, light intensity, humidity and wind conditions. By building a model that can process these variables into a percentage, this web service will be able to report live feedback on places where this data is readily accessible. Furthermore, devices are included that can read analog environmental data and report it into a local server used to process that information and display it on the website.
This project was developed on a Vue.js-based frontend, and a Golang and Python-based backend. It was deployed using the Kubernetes software "Docker" and incorporated the Solace PubSub+ WebBroker to communicate between a Qualcomm Dragonboard 410c and the WebBroker over WiFi. The Dragonboard acted as an all-in-one meteorological unit that can observe the environment around it using the Grove temperature and humidifier sensor, as well as a photoresistor. The Dragonboard was also used to communicate its GPS position for planned scalability involving autonomous deployed drones.
Currently the challenge faced by our team was being able to deploy such a robust live service within 24 hours whilst attempting to train a machine learning model using limited. Data was very hard to find, and the hurdles we encountered attempting to use foreign software unknown to us led up to a very rushed and last minute project, albeit pretty much done.