Solution
The Approach
Project SWEDES addresses the rampant issue of wildfires by maintaining a network of points that can each report the distance to a wildfire through machine learning. Current machine learning approaches focus on severity prediction and draw heavily upon spatial data, but such information is not available continuously. I hypothesized that a machine learning model could map continuously measurable meteorological data, such as temperature, humidity, pressure, and wind speed, to wildfire distance. If each distinct point can obtain meteorological data and report an estimated wildfire distance, the precise wildfire location can be triangulated efficiently. A single server can use the coordinates of the points and their individual distances to wildfires to decipher the fire's location.
This solution is continuous, automatic, and accounts for potential failures in its design. Meteorological data is always readily accessible and sensors can automatically pick up variable changes to tweak their estimation of the wildfire distance. The large number of separated points minimizes the risk of a few failed units, which often ruins other systems. If one point is corrupted, the remaining points can still support the system as a whole. In addition, reliability can be ensured by comparing the estimations of nearby points.
Current Results
This project was conducted with Napa County, California as a case study since it is one of the most wildfire-prone regions in the Western United States. With 20 years of meteorological and burned area data, multiple ML models were trained hundreds of times with constant parameter adjustments to minimize error. The Random Forest Algorithm was able to establish the existence of a wildfire with a true positive accuracy of 96%. In the 111km x 111km coordinate grid of Napa, with 121 detection points, an LSTM Neural Network was able to estimate distances to wildfires with an error of 3.6km.
