Title: Wildfire Spread Burned Area Prediction using Deep Learning
This project aims to predict final burned area from a wildfire. Using a spatial and temporal dataset with 28 remote sensing variables in 64x64km patches where each pixel has a 1x1km resolution and 1-day temporal resolution. Deep Learning Models are made in Pytorch. Models are based on UNet2D and UNet3D. A baseline model is trained with only 1 day which consists of 28 variables and the burned area mask. A UNet2D model is trained using 10 days of spatial and temporal data, each sample having 5 days before the fire and 5 days after the fire, totaling 10 days. Finally, a UNet3D model is trained with the key difference being that for every feature map UNet3D generates, data from 1 day before and 1 day after are used in combination with the current day the model is processing. To see results and project structure overall, visit the GitHub page. You can download the test results of every model in binary form or shapefile form and open them with QGIS.