Modelling forest fire dynamics using conditional variational autoencoders

Forest fires have far-reaching consequences, threatening human life, economic stability, and the environment. Understanding the dynamics of forest fires is crucial, especially in high-incidence regions. In this work, we apply deep networks to simulate the spatiotemporal progression of the area burnt in a forest fire. We tackle the region interpolation problem challenge by using a Conditional Variational Autoencoder (CVAE) model and generate in-between representations on the evolution of the burnt area. We also apply a CVAE model to forecast the progression of fire propagation, estimating the burnt area at distinct horizons and propagation stages. We evaluate our approach against other established techniques using real-world data. The results demonstrate that our method is competitive in geometric similarity metrics and exhibits superior temporal consistency for in-between representation generation. In the context of burnt area forecasting, our approach achieves scores of 90% for similarity and 99% for temporal consistency. These findings suggest that CVAE models may be a viable alternative for modeling the spatiotemporal evolution of 2D moving regions of forest fire evolution.