Predicting the duration of wildfires using machine learning techniques

In the US Wildfires have become more frequent over the previous decade increasing in size and frequency. Alongside many computational efforts to culminate the factors effecting the spread of these wildfires we utilise Convolutional Neural Networks (CNN) alongside Autoencoder/Random Forest (AE)(RF) ensembles to predict the duration of these wildfire and areas burnt. Our data is centred around the western state of California with the region providing the various conditions for frequent Wildfires from previous literature. We took data from Google Earth of ecological features alongside fire spread data to construct the various training/testing datasets. Various CNN architectures were tested to get the optimal predictability over a given time frame and random search was utilised for hyperprameter optimisation for prediction accuracy. The Fire Image dataset was pre-processed to create a regression problem for number of days left for fire to burn with further ensembles for investigation with ecological factors in the form of 3d tensors-where each ecological factors was concatenated as a layer to tensor. To rival the state of the art CNN architecture an auto-encoder/ random forest ensemble was constructed, the autoencoder was utilised for dimensionality reduction to image data reducing the X and Y dimensions from 200x200 to 10x10-tested as optimal dimension for predictability with random forest