Modelling global burned area with deep learning

The temporal coverage from ˜2000 to present of global burned area satellite observations limits many aspects of fire research. As a result, global fire models are often being used to investigate past and future fire behaviour. Unfortunately, the limited temporal coverage of the observations also hinders the development and evaluation of these fire models. The current generation of global fire models are capable of simulating some characteristics of regional fire behaviour, such as mean state and seasonality, well. However, the performance of these models differs greatly from region to region, and aspects such as extreme fire behaviour are not well represented yet.

Here, we propose a new, data-driven fire model that predicts burned area from the same input parameters that are passed to global fire models. We trained LSTMs to model burned area from GFED5. We split our data according to the IPCC regions and perform a region-based cross-validation, that is, we train different LSTMs on different region-splits of the data. We then compose the predictions of these different models so that for each region the predictions are made by LSTMs that have never seen any data during training and validation from that region before. Our model outperforms all fire models on a global scale and in most IPCC regions. With our model, we can improve our understanding of past fire behavior and simulate future fire trends.