Canopy Fire Effects Estimation Using Sentinel-2 Imagery and Deep Learning Approach. A Case Study on the Aspromonte National Park

The accurate estimation of fire severity, in terms of physical effects that occurred on the tree’s canopies, as well as the accurate mapping of its spatial distribution, is necessary information to optimally quantify and qualify the damage caused by the fire to ecosystems and address the most correct remedial procedures. The development of even more accurate learning algorithms and higher resolution satellite multispectral data have become essential resources in this framework. This study proposes a deep learning approach, exploiting remotely sensed satellite data, to produce an accurate severity map of the effects caused by the devasting fires affecting the Aspromonte National Park forests during the 2021 fire season. Two multispectral Sentinel-2 data, acquired before and after the fires, were classified using an artificial neural network-based model. All the multispectral fire-sensitive bands (visible, near-infra-red, short-infrared) and the respective temporal difference (post-fire—pre-fire) were involved, while the selection of the training pixels was based on field-based observations. Despite the preliminary nature of this study, the map accuracy reached high values (>95%) of F-scoreM (representing the overall accuracy) already since the first test of this configuration, confirming the validity of this approach. The quanti-qualification of the fire effects reported that 35.26 km2 of forest cover was affected, of which: 41.03% and the 26.04% of tree’s canopies were low and moderately affected respectively; the canopies killed but structurally preserved were the 12.88%; the destroyed trees (very-high severity), instead, were the 20.05%.