Figure 1.

Ensemble deep learning mechanism for the weekly bush-fire frequency estimation and bush-fire weekly hot-spot estimation. This mechanism has two stages: the first one is the unsupervised deep learning phase and the second one is an ensemble supervised classification phase. In the unsupervised stage, multi-layered deep neural networks were employed to learn about the given data and generate simplified features without any prior information or training targets. In the ensemble classification stage, multiple supervised classifiers were used to learn the extracted features against the ground truth bush-fire host spot maps. Training inputs to the second stage are features (extracted in the first stage) from the climate maps along with the two estimated bush-fire frequencies (depicted in figure 2) based on the climatic zone boundaries and the administrative state boundaries on the Australian map. Outcome from this ensemble deep learning mechanism is a map with pixel (defined by latitude and longitude) based predicted bush-fire hot spots in multiple colours, according to the most probable category of the fire intensity or severity level (as defined in the electronic supplementary material, figures S2–S17). Maps and figures were generated using Matlab software packages. Copyright © CSIRO, Australia.