Wildfires are one of the natural hazards that the European Union is actively monitoring through the Copernicus EMS Earth observation program which continuously releases public information related to such catastrophic events. Such occurrences are the cause of both short- and long-term damages. Thus, to limit their impact and plan the restoration process, a rapid intervention by authorities is needed, which can be enhanced by the use of satellite imagery and automatic burned area delineation methodologies, accelerating the response and the decision-making processes. In this context, we analyze the burned area severity estimation problem by exploiting a state-of-the-art deep learning framework. Experimental results compare different model architectures and loss functions on a very large real-world Sentinel2 satellite dataset. Furthermore, a novel multi-channel attention-based analysis is presented to uncover the prediction behaviour and provide model interpretability. A perturbation mechanism is applied to an attention-based DS-UNet to evaluate the contribution of different domain-driven groups of channels to the severity estimation problem.

The use of theory-based knowledge in machine learning models has a major impact on many engineering and physics problems. The growth of deep learning algorithms is closely related to an increasing demand for data that is not acceptable or available in many use cases. In this context, the incorporation of physical knowledge or a priori constraints has proven beneficial in many tasks. On the other hand, this collection of approaches is context-specific, and it is difficult to generalize them to new problems. In this paper, we experimentally compare some of the most commonly used theory-injection strategies to perform a systematic analysis of their advantages. Selected state-of-the-art algorithms were reproduced for different use cases to evaluate their effectiveness with smaller training data and to discuss how the underlined strategies can fit into new application contexts.