Probabilistic Fusion Framework Based on Fully Convolutional Networks and Graphical Models for Burnt Area Detection from Multiresolution Satellite and UAV Imagery

Mapping areas affected by wildfires is an important task for fire disaster management. When dealing with these applications, it is common to have access to images of the same scene collected by very different acquisition systems. The development of processing methods capable to benefit from these multimodal datasets is promising for wildfire management but is generally challenging. This paper addresses the semantic segmentation of zones affected by forest fires from input multimodal imagery collected by both unmanned aerial vehicles (UAVs) and satellite platforms. The multiresolution fusion task is especially challenging in this case because the difference between the involved spatial resolutions is very large – a situation that is normally not addressed by traditional multiresolution schemes. Two novel multiresolution fusion approaches, based on deep learning, Bayesian and probabilistic graphical fusion models are proposed. The first method performs the multiresolution fusion of the multimodal imagery via a probabilistic decision fusion framework, after computing posteriors on the multiresolution data separately with deep neural networks or decision tree ensembles. The optimization of the parameters of the model is fully automated through an approximate formulation of the expectation maximization (EM) algorithm. The second proposed method performs the multiresolution fusion of the multimodal data through a probabilistic graphical model defined on a pyramidal tree structure, where the imagery can be inserted, modeled, and analyzed at their native resolutions. The experimental validation was conducted with three real datasets consisting of UAV and satellite data collected over the South of France over areas affected by wildfires.