Arduino Genuino/Uno with sensors
Machine learning parameter correlations and probability of forest fire occurrences
In the past few years, California has been subject to intense wildfires that have devastated so many lives and homes. Most recently in November 2018, we had Camp Fire (originating in Butte County) strike down 18,000 structures and leave 85 people dead. The tragedy of wildfires is often difficult to deal with and poor air quality can cause extremely harmful impacts on human and environmental health as well. Whether people have just watched their house go up in flames or are seeing hazy skies up above, we want to help victims of wildfire tragedy or those at risk of fires know about the nearest safety zones, wildfire locations, and notifications on how to proceed. Fires always start as a small ember, and we want to help society before it blazes into a flame.
What it does
Forest fires have environmental impacts that create economic problems as well as ecological damage. Developing a means to predict the occurrence of a forest fire shortly after it first breaks out has the potential to guide proper resource allocation for improved fire control. In this study, the possibilities of predictions resulting from possible forest fires were estimated using historical forest fire records. These contained parameters like geographical conditions of the existing environment, date and time when the fire broke out, meteorological data such as temperature, humidity and wind speed, and the type and number of trees in a unit area.
How I built it
We've developed our idea into 3 zones! 1) Device Prototyping - By using an Arduino/Geniuno Uno foundation, we configured our board to collect real-time data for temperature, air quality via CO2 emissions monitoring, and soil moisture. We envision scattering these devices throughout the state and in other high-risk fire regions to constantly monitor these parameters and inform locals about the current status quo of wildfires in the area. Data collected from the hardware can be added to the current dataset’s csv file for continuous intake of data. 2) Machine Learning Model - This model was trained using historical forest fire records in California which contained geographical conditions of the existing environment such as average temperature, relative humidity, wind speed. At certain conditions, historically, for example high temperatures, low wind speed and low humidity, there have been a higher occurrence of forest fires. Hence, we can use this pattern to predict within a certain confidence level when a forest fire could occur based on recurring weather/atmospheric conditions. Previously, experts confirmed an artificially intelligent model during 13th EPIA 2007 using Gaussian SVM with just 4 parameters to predict forest fire possibilities based on RMSE and MAD. This model works best for small fires but alternatively for forest fires affecting wider regions, so we approached it differently. We set up a Dense Neural Network using TensorFlow Regressor to prepare our model to predict widespread forest fires. Hence, we considered multiple aspects of the available features to classify. Our network composed of three layers associated with a hundred thousand steps to achieve better results, to evaluate with RMSE of Scikit-Learn. 3) Web App Visualization - The web app seamlessly acts as a user friendly approach to the wildfire problem. It offers a simple digital platform to see the locations of live fires, nearby fire relief areas, and offers a notification-subscription feature implemented through esri APIs and ArcGIS. The app itself allows for on-the-go updates of possible fires based on a person’s location, creates a route from where the person is to the nearest relief camp, in addition to humanitarian organization links that the user can easily navigate to for donating to wildfire aid relief.
Accomplishments that I'm proud of
For us four founders of ember, with eclectic skills and too many inside jokes between us, we are so proud of the product we developed. Making our ember prototype taught us all so much and allowed us to teach each other new skills: designing machine learning algorithms, working together in hardware hacking, and training our model to predict wildfire locations. We're also proud to have created a very practical tool that has a widespread, real-world application to enforce better wildfire safety and awareness. All in all, we really just wanted to create an innovative, problem-solving system that could better the world and make a difference, and we know this is the start of something big.
What I learned
First and foremost, there is so much involved in hacking an integrated software-hardware device! We learned how to better apply our individual strengths towards different tasks, integrate all the components on a time crunch, and just find efficient alternatives when things got hard. So being resourceful was key to this whole project!
What's next for ember
ember is only going to burn brighter as we take over wildfires! Over time, we'd love to work on more sophisticated machine learning algorithms, refine the hardware for our device to be resistant against environmental degradation, and make our product ultra-affordable and accessible to communities around the world. ember would be an amazing addition to regions most susceptible to fire risks, providing pre-emptive alerts about upcoming wildfire dangers and making sure that people are safe before the fire spreads.