摘要
Since 2000, seven million acres have burned every year. Yet, since robust analytics are scarce, capitalizing on machine learning algorithms have the capability to bridge gaps in decision making and effective deployment. Despite this, a major limitation in current research is resolution and accuracy. Utilizing public data from NASA’s MODIS, LP DAAC, University of Idaho, and UC Irvine, 12 input features and 18,545 samples, the fire mask at day t+1 is predicted. Compared to existing datasets (FRY, FireAtlas, UCI Forest Fires, US Wildfires Catalog, Globfire and European Forest Fire Information System), this dataset contains the most variables at 1 km. resolution with the most input features. By treating the fire mask as binary and probability maps, regression and classification were performed. Several novel architectures were tested (ResNet, EfficientNet, RegNet and VGG19). A dataset scaling algorithm helped improve resolution by predicting data from existing points. The most optimal models were ResNet and Efficient Net, achieving a binary accuracy of 96.58%, precision of 72.37% and mean absolute error of 0.036. Compared to current studies, this study is around 38% more precise with 0.0142 lower mean absolute error, a significant improvement. Implementation regarding spread was implemented in two ways. With classification data and substituting resulting fire masks for previous ones, the spread of a wildfire could be mapped for various days. Additionally, with population density data and this spread, escape routes were also predicted.